MANNING Chuck Lam IN ACTION Hadoop in Action Hadoop in Action CHUCK LAM MANNING Greenwich (74° w. long.) For online information and ordering of this and other Manning books, please visit www.manning.com. The publisher offers discounts on this book when ordered in quantity. For more information, please contact Special Sales Department Manning Publications Co. 180 Broad St. Suite 1323 Stamford, CT 06901 Email: orders@manning.com ©2011 by Manning Publications Co. All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by means electronic, mechanical, photocopying, or otherwise, without prior written permission of the publisher. Many of the designations used by manufacturers and sellers to distinguish their products are claimed as trademarks. Where those designations appear in the book, and Manning Publications was aware of a trademark claim, the designations have been printed in initial caps or all caps. Recognizing the importance of preserving what has been written, it is Manning’s policy to have the books we publish printed on acid-free paper, and we exert our best efforts to that end. Recognizing also our responsibility to conserve the resources of our planet, Manning books are printed on paper that is at least 15 percent recycled and processed without the use of elemental chlorine. Manning Publications Co. Development editor: Cynthia Kane 180 Broad St. Copyeditor: Composure Graphics Suite 1323 Proofreader: Katie Tennant Stamford, CT 06901 Composition: Composure Graphics Cover designer: Marija Tudor ISBN: 9781935182191 Printed in the United States of America 1 2 3 4 5 6 7 8 9 10 -- MAL -- 15 14 13 12 11 10 v brief contents PART I HADOOP–A DISTRIBUTED PROGRAMMING FRAMEWORK .................................................. 1 1 ■ Introducing Hadoop 3 2 ■ Starting Hadoop 21 3 ■ Components of Hadoop 37 PART II HADOOP IN ACTION .................................... 61 4 ■ Writing basic MapReduce programs 63 5 ■ Advanced MapReduce 102 6 ■ Programming Practices 134 7 ■ Cookbook 160 8 ■ Managing Hadoop 173 PART III HADOOP GONE WILD.................................191 9 ■ Running Hadoop in the cloud 193 10 ■ Programming with Pig 212 11 ■ Hive and the Hadoop herd 246 12 ■ Case studies 266 vii contents preface xiii acknowledgments xv about this book xvii Author Online xix About the author xx About the cover illustration xxi PART I HADOOP–A DISTRIBUTED PROGRAMMING FRAMEWORK....................... ................................ 1 1 Introducing Hadoop 3 1.1 Why “Hadoop in Action”? 4 1.2 What is Hadoop? 4 1.3 Understanding distributed systems and Hadoop 6 1.4 Comparing SQL databases and Hadoop 7 1.5 Understanding MapReduce 8 Scaling a simple program manually 9 ■ Scaling the same program in MapReduce 12 1.6 Counting words with Hadoop—running your first program 14 1.7 History of Hadoop 19 viii CONTENTS 1.8 Summary 20 1.9 Resources 20 2 Starting Hadoop 21 2.1 The building blocks of Hadoop 21 NameNode 22 ■ DataNode 22 ■ Secondary NameNode 23 JobTracker 24 ■ TaskTracker 24 2.2 Setting up SSH for a Hadoop cluster 25 Define a common account 26 ■ Verify SSH installation 26 ■ Generate SSH key pair 26 ■ Distribute public key and validate logins 27 2.3 Running Hadoop 27 Local (standalone) mode 28 ■ Pseudo-distributed mode 29 ■ Fully distributed mode 31 2.4 Web-based cluster UI 34 2.5 Summary 36 3 Components of Hadoop 37 3.1 Working with files in HDFS 38 Basic file commands 38 ■ Reading and writing to HDFS programmatically 42 3.2 Anatomy of a MapReduce program 44 Hadoop data types 46 ■ Mapper 47 ■ Reducer 48 ■ Partitioner—redirecting output from Mapper 49 ■ Combiner—local reduce 50 ■ Word counting with predefined mapper and reducer classes 51 3.3 Reading and writing 51 InputFormat 52 ■ OutputFormat 57 3.4 Summary 58 PART II HADOOP IN ACTION....................... ..................61 4 Writing basic MapReduce programs 63 4.1 Getting the patent data set 64 The patent citation data 64 ■ The patent description data 65 4.2 Constructing the basic template of a MapReduce program 67 4.3 Counting things 72 4.4 Adapting for Hadoop’s API changes 77 4.5 Streaming in Hadoop 80 Streaming with Unix commands 81 ■ Streaming with scripts 82 Streaming with key/value pairs 86 ■ Streaming with the Aggregate package 90 CONTENTS ix 4.6 Improving performance with combiners 95 4.7 Exercising what you’ve learned 98 4.8 Summary 100 4.9 Further resources 101 5 Advanced MapReduce 102 5.1 Chaining MapReduce jobs 103 Chaining MapReduce jobs in a sequence 103 ■ Chaining MapReduce jobs with complex dependency 103 ■ Chaining preprocessing and postprocessing steps 104 5.2 Joining data from different sources 107 Reduce-side joining 108 ■ Replicated joins using DistributedCache 117 Semijoin: reduce-side join with map-side filtering 121 5.3 Creating a Bloom filter 122 What does a Bloom filter do? 122 ■ Implementing a Bloom filter 124 Bloom filter in Hadoop version 0.20+ 131 5.4 Exercising what you’ve learned 131 5.5 Summary 132 5.6 Further resources 133 6 Programming Practices 134 6.1 Developing MapReduce programs 135 Local mode 135 ■ Pseudo-distributed mode 140 6.2 Monitoring and debugging on a production cluster 145 Counters 146 ■ Skipping bad records 148 ■ Rerunning failed tasks with IsolationRunner 151 6.3 Tuning for performance 152 Reducing network traffic with combiner 152 ■ Reducing the amount of input data 152 ■ Using compression 153 ■ Reusing the JVM 155 ■ Running with speculative execution 156 ■ Refactoring code and rewriting algorithms 157 6.4 Summary 158 7 Cookbook 160 7.1 Passing job-specific parameters to your tasks 160 7.2 Probing for task-specific information 163 7.3 Partitioning into multiple output files 164 7.4 Inputting from and outputting to a database 169 7.5 Keeping all output in sorted order 171 7.6 Summary 172 x CONTENTS 8 Managing Hadoop 173 8.1 Setting up parameter values for practical use 174 8.2 Checking system’s health 176 8.3 Setting permissions 178 8.4 Managing quotas 179 8.5 Enabling trash 179 8.6 Removing DataNodes 180 8.7 Adding DataNodes 180 8.8 Managing NameNode and Secondary NameNode 181 8.9 Recovering from a failed NameNode 183 8.10 Designing network layout and rack awareness 184 8.11 Scheduling jobs from multiple users 186 Multiple JobTrackers 186 ■ Fair Scheduler 187 8.12 Summary 189 PART III HADOOP GONE WILD..................... ...............191 9 Running Hadoop in the cloud 193 9.1 Introducing Amazon Web Services 194 9.2 Setting up AWS 194 Getting your AWS authentication credentials 195 ■ Getting command line tools 198 ■ Preparing an SSH key pair 200 9.3 Setting up Hadoop on EC2 201 Configuring security parameters 201 ■ Configuring cluster type 202 9.4 Running MapReduce programs on EC2 203 Moving your code to the Hadoop cluster 204 ■ Accessing your data from the Hadoop cluster 204 9.5 Cleaning up and shutting down your EC2 instances 209 9.6 Amazon Elastic MapReduce and other AWS services 209 Amazon Elastic MapReduce 209 ■ AWS Import/Export 210 9.7 Summary 211 10 Programming with Pig 212 10.1 Thinking like a Pig 213 Data flow language 213 ■ Data types 213 ■ User-defined functions 214 CONTENTS xi 10.2 Installing Pig 214 10.3 Running Pig 215 Managing the Grunt shell 216 10.4 Learning Pig Latin through Grunt 217 10.5 Speaking Pig Latin 221 Data types and schemas 222 ■ Expressions and functions 223 Relational operators 225 ■ Execution optimization 233 10.6 Working with user-defined functions 233 Using UDFs 234 ■ Writing UDFs 234 10.7 Working with scripts 237 Comments 237 ■ Parameter substitution 238 ■ Multiquery execution 239 10.8 Seeing Pig in action—example of computing similar patents 240 10.9 Summary 245 11 Hive and the Hadoop herd 246 11.1 Hive 247 Installing and configuring Hive 248 ■ Example queries 250 ■ HiveQL in details 254 ■ Hive Sum-up 260 11.2 Other Hadoop-related stuff 262 HBase 262 ■ ZooKeeper 262 ■ Cascading 263 ■ Cloudera 263 Katta 263 ■ CloudBase 264 ■ Aster Data and Greenplum 264 Hama and Mahout 264 ■ search-hadoop.com 264 11.3 Summary 265 12 Case studies 266 12.1 Converting 11 million image documents from the New York Times archive 267 12.2 Mining data at China Mobile 267 12.3 Recommending the best websites at StumbleUpon 272 Distributed beginnings at StumbleUpon 273 ■ HBase and StumbleUpon 274 ■ More Hadoop at StumbleUpon 281 12.4 Building analytics for enterprise search—IBM’s Project ES2 282 ES2 architecture 285 ■ ES2 crawler 287 ■ ES2 analytics 288 Conclusions 296 ■ References 297 appendix HDFS file commands 298 index 302 xiii preface I’ve been fascinated by data for a long time. When I was an undergrad in electrical engineering, I discovered digital signal processing and gravitated toward it. I found out that music, video, photos, and lots of other stuff could be viewed as data. Com- putation was creating and enhancing those emotional experiences. I thought that was the coolest thing ever. Over time, I continued to get excited by new aspects of data. The last few years had exposed me to social and big data. Big data was especially intellectually challenging for me. Previously I had learned to look at data from a statistician’s point of view, and new types of data had “only” asked for new mathematical methods. It wasn’t simple, but at least I had been trained for that, and there was also a wealth of resources to tap into. Big data, on the other hand, was about system-level innovations and new ways of programming. I wasn’t trained for it, and more importantly, I wasn’t alone. Knowledge about handling big data in practice was somewhat of a black art. This was true of many tools and techniques for scaling data processing, including caching (for example, memcached), replication, sharding, and, of course, MapReduce/ Hadoop. I had spent the last few years getting up to speed on many of these skills. Personally I have found the biggest hurdle to learning these techniques is in the middle of the learning curve. In the beginning it’s pretty easy to find introductory blogs and presentations teaching you how to do a “Hello World” example. And when you’re sufficiently well-versed, you’ll know how to ask additional questions to the mailing lists, meet experts at meetups or conferences, and even read the source code yourself. But there’s a huge knowledge gap in the middle, when your xiv PREFACE appetite is whetted but you don’t quite know what questions to ask next. This problem is especially acute for the newest technologies, like Hadoop. An organized exposition that starts with “Hello World” and takes you to the point where you can comfortably apply Hadoop in practice is needed. That’s what I intend this book to be. Fortunately I’ve found Manning’s In Action series to be consistent with this objective, and they have excellent editors that helped me along the way. I had a fun time writing this book, and I hope this is the beginning of your wonderful journey with Hadoop. xv acknowledgments Many people have inspired and made contributions to this book. First and foremost I want to thank James Warren. He led analytics at RockYou, and together we evange- lized the use of Hadoop throughout the organization. I learned a lot from him, and he even lent a hand in some of the initial writings. I’ve been lucky to have many people contribute interesting case studies from outside the Web 2.0 industry. For this I need to thank Zhiguo Luo, Meng Xu, Shaoling Sun, Ken MacInnis, Ryan Rawson, Vuk Ercegovac, Rajasekar Krishnamurthy, Sriram Raghavan, Frederick Reiss, Eugene Shekita, Sandeep Tata, Shivakumar Vaithyanathan, and Huaiyu Zhu. I also want to thank the many reviewers of this book. They provided valuable feedback to my early drafts. In particular, Paul O’Rorke was the technical reviewer that went beyond his call of duty and made some wonderful suggestions on how to make the manuscript better. I look forward to seeing him author his own book at some point. I also enjoyed many long chats with Jonathan Cao. His expertise in databases and large-scale systems provided a broad perspective to understanding the capabilities of Hadoop. The other reviewers who read the manuscript numerous times during development and whom I’d like to thank for their invaluable feedback include the following: Paul Stusiak, Philipp K. Janert, Amin Mohammed-Coleman, John S. Griffin, Marco Ughetti, Rick Wagner, Kenneth DeLong, Josh Patterson, Srini Penchikala, Costantino Cerbo, Steve Loughran, Ara Abrahamian, Ben Hall, Andrew Siemer, Robert Hanson, Keith Kim , Sopan Shewale, Marion Sturtevant, Chris Chandler, Eric Raymond, and Jeroen Benckhuijsen. ACKNOWLEDGMENTS xvi I’ve been blessed to work with a wonderful group of people at Manning. Special thanks go out to Troy Mott, who got me started on this writing project and has been patient enough to see me finish it. Thanks also to Tara Walsh, Karen Tegtmeyer, Marjan Bace, Mary Piergies, Cynthia Kane, Steven Hong, Rachel Schroeder, Katie Tennant, and Maureen Spencer. Their support is simply phenomenal. I couldn’t imagine a better group of people to work with. Needless to say, all the people who contribute to Hadoop and help grow its ecosystem deserve praise. Doug Cutting got it all started, and Yahoo had the foresight to support it early on. Cloudera is now bringing Hadoop to a broader enterprise audience. It’s an exciting time to be part of this growing Hadoop community. Last, but not least, I want to thank all my friends, family, and colleagues for supporting me throughout the writing of this book. xvii about this book Hadoop is an open source framework implementing the MapReduce algorithm be- hind Google’s approach to querying the distributed data sets that constitute the internet. This definition naturally leads to an obvious question: What are maps and why do they need to be reduced ? Massive data sets can be extremely difficult to ana- lyze and query using traditional mechanisms, especially when the queries them- selves are quite complicated. In effect, the MapReduce algorithm breaks up both the query and the data set into constituent parts—that’s the mapping. The mapped components of the query can be processed simultaneously—or reduced—to rapidly return results. This book teaches readers how to use Hadoop and write MapReduce programs. The intended readers are programmers, architects, and project managers who have to process large amounts of data offline. This book guides the reader from obtaining a copy of Hadoop to setting it up in a cluster and writing data analytic programs. The book begins by making the basic idea of Hadoop and MapReduce easier to grasp by applying the default Hadoop installation to a few easy-to-follow tasks, such as analyzing changes in word frequency across a body of documents. The book continues through the basic concepts of MapReduce applications developed using Hadoop, including a close look at framework components, use of Hadoop for a variety of data analysis tasks, and numerous examples of Hadoop in action. MapReduce is a complex idea both conceptually and in its implementation, and Hadoop users are challenged to learn all the knobs and levers for running Hadoop. This book takes you beyond the mechanics of running Hadoop, teaching you to write meaningful programs in a MapReduce framework. ABOUT THIS BOOK xviii This book assumes the reader will have a basic familiarity with Java, as most code examples will be written in Java. Familiarity with basic statistical concepts (e.g., histogram, correlation) will help the reader appreciate the more advanced data processing examples. Roadmap The book has 12 chapters divided into three parts. Part 1 consists of three chapters which introduce the Hadoop framework, covering the basics you’ll need to understand and use Hadoop. The chapters describe the hardware components that make up a Hadoop cluster, as well as the installation and configuration to create a working system. Part 1 also covers the MapReduce framework at a high level and gets your first MapReduce program up and running. Part 2, “Hadoop in action,” consists of five chapters that teach the practical skills required to write and run data processing programs in Hadoop. In these chapters we explore various examples of using Hadoop to analyze a patent data set, including advanced algorithms such as the Bloom filter. We also cover programming and administration techniques that are uniquely useful to working with Hadoop in production. Part 3 is called “Hadoop gone wild” and the final four chapters of the book explore the larger ecosystem around Hadoop. Cloud services provide an alternative to buying and hosting your own hardware to create a Hadoop cluster and any add-on packages provide higher-level programming abstractions over MapReduce. Finally, we look at several case studies where Hadoop solves real business problems in practice. An appendix contains a listing of HDFS commands along with their descriptions and usage. Code conventions and downloads All source code in listings or in text is in a fixed-width font like this to separate it from ordinary text. Code annotations accompany many of the listings, highlighting important concepts. In some cases, numbered bullets link to explanations that follow the listing. The code for the examples in this book is available for download from the publisher’s website at www.manning.com/HadoopinAction. xix Author Online The purchase of Hadoop in Action includes free access to a private forum run by Manning Publications where you can make comments about the book, ask techni- cal questions, and receive help from the author and other users. You can access and subscribe to the forum at www.manning.com/HadoopinAction. This page provides information on how to get on the forum once you’re registered, what kind of help is available, and the rules of conduct in the forum. Manning’s commitment to our readers is to provide a venue where a meaningful dialogue between individual readers and between readers and the author can take place. It isn’t a commitment to any specific amount of participation on the part of the author, whose contribution to the book’s forum remains voluntary (and unpaid). We suggest you try asking the author some challenging questions, lest his interest stray! The Author Online forum and the archives of previous discussions will be accessible from the publisher’s website as long as the book is in print. xx About the author Chuck Lam is currently founding a mobile social networking startup called RollCall. It’s building the social secretary for active individuals. Previously Chuck was a Senior Tech Lead at RockYou. There he developed social applications and data processing infrastructure handling hundreds of millions of users. He applied A/B testing and statistical analysis to tune the virality of social apps. He also optimized RockYou’s ad network with the use of real-time and social data. He was able to improve response rates dramatically, sometimes by an order of magnitude. Chuck first got interested in big data when he began his PhD study at Stanford. He learned of the significant effect big data has on machine learning and began to explore its consequences. His thesis, “Computational Data Acquisition” was the first to investigate creative approaches to data acquisition for machine learning, adopting ideas from areas such as open source software and online games. xxi About the cover illustration The figure on the cover of Hadoop in Action is captioned “A young man from Kistan- ja, Dalmatia.” The illustration is taken from a reproduction of an album of Croa- tian traditional costumes from the mid-nineteenth century by Nikola Arsenovic, published by the Ethnographic Museum in Split, Croatia, in 2003. The illustrations were obtained from a helpful librarian at the Ethnographic Museum in Split, itself situated in the Roman core of the medieval center of the town: the ruins of Em- peror Diocletian’s retirement palace from around AD 304. The book includes finely colored illustrations of figures from different regions of Croatia, accompanied by descriptions of the costumes and of everyday life. Kistanja is a small town located in Bukovica, a geographical region in Croatia. It is situated in northern Dalmatia, an area rich in Roman and Venetian history. The word “mamok” in Croatian means a bachelor, beau, or suitor—a single young man who is of courting age—and the young man on the cover, looking dapper in a crisp white linen shirt and a colorful, embroidered vest, is clearly dressed in his finest clothes, which would be worn to church and for festive occasions—or to go calling on a young lady. Dress codes and lifestyles have changed over the last 200 years, and the diversity by region, so rich at the time, has faded away. It is now hard to tell apart the inhabitants of different continents, let alone of different hamlets or towns separated by only a few miles. Perhaps we have traded cultural diversity for a more varied personal life—certainly for a more varied and fast-paced technological life. Manning celebrates the inventiveness and initiative of the computer business with book covers based on the rich diversity of regional life of two centuries ago, brought back to life by illustrations from old books and collections like this one. Part 1 Hadoop–A Distributed Programming Framework Part 1 of this book introduces the basics for understanding and using Hadoop. We describe the hardware components that make up a Hadoop cluster, as well as the installation and configuration to create a working system. We cover the MapReduce framework at a high level and get your first MapReduce program up and running. 1 3 Introducing Hadoop This chapter covers ■ The basics of writing a scalable, distributed data-intensive program ■ Writing and running a basic MapReduce program Understanding Hadoop and MapReduce ■ Today, we’re surrounded by data. People upload videos, take pictures on their cell phones, text friends, update their Facebook status, leave comments around the web, click on ads, and so forth. Machines, too, are generating and keeping more and more data. You may even be reading this book as digital data on your computer screen, and certainly your purchase of this book is recorded as data with some retailer.1 The exponential growth of data first presented challenges to cutting-edge businesses such as Google, Yahoo, Amazon, and Microsoft. They needed to go through terabytes and petabytes of data to figure out which websites were popular, what books were in demand, and what kinds of ads appealed to people. Existing tools were becoming inadequate to process such large data sets. Google was the first to publicize MapReduce—a system they had used to scale their data processing needs. 1 Of course, you’re reading a legitimate copy of this, right? 4 CHAPTER 1 Introducing Hadoop This system aroused a lot of interest because many other businesses were facing similar scaling challenges, and it wasn’t feasible for everyone to reinvent their own proprietary tool. Doug Cutting saw an opportunity and led the charge to develop an open source version of this MapReduce system called Hadoop . Soon after, Yahoo and others rallied around to support this effort. Today, Hadoop is a core part of the computing infrastructure for many web companies, such as Yahoo , Facebook , LinkedIn , and Twitter . Many more traditional businesses, such as media and telecom, are beginning to adopt this system too. Our case studies in chapter 12 will describe how companies including New York Times , China Mobile , and IBM are using Hadoop . Hadoop, and large-scale distributed data processing in general, is rapidly becoming an important skill set for many programmers. An effective programmer, today, must have knowledge of relational databases, networking, and security, all of which were considered optional skills a couple decades ago. Similarly, basic understanding of distributed data processing will soon become an essential part of every programmer’s toolbox. Leading universities, such as Stanford and CMU, have already started introducing Hadoop into their computer science curriculum. This book will help you, the practicing programmer, get up to speed on Hadoop quickly and start using it to process your data sets. This chapter introduces Hadoop more formally, positioning it in terms of distributed systems and data processing systems. It gives an overview of the MapReduce programming model. A simple word counting example with existing tools highlights the challenges around processing data at large scale. You’ll implement that example using Hadoop to gain a deeper appreciation of Hadoop’s simplicity. We’ll also discuss the history of Hadoop and some perspectives on the MapReduce paradigm. But let me first briefly explain why I wrote this book and why it’s useful to you. 1.1 Why “Hadoop in Action”? Speaking from experience, I first found Hadoop to be tantalizing in its possibilities, yet frustrating to progress beyond coding the basic examples. The documentation at the official Hadoop site is fairly comprehensive, but it isn’t always easy to find straightfor- ward answers to straightforward questions. The purpose of writing the book is to address this problem. I won’t focus on the nitty- gritty details. Instead I will provide the information that will allow you to quickly create useful code, along with more advanced topics most often encountered in practice. 1.2 What is Hadoop ? Formally speaking, Hadoop is an open source framework for writing and running dis- tributed applications that process large amounts of data. Distributed computing is a wide and varied field, but the key distinctions of Hadoop are that it is ■ Accessible—Hadoop runs on large clusters of commodity machines or on cloud computing services such as Amazon’s Elastic Compute Cloud (EC2 ). What is Hadoop? 5 Hadoop cluster Client Client Client Figure 1.1 A Hadoop cluster has many parallel machines that store and process large data sets. Client computers send jobs into this computer cloud and obtain results. ■ Robust—Because it is intended to run on commodity hardware, Hadoop is archi- tected with the assumption of frequent hardware malfunctions. It can gracefully handle most such failures. ■ Scalable—Hadoop scales linearly to handle larger data by adding more nodes to the cluster. ■ Simple—Hadoop allows users to quickly write efficient parallel code. Hadoop’s accessibility and simplicity give it an edge over writing and running large distributed programs. Even college students can quickly and cheaply create their own Hadoop cluster . On the other hand, its robustness and scalability make it suitable for even the most demanding jobs at Yahoo and Facebook. These features make Hadoop popular in both academia and industry. Figure 1.1 illustrates how one interacts with a Hadoop cluster. As you can see, a Hadoop cluster is a set of commodity machines networked together in one location.2 Data storage and processing all occur within this “cloud” of machines . Different users can submit computing “jobs” to Hadoop from individual clients, which can be their own desktop machines in remote locations from the Hadoop cluster. Not all distributed systems are set up as shown in figure 1.1. A brief introduction to other distributed systems will better showcase the design philosophy behind Hadoop. 2 While not strictly necessary, machines in a Hadoop cluster are usually relatively homogeneous x86 Linux boxes. And they’re almost always located in the same data center, often in the same set of racks. 6 CHAPTER 1 Introducing Hadoop 1.3 Understanding distributed systems and Hadoop Moore’s law suited us well for the past decades, but building bigger and bigger servers is no longer necessarily the best solution to large-scale problems. An alternative that has gained popularity is to tie together many low-end/commodity machines together as a single functional distributed system . To understand the popularity of distributed systems (scale-out) vis-à-vis huge monolithic servers (scale-up), consider the price performance of current I/O technology. A high-end machine with four I/O channels each having a throughput of 100 MB/sec will require three hours to read a 4 TB data set! With Hadoop, this same data set will be divided into smaller (typically 64 MB) blocks that are spread among many machines in the cluster via the Hadoop Distributed File System (HDFS ). With a modest degree of replication, the cluster machines can read the data set in parallel and provide a much higher throughput. And such a cluster of commodity machines turns out to be cheaper than one high-end server! The preceding explanation showcases the efficacy of Hadoop relative to monolithic systems. Now let’s compare Hadoop to other architectures for distributed systems. SETI@home , where screensavers around the globe assist in the search for extraterrestrial life, represents one well-known approach. In SETI@home, a central server stores radio signals from space and serves them out over the internet to client desktop machines to look for anomalous signs. This approach moves the data to where computation will take place (the desktop screensavers). After the computation, the resulting data is moved back for storage. Hadoop differs from schemes such as SETI@home in its philosophy toward data. SETI@home requires repeat transmissions of data between clients and servers. This works fine for computationally intensive work, but for data-intensive processing, the size of data becomes too large to be moved around easily. Hadoop focuses on moving code to data instead of vice versa. Referring to figure 1.1 again, we see both the data and the computation exist within the Hadoop cluster. The clients send only the MapReduce programs to be executed, and these programs are usually small (often in kilobytes). More importantly, the move-code-to-data philosophy applies within the Hadoop cluster itself. Data is broken up and distributed across the cluster, and as much as possible, computation on a piece of data takes place on the same machine where that piece of data resides. This move-code-to-data philosophy makes sense for the type of data-intensive processing Hadoop is designed for. The programs to run (“code”) are orders of magnitude smaller than the data and are easier to move around. Also, it takes more time to move data across a network than to apply the computation to it. Let the data remain where it is and move the executable code to its hosting machine. Now that you know how Hadoop fits into the design of distributed systems, let’s see how it compares to data processing systems, which usually means SQL databases. Comparing SQL databases and Hadoop 7 1.4 Comparing SQL databases and Hadoop Given that Hadoop is a framework for processing data, what makes it better than standard relational databases, the workhorse of data processing in most of today’s applications? One reason is that SQL (structured query language) is by design targeted at structured data. Many of Hadoop’s initial applications deal with unstructured data such as text. From this perspective Hadoop provides a more general paradigm than SQL. For working only with structured data, the comparison is more nuanced. In principle, SQL and Hadoop can be complementary, as SQL is a query language which can be implemented on top of Hadoop as the execution engine. 3 But in practice, SQL databases tend to refer to a whole set of legacy technologies, with several dominant vendors, optimized for a historical set of applications. Many of these existing commercial databases are a mismatch to the requirements that Hadoop targets. With that in mind, let’s make a more detailed comparison of Hadoop with typical SQL databases on specific dimensions. SCALE-OUT INSTEAD OF SCALE-UP Scaling commercial relational databases is expensive. Their design is more friendly to scaling up. To run a bigger database you need to buy a bigger machine. In fact, it’s not unusual to see server vendors market their expensive high-end machines as “database-class servers.” Unfortunately, at some point there won’t be a big enough machine available for the larger data sets. More importantly, the high-end machines are not cost effective for many applications. For example, a machine with four times the power of a standard PC costs a lot more than putting four such PCs in a cluster. Hadoop is designed to be a scale-out architecture operating on a cluster of com- modity PC machines. Adding more resources means adding more machines to the Hadoop cluster. Hadoop clusters with ten to hundreds of machines is standard. In fact, other than for development purposes, there’s no reason to run Hadoop on a single server. KEY/VALUE PAIRS INSTEAD OF RELATIONAL TABLES A fundamental tenet of relational databases is that data resides in tables having rela- tional structure defined by a schema . Although the relational model has great formal properties, many modern applications deal with data types that don’t fit well into this model. Text documents, images, and XML files are popular examples. Also, large data sets are often unstructured or semistructured. Hadoop uses key/value pairs as its ba- sic data unit, which is flexible enough to work with the less-structured data types. In Hadoop, data can originate in any form, but it eventually transforms into (key/value) pairs for the processing functions to work on. FUNCTIONAL PROGRAMMING (MAPREDUCE) INSTEAD OF DECLARATIVE QUERIES (SQL) SQL is fundamentally a high-level declarative language. You query data by stating the result you want and let the database engine figure out how to derive it. Under MapReduce you 3 This is in fact a hot area within the Hadoop community, and we’ll cover some of the leading projects in chapter 11. 8 CHAPTER 1 Introducing Hadoop specify the actual steps in processing the data, which is more analogous to an execution plan for a SQL engine . Under SQL you have query statements; under MapReduce you have scripts and codes. MapReduce allows you to process data in a more general fashion than SQL queries. For example, you can build complex statistical models from your data or reformat your image data. SQL is not well designed for such tasks. On the other hand, when working with data that do fit well into relational structures, some people may find MapReduce less natural to use. Those who are accustomed to the SQL paradigm may find it challenging to think in the MapReduce way. I hope the exercises and the examples in this book will help make MapReduce programming more intuitive. But note that many extensions are available to allow one to take advantage of the scalability of Hadoop while programming in more familiar paradigms. In fact, some enable you to write queries in a SQL-like language, and your query is automatically compiled into MapReduce code for execution. We’ll cover some of these tools in chapters 10 and 11. OFFLINE BATCH PROCESSING INSTEAD OF ONLINE TRANSACTIONS Hadoop is designed for offline processing and analysis of large-scale data. It doesn’t work for random reading and writing of a few records, which is the type of load for online transaction processing. In fact, as of this writing (and in the foreseeable future), Hadoop is best used as a write-once , read-many-times type of data store. In this aspect it’s similar to data warehouses in the SQL world. You have seen how Hadoop relates to distributed systems and SQL databases at a high level. Let’s learn how to program in it. For that, we need to understand Hadoop’s MapReduce paradigm. 1.5 Understanding MapReduce You’re probably aware of data processing models such as pipelines and message queues . These models provide specific capabilities in developing different aspects of data processing applications. The most familiar pipelines are the Unix pipes . Pipelines can help the reuse of processing primitives; simple chaining of existing modules cre- ates new ones. Message queues can help the synchronization of processing primitives . The programmer writes her data processing task as processing primitives in the form of either a producer or a consumer. The timing of their execution is managed by the system. Similarly, MapReduce is also a data processing model . Its greatest advantage is the easy scaling of data processing over multiple computing nodes. Under the MapReduce model, the data processing primitives are called mappers and reducers . Decomposing a data processing application into mappers and reducers is sometimes nontrivial. But, once you write an application in the MapReduce form, scaling the application to run over hundreds, thousands, or even tens of thousands of machines in a cluster is merely a configuration change. This simple scalability is what has attracted many programmers to the MapReduce model. Understanding MapReduce 9 1.5.1 Scaling a simple program manually Before going through a formal treatment of MapReduce, let’s go through an exercise of scaling a simple program to process a large data set. You’ll see the challenges of scaling a data processing program and will better appreciate the benefits of using a framework such as MapReduce to handle the tedious chores for you. Our exercise is to count the number of times each word occurs in a set of documents. In this example, we have a set of documents having only one document with only one sentence: Do as I say, not as I do. We derive the word counts shown to the right. We’ll call this particular exercise word counting . When the set of documents is small, a straightforward program will do the job. Let’s write one here in pseudo-code : define wordCount as Multiset; for each document in documentSet { T = tokenize(document); for each token in T { wordCount[token]++; } } display(wordCount); The program loops through all the documents. For each document, the words are extracted one by one using a tokenization process. For each word, its corresponding entry in a multiset called wordCount is incremented by one. At the end, a display() function prints out all the entries in wordCount. Many ways to say MapReduce Even though much has been written about MapReduce, one does not find the name itself written the same everywhere. The original Google paper and the Wikipedia entry use the CamelCase version MapReduce. However, Google itself has used Map Reduce in some pages on its website (for example, http://research.google.com/ roundtable/MR.html). At the official Hadoop documentation site, one can find links pointing to a Map-Reduce Tutorial. Clicking on the link brings one to a Hadoop Map/Reduce Tutorial (http://hadoop.apache.org/core/docs/current/mapred_ tutorial.html) explaining the Map/Reduce framework. Writing variations also exist for the different Hadoop components such as NameNode (name node, name- node, and namenode), DataNode, JobTracker, and TaskTracker. For the sake of consistency, we’ll go with CamelCase for all those terms in this book. (That is, we will use MapReduce, NameNode, DataNode, JobTracker, and TaskTracker.) Word Count as 2 do 2 i 2 not 1 say 1 10 CHAPTER 1 Introducing Hadoop NOTE A multiset is a set where each element also has a count. The word count we’re trying to generate is a canonical example of a multiset. In practice, it’s usually implemented as a hash table . This program works fine until the set of documents you want to process becomes large. For example, you want to build a spam filter to know the words frequently used in the millions of spam emails you’ve received. Looping through all the documents using a single computer will be extremely time consuming. You speed it up by rewriting the program so that it distributes the work over several machines. Each machine will pro- cess a distinct fraction of the documents. When all the machines have completed this, a second phase of processing will combine the result of all the machines. The pseudo- code for the first phase, to be distributed over many machines, is define wordCount as Multiset; for each document in documentSubset { T = tokenize(document); for each token in T { wordCount[token]++; } } sendToSecondPhase(wordCount); And the pseudo-code for the second phase is define totalWordCount as Multiset; for each wordCount received from firstPhase { multisetAdd (totalWordCount, wordCount); } That wasn’t too hard, right? But a few details may prevent it from working as expected. First of all, we ignore the performance requirement of reading in the documents. If the documents are all stored in one central storage server, then the bottleneck is in the bandwidth of that server. Having more machines for processing only helps up to a certain point—until the storage server can’t keep up. You’ll also need to split up the documents among the set of processing machines such that each machine will process only those documents that are stored in it. This will remove the bottleneck of a central storage server. This reiterates the point made earlier about storage and processing hav- ing to be tightly coupled in data-intensive distributed applications. Another flaw with the program is that wordCount (and totalWordCount) are stored in memory. When processing large document sets, the number of unique words can exceed the RAM storage of a machine. The English language has about one million words, a size that fits comfortably into an iPod, but our word counting program will deal with many unique words not found in any standard English dictionary. For example, we must deal with unique names such as Hadoop. We have to count misspellings even if they are not real words (for example, exampel), and we count all different forms of a word separately (for example, eat, ate, eaten, and eating ). Even if the number of unique words in the document set is manageable in memory, a slight change in the problem definition can explode the space complexity. For example, instead of words in documents, we may want to count IP addresses in a log file, or the frequency of bigrams. In the case of the latter, we’ll work with a multiset with billions of entries, which exceeds the RAM storage of most commodity computers. NOTE A bigram is a pair of consecutive words. The sentence “Do as I say, not as I do” can be broken into the following bigrams: Do as, as I, I say, say not, not as, as I, I do. Analogously, trigrams are groups of three consecutive words. Both bigrams and trigrams are important in natural language processing . wordCount may not fit in memory; we’ll have to rewrite our program to store this hash table on disk. This means we’ll implement a disk-based hash table, which involves a substantial amount of coding. Furthermore, remember that phase two has only one machine, which will process wordCount sent from all the machines in phase one. Processing one wordCount is itself quite unwieldy. After we have added enough machines to phase one processing, the single machine in phase two will become the bottleneck. The obvious question is, can we rewrite phase two in a distributed fashion so that it can scale by adding more machines? The answer is, yes. To make phase two work in a distributed fashion, you must somehow divide its work among multiple machines such that they can run independently. You need to partition wordCount after phase one such that each machine in phase two only has to handle one partition. In one example, let’s say we have 26 machines for phase two. We assign each machine to only handle wordCount for words beginning with a particular letter in the alphabet. For example, machine A in phase two will only handle word counting for words beginning with the letter a. To enable this partitioning in phase two, we need a slight modification in phase one. Instead of a single disk-based hash table for wordCount, we will need 26 of them: wordCount-a, wordCount-b, and so on. Each one counts words starting with a particular letter. After phase one, wordCount-a from each of the phase one machines will be sent to machine A of phase two, all the wordCount-b’s will be sent to machine B, and so on. Each machine in phase one will shuffle its results among the machines in phase two. Looking back, this word counting program is getting complicated. To make it work across a cluster of distributed machines, we find that we need to add a number of functionalities: ■ Store files over many processing machines (of phase one). ■ Write a disk-based hash table permitting processing without being limited by RAM capacity. ■ Shuffle the partitions to the appropriate machines in phase two. Partition the intermediate data (that is, wordCount) from phase one. ■ This is a lot of work for something as simple as word counting, and we haven’t even touched upon issues like fault tolerance. (What if a machine fails in the middle of its task?) This is the reason why you would want a framework like Hadoop. When you Understanding MapReduce 11 12 CHAPTER 1 Introducing Hadoop write your application in the MapReduce model, Hadoop will take care of all that scalability “plumbing” for you. 1.5.2 Scaling the same program in MapReduce MapReduce programs are executed in two main phases, called mapping and reducing . Each phase is defined by a data processing function, and these functions are called mapper and reducer, respectively. In the mapping phase, MapReduce takes the input data and feeds each data element to the mapper. In the reducing phase, the reducer processes all the outputs from the mapper and arrives at a final result. In simple terms, the mapper is meant to filter and transform the input into something that the reducer can aggregate over. You may see a striking similarity here with the two phases we had to develop in scaling up word counting. The similarity is not accidental. The MapReduce framework was designed after a lot of experience in writing scalable, distributed programs. This two-phase design pattern was seen in scaling many programs, and became the basis of the framework. In scaling our distributed word counting program in the last section, we also had to write the partitioning and shuffling functions. Partitioning and shuffling are common design patterns that go along with mapping and reducing. Unlike mapping and reducing, though, partitioning and shuffling are generic functionalities that are not too dependent on the particular data processing application. The MapReduce framework provides a default implementation that works in most situations. In order for mapping, reducing, partitioning, and shuffling (and a few others we haven’t mentioned) to seamlessly work together, we need to agree on a common structure for the data being processed. It should be flexible and powerful enough to handle most of the targeted data processing applications. MapReduce uses lists and (key/value) pairs as its main data primitives. The keys and values are often integers or strings but can also be dummy values to be ignored or complex object types. The map and reduce functions must obey the following constraint on the types of keys and values. In the MapReduce framework you write applications by specifying the mapper and reducer. Let’s look at the complete data flow: 1 The input to your application must be structured as a list of (key/value) pairs , list(). This input format may seem open-ended but is often quite simple in practice. The input format for processing multiple files is usually list(). The input format for processing one large file, such as a log file, is list(). Input Output map list() reduce list() 2 The list of (key/value) pairs is broken up and each individual (key/value) pair, , is processed by calling the map function of the mapper. In practice, the key k1 is often ignored by the mapper. The mapper transforms each pair into a list of pairs. The details of this transformation largely determine what the MapReduce program does. Note that the (key/value) pairs are processed in arbitrary order. The transformation must be self-contained in that its output is dependent only on one single (key/value) pair. For word counting, our mapper takes and promptly ignores filename. It can output a list of but can be even simpler. As we know the counts will be aggregated in a later stage, we can output a list of with repeated entries and let the complete aggregation be done later. That is, in the output list we can have the (key/value) pair <"foo", 3> once or we can have the pair <"foo", 1> three times. As we’ll see, the latter approach is much easier to program. The former approach may have some performance benefits, but let’s leave such optimization alone until we have fully grasped the MapReduce framework. 3 The output of all the mappers are (conceptually) aggregated into one giant list of pairs. All pairs sharing the same k2 are grouped together into a new (key/value) pair, . The framework asks the reducer to process each one of these aggregated (key/value) pairs individually. Following our word counting example, the map output for one document may be a list with pair <"foo", 1> three times, and the map output for another document may be a list with pair <"foo", 1> twice. The aggregated pair the reducer will see is <"foo", list(1,1,1,1,1)>. In word counting, the output of our reducer is <"foo", 5>, which is the total number of times “foo” has occurred in our document set. Each reducer works on a different word. The MapReduce framework automatically collects all the pairs and writes them to file(s). Note that for the word counting example, the data types k2 and k3 are the same and v2 and v3 are also the same. This will not always be the case for other data processing applications. Let’s rewrite the word counting program in MapReduce to see how all this fits together Listing 1.1 shows the pseudo-code. Listing 1.1 Pseudo-code for map and reduce functions for word counting map(String filename, String document) { List T = tokenize(document); for each token in T { emit ((String)token, (Integer) 1); } } reduce(String token, List values) { Integer sum = 0; Understanding MapReduce 13 14 CHAPTER 1 Introducing Hadoop for each value in values { sum = sum + value; } emit ((String)token, (Integer) sum); } We’ve said before that the output of both map and reduce function are lists. As you can see from the pseudo-code, in practice we use a special function in the framework called emit() to generate the elements in the list one at a time. This emit() function further relieves the programmer from managing a large list. The code looks similar to what we have in section 1.5.1, except this time it will actually work at scale. Hadoop makes building scalable distributed programs easy, doesn’t it? Now let’s turn this pseudo-code into a Hadoop program. 1.6 Counting words with Hadoop—running your first program Now that you know what the Hadoop and MapReduce framework is about, let’s get it running. In this chapter, we’ll run Hadoop only on a single machine, which can be your desktop or laptop computer. The next chapter will show you how to run Hadoop over a cluster of machines, which is what you’d want for practical deployment. Run- ning Hadoop on a single machine is mainly useful for development work. Linux is the official development and production platform for Hadoop, although Windows is a supported development platform as well. For a Windows box, you’ll need to install cygwin (http://www-cygwin.com/) to enable shell and Unix scripts. NOTE Many people have reported success in running Hadoop in development mode on other variants of Unix, such as Solaris and Mac OS X . In fact, MacBook Pro seems to be the laptop of choice among Hadoop developers, as they’re ubiquitous in Hadoop conferences and user group meetings. Running Hadoop requires Java (version 1.6 or higher). Mac users should get it from Apple. You can download the latest JDK for other operating systems from Sun at http://java.sun.com/javase/downloads/index.jsp. Install it and remember the root of the Java installation, which we’ll need later. To install Hadoop, first get the latest stable release at http://hadoop.apache.org/ core/releases.html. After you unpack the distribution, edit the script conf/hadoop- env.sh to set JAVA_HOME to the root of the Java installation you have remembered from earlier. For example, in Mac OS X, you’ll replace this line # export JAVA_HOME=/usr/lib/j2sdk1.5-sun with this line export JAVA_HOME=/Library/Java/Home You’ll be using the Hadoop script quite often. Let’s run it without any arguments to see its usage documentation: Counting words with Hadoop—running your first program 15 bin/hadoop We get Usage: hadoop [--config confdir] COMMAND where COMMAND is one of: namenode -format format the DFS filesystem secondarynamenode run the DFS secondary namenode namenode run the DFS namenode datanode run a DFS datanode dfsadmin run a DFS admin client fsck run a DFS filesystem checking utility fs run a generic filesystem user client balancer run a cluster balancing utility jobtracker run the MapReduce job Tracker node pipes run a Pipes job tasktracker run a MapReduce task Tracker node job manipulate MapReduce jobs version print the version jar run a jar file distcp copy file or directories recursively archive -archiveName NAME * create a hadoop archive daemonlog get/set the log level for each daemon or CLASSNAME run the class named CLASSNAME Most commands print help when invoked w/o parameters. We’ll cover the various Hadoop commands in the course of this book. For our current purpose, we only need to know that the command to run a ( Java) Hadoop program is bin/hadoop jar . As the command implies, Hadoop programs written in Java are packaged in jar files for execution. Fortunately for us, we don’t need to write a Hadoop program first; the default installation already has several sample programs we can use. The following command shows what is available in the examples jar file: bin/hadoop jar hadoop-*-examples.jar You’ll see about a dozen example programs prepackaged with Hadoop, and one of them is a word counting program called... wordcount! The important (inner) classes of that program are shown in listing 1.2. We’ll see how this Java program implements the word counting map and reduce functions we had in pseudo-code in listing 1.1. We’ll modify this program to understand how to vary its behavior. For now we’ll assume it works as expected and only follow the mechanics of executing a Hadoop program. Without specifying any arguments, executing wordcount will show its usage information: bin/hadoop jar hadoop-*-examples.jar wordcount which shows the arguments list: wordcount [-m ] [-r ] 16 CHAPTER 1 Introducing Hadoop The only parameters are an input directory () of text documents you want to analyze and an output directory () where the program will dump its output. To execute wordcount, we need to first create an input directory: mkdir input and put some documents in it. You can add any text document to the directory. For illustration, let’s put the text version of the 2002 State of the Union address, obtained from http://www.gpoaccess.gov/sou/. We now analyze its word counts and see the results: bin/hadoop jar hadoop-*-examples.jar wordcount input output more output/* You’ll see a word count of every word used in the document, listed in alphabetical or- der. This is not bad considering you have not written a single line of code yet! But, also note a number of shortcomings in the included wordcount program. Tokenization is based purely on whitespace characters and not punctuation marks, making States, States., and States: separate words. The same is true for capitalization, where States and states appear as separate words. Furthermore, we would like to leave out words that show up in the document only once or twice. Fortunately, the source code for wordcount is available and included in the installation at src/examples/org/apache/hadoop/examples/WordCount.java. We can modify it as per our requirements. Let’s first set up a directory structure for our playground and make a copy of the program. mkdir playground mkdir playground/src mkdir playground/classes cp src/examples/org/apache/hadoop/examples/WordCount.java ➥ playground/src/WordCount.java Before we make changes to the program, let’s go through compiling and executing this new copy in the Hadoop framework. javac -classpath hadoop-*-core.jar -d playground/classes ➥ playground/src/WordCount.java jar -cvf playground/wordcount.jar -C playground/classes/ . You’ll have to remove the output directory each time you run this Hadoop command, because it is created automatically. bin/hadoop jar playground/wordcount.jar ➥ org.apache.hadoop.examples.WordCount input output Look at the files in your output directory again. As we haven’t changed any program code, the result should be the same as before. We’ve only compiled our own copy rather than running the precompiled version. Now we are ready to modify WordCount to add some extra features. Listing 1.2 is a partial view of the WordCount.java program. Comments and supporting code are stripped out. Listing 1.2 WordCount.java public class WordCount extends Configured implements Tool { public static class MapClass extends MapReduceBase implements Mapper { private final static IntWritable one = new IntWritable(1); private Text word = new Text(); public void map(LongWritable key, Text value, OutputCollector output, Reporter reporter) throws IOException { String line = value.toString(); StringTokenizer itr = new StringTokenizer(line); q while (itr.hasMoreTokens()) { word.set(itr.nextToken()); w output.collect(word, one); } } } public static class Reduce extends MapReduceBase implements Reducer { public void reduce(Text key, Iterator values, OutputCollector output, Reporter reporter) throws IOException { int sum = 0; while (values.hasNext()) { sum += values.next().get(); } output.collect(key, new IntWritable(sum)); e } } ... } The main functional distinction between WordCount.java and our MapReduce pseudo- code is that in WordCount.java, map() processes one line of text at a time whereas our pseudo-code processes a document at a time. This distinction may not even be apparent from looking at WordCount.java as it’s Hadoop’s default configuration. The code in listing 1.2 is virtually identical to our pseudo-code in listing 1.1 though the Java syntax makes it more verbose. The map and reduce functions are inside inner classes of WordCount. You may notice we use special classes such as LongWritable , IntWritable , and Text instead of the more familiar Long, Integer, and String classes of Java. Consider these implementation details for now. The new classes have additional serialization capabilities needed by Hadoop’s internal. The changes we want to make to the program are easy to spot. We see q that WordCount uses Java’s StringTokenizer in its default setting, which tokenizes based only on whitespaces. To ignore standard punctuation marks, we add them to the StringTokenizer’s list of delimiter characters: StringTokenizer itr = new StringTokenizer(line, " \t\n\r\f,.:;?![]'"); Tokenize using white spaces Cast token into Text object Output count of each token Counting words with Hadoop—running your first program 17 18 CHAPTER 1 Introducing Hadoop When looping through the set of tokens, each token is extracted and cast into a Text object w. (Again, in Hadoop, the special class Text is used in place of String.) We want the word count to ignore capitalization, so we lowercase all the words before turn- ing them into Text objects. word.set(itr.nextToken().toLowerCase()); Finally, we want only words that appear more than four times. We modify e to collect the word count into the output only if that condition is met. (This is Hadoop’s equiva- lent of the emit() function in our pseudo-code.) if (sum > 4) output.collect(key, new IntWritable(sum)); After making changes to those three lines, you can recompile the program and ex- ecute it again. The results are shown in table 1.1. Table 1.1 Words with a count higher than 4 in the 2002 State of the Union Address 11th (5) citizens (9) its (6) over (6) to (123) a (69) congress (10) jobs (11) own (5) together (5) about (5) corps (6) join (7) page (7) tonight (5) act (7) country (10) know (6) people (12) training (5) afghanistan (10) destruction (5) last (6) protect (5) united (6) all (10) do (6) lives (6) regime (5) us (6) allies (8) every (8) long (5) regimes (6) want (5) also (5) evil (5) make (7) security (19) war (12) America (33) for (27) many (5) september (5) was (11) American (15) free (6) more (11) so (12) we (76) americans (8) freedom (10) most (5) some (6) we’ve (5) an (7) from (15) must (18) states (9) weapons (12) and (210) good (13) my (13) tax (7) were (7) are (17) great (8) nation (11) terror (13) while (5) as (18) has (12) need (7) terrorist (12) who (18) ask (5) have (32) never (7) terrorists (10) will (49) at (16) health (5) new (13) than (6) with (22) be (23) help (7) no (7) that (29) women (5) been (8) home (5) not (15) the (184) work (7) best (6) homeland (7) now (10) their (17) workers (5) budget (7) hope (5) of (130) them (8) world (17) but (7) i (29) on (32) these (18) would (5) by (13) if (8) one (5) they (12) yet (8) History of Hadoop 19 Table 1.1 Words with a count higher than 4 in the 2002 State of the Union Address (continued) camps (8) in (79) opportunity (5) this (28) you (12) can (7) is (44) or (8) thousands (5) children (6) it (21) our (78) time (7) We see that 128 words have a frequency count greater than 4. Many of these words appear frequently in almost any English text. For example, there is a (69), and (210), i (29), in (79), the (184) and many others. We also see words that summarize the issues facing the United States at that time: terror (13), terrorist (12), terrorists (10), security (19), weapons (12), destruction (5), afghanistan (10), freedom (10), jobs (11), budget (7), and many others. 1.7 History of Hadoop Hadoop started out as a subproject of Nutch , which in turn was a subproject of Apache Lucene . Doug Cutting founded all three projects, and each project was a logical pro- gression of the previous one. Lucene is a full-featured text indexing and searching library. Given a text collection, a developer can easily add search capability to the documents using the Lucene engine . Desktop search, enterprise search, and many domain-specific search engines have been built using Lucene. Nutch is the most ambitious extension of Lucene. It tries to build a complete web search engine using Lucene as its core component. Nutch has parsers for HTML, a web crawler, a link-graph database, and other extra components necessary for a web search engine. Doug Cutting envisions Nutch to be an open democratic alternative to the proprietary technologies in commercial offerings such as Google. Besides having added components like a crawler and a parser, a web search engine differs from a basic document search engine in terms of scale. Whereas Lucene is targeted at indexing millions of documents, Nutch should be able to handle billions of web pages without becoming exorbitantly expensive to operate. Nutch will have to run on a distributed cluster of commodity hardware. The challenge for the Nutch team is to address scalability issues in software. Nutch needs a layer to handle distributed processing, redundancy, automatic failover, and load balancing. These challenges are by no means trivial. Around 2004, Google published two papers describing the Google File System (GFS) and the MapReduce framework . Google claimed to use these two technologies for scaling its own search system. Doug Cutting immediately saw the applicability of these technologies to Nutch, and his team implemented the new framework and ported Nutch to it. The new implementation immediately boosted Nutch’s scalability. It started to handle several hundred million web pages and could run on clusters of dozens of nodes. Doug realized that a dedicated project to flesh out the two technologies was needed to get to web scale, and Hadoop was born. Yahoo! hired Doug in January 20 CHAPTER 1 Introducing Hadoop 2006 to work with a dedicated team on improving Hadoop as an open source project. Two years later, Hadoop achieved the status of an Apache Top Level Project . Later, on February 19, 2008, Yahoo! announced that Hadoop running on a 10,000+ core Linux cluster was its production system for indexing the Web (http://developer.yahoo. net/blogs/hadoop/2008/02/yahoo-worlds-largest-production-hadoop.html). Hadoop had truly hit web scale! What’s up with the names? When naming software projects, Doug Cutting seems to have been inspired by his family. Lucene is his wife’s middle name, and her maternal grandmother’s first name. His son, as a toddler, used Nutch as the all-purpose word for meal and later named a yellow stuffed elephant Hadoop. Doug said he “was looking for a name that wasn’t already a web domain and wasn’t trademarked, so I tried various words that were in my life but not used by anybody else. Kids are pretty good at making up words.” 1.8 Summary Hadoop is a versatile tool that allows new users to access the power of distributed com- puting. By using distributed storage and transferring code instead of data, Hadoop avoids the costly transmission step when working with large data sets. Moreover, the redundancy of data allows Hadoop to recover should a single node fail. You have seen the ease of creating programs with Hadoop using the MapReduce framework. What is equally important is what you didn’t have to do—worry about partitioning the data, determining which nodes will perform which tasks, or handling communication be- tween nodes. Hadoop handles this for you, leaving you free to focus on what’s most important to you—your data and what you want to do with it. In the next chapter we’ll go into further details about the internals of Hadoop and setting up a working Hadoop cluster. 1.9 Resources The official Hadoop website is at http://hadoop.apache.org/. The original papers on the Google File System and MapReduce are well worth reading. Appreciate their underlying design and architecture: ■ The Google File System—http://labs.google.com/papers/gfs.html ■ MapReduce: Simplified Data Processing on Large Clusters—http://labs.google.com/ papers/mapreduce.html 2 21 Starting Hadoop This chapter covers ■ The architectural components of Hadoop ■ Setting up Hadoop and its three operating modes: standalone, pseudo-distributed, and fully distributed Web-based tools to monitor your Hadoop setup■ This chapter will serve as a roadmap to guide you through setting up Hadoop. If you work in an environment where someone else sets up the Hadoop cluster for you, you may want to skim through this chapter. You’ll want to understand enough to set up your personal development machine, but you can skip through the details of configuring the communication and coordination of various nodes. After discussing the physical components of Hadoop in section 2.1, we’ll progress to setting up your cluster in sections 2.2. and 2.3. Section 2.3 will focus on the three operational modes of Hadoop and how to set them up. You’ll read about web-based tools that assist monitoring your cluster in section 2.4. 2.1 The building blocks of Hadoop We’ve discussed the concepts of distributed storage and distributed computation in the previous chapter. Now let’s see how Hadoop implements those ideas. On 22 CHAPTER 2 Starting Hadoop a fully configured cluster, “running Hadoop” means running a set of daemons, or resident programs, on the different servers in your network. These daemons have specific roles; some exist only on one server, some exist across multiple servers. The daemons include NameNode■ DataNode■ Secondary NameNode■ JobTracker■ TaskTracker■ We’ll discuss each one and its role within Hadoop. 2.1.1 NameNode Let’s begin with arguably the most vital of the Hadoop daemons—the NameNode . Hadoop employs a master/slave architecture for both distributed storage and distrib- uted computation. The distributed storage system is called the Hadoop File System , or HDFS. The NameNode is the master of HDFS that directs the slave DataNode daemons to perform the low-level I/O tasks. The NameNode is the bookkeeper of HDFS; it keeps track of how your files are broken down into file blocks, which nodes store those blocks, and the overall health of the distributed filesystem. The function of the NameNode is memory and I/O intensive. As such, the server hosting the NameNode typically doesn’t store any user data or perform any computations for a MapReduce program to lower the workload on the machine. This means that the NameNode server doesn’t double as a DataNode or a TaskTracker. There is unfortunately a negative aspect to the importance of the NameNode—it’s a single point of failure of your Hadoop cluster. For any of the other daemons, if their host nodes fail for software or hardware reasons, the Hadoop cluster will likely continue to function smoothly or you can quickly restart it. Not so for the NameNode. 2.1.2 DataNode Each slave machine in your cluster will host a DataNode daemon to perform the grunt work of the distributed filesystem—reading and writing HDFS blocks to actual files on the local filesystem. When you want to read or write a HDFS file, the file is broken into blocks and the NameNode will tell your client which DataNode each block resides in. Your client communicates directly with the DataNode daemons to process the local files corresponding to the blocks. Furthermore, a DataNode may communicate with other DataNodes to replicate its data blocks for redundancy. Figure 2.1 illustrates the roles of the NameNode and DataNodes. In this figure, we show two data files, one at /user/chuck/data1 and another at /user/james/data2. The data1 file takes up three blocks, which we denote 1, 2, and 3, and the data2 file consists of blocks 4 and 5. The content of the files are distributed among the DataNodes. In The building blocks of Hadoop 23 3 5 4 2 3 5 1 5 3 2 4 1 1 4 2 File metadata: /user/chuck/data1 -> 1,2,3 /user/james/data2 -> 4,5 NameNode DataNodes Figure 2.1 NameNode /DataNode interaction in HDFS. The NameNode keeps track of the file metadata—which files are in the system and how each file is broken down into blocks. The DataNodes provide backup store of the blocks and constantly report to the NameNode to keep the metadata current. this illustration, each block has three replicas. For example, block 1 (used for data1) is replicated over the three rightmost DataNodes. This ensures that if any one DataNode crashes or becomes inaccessible over the network, you’ll still be able to read the files. DataNodes are constantly reporting to the NameNode. Upon initialization, each of the DataNodes informs the NameNode of the blocks it’s currently storing. After this mapping is complete, the DataNodes continually poll the NameNode to provide information regarding local changes as well as receive instructions to create, move, or delete blocks from the local disk. 2.1.3 Secondary NameNode The Secondary NameNode (SNN) is an assistant daemon for monitoring the state of the cluster HDFS. Like the NameNode, each cluster has one SNN, and it typically resides on its own machine as well. No other DataNode or TaskTracker daemons run on the same server. The SNN differs from the NameNode in that this process doesn’t receive or record any real-time changes to HDFS. Instead, it communicates with the NameNode to take snapshots of the HDFS metadata at intervals defined by the cluster configuration. As mentioned earlier, the NameNode is a single point of failure for a Hadoop cluster, and the SNN snapshots help minimize the downtime and loss of data. Nevertheless, a NameNode failure requires human intervention to reconfigure the cluster to use the SNN as the primary NameNode. We’ll discuss the recovery process in chapter 8 when we cover best practices for managing your cluster. 24 CHAPTER 2 Starting Hadoop 2.1.4 JobTracker The JobTracker daemon is the liaison between your application and Hadoop. Once you submit your code to your cluster, the JobTracker determines the execution plan by determining which files to process, assigns nodes to different tasks, and monitors all tasks as they’re running. Should a task fail, the JobTracker will automatically relaunch the task, possibly on a different node, up to a predefined limit of retries. There is only one JobTracker daemon per Hadoop cluster. It’s typically run on a server as a master node of the cluster. 2.1.5 TaskTracker As with the storage daemons, the computing daemons also follow a master/slave archi- tecture: the JobTracker is the master overseeing the overall execution of a MapReduce job and the TaskTrackers manage the execution of individual tasks on each slave node. Figure 2.2 illustrates this interaction. Each TaskTracker is responsible for executing the individual tasks that the JobTracker assigns. Although there is a single TaskTracker per slave node, each TaskTracker can spawn multiple JVMs to handle many map or reduce tasks in parallel. One responsibility of the TaskTracker is to constantly communicate with the JobTracker. If the JobTracker fails to receive a heartbeat from a TaskTracker within a specified amount of time, it will assume the TaskTracker has crashed and will resubmit the corresponding tasks to other nodes in the cluster. JobTracker Client TaskTracker Reduce Map TaskTracker Reduce Map TaskTracker Reduce Map TaskTracker Reduce Map Figure 2.2 JobTracker and TaskTracker interaction. After a client calls the JobTracker to begin a data processing job, the JobTracker partitions the work and assigns different map and reduce tasks to each TaskTracker in the cluster. Setting up SSH for a Hadoop cluster 25 JobTracker NameNode DataNode TaskTracker DataNode TaskTracker DataNode TaskTracker DataNode TaskTracker Secondary NameNode Figure 2.3 Topology of a typical Hadoop cluster . It’s a master/slave architecture in which the NameNode and JobTracker are masters and the DataNodes and TaskTrackers are slaves. Having covered each of the Hadoop daemons, we depict the topology of one typical Hadoop cluster in figure 2.3. This topology features a master node running the NameNode and JobTracker daemons and a standalone node with the SNN in case the master node fails. For small clusters, the SNN can reside on one of the slave nodes. On the other hand, for large clusters, separate the NameNode and JobTracker on two machines. The slave machines each host a DataNode and TaskTracker, for running tasks on the same node where their data is stored. We’ll work toward setting up a complete Hadoop cluster of this form by first establishing the master node and the control channels between nodes. If a Hadoop cluster is already available to you, you can skip the next section on how to set up Secure Shell (SSH) channels between nodes. You also have a couple of options to run Hadoop using only a single machine, in what are known as standalone and pseudo-distributed modes. They’re useful for development. Configuring Hadoop to run in these two modes or the standard cluster setup (fully distributed mode) is covered in section 2.3. 2.2 Setting up SSH for a Hadoop cluster When setting up a Hadoop cluster , you’ll need to designate one specific node as the master node. As shown in figure 2.3, this server will typically host the NameNode and 26 CHAPTER 2 Starting Hadoop JobTracker daemons. It’ll also serve as the base station contacting and activating the DataNode and TaskTracker daemons on all of the slave nodes. As such, we need to define a means for the master node to remotely access every node in your cluster. Hadoop uses passphraseless SSH for this purpose. SSH utilizes standard public key cryptography to create a pair of keys for user verification—one public, one private. The public key is stored locally on every node in the cluster, and the master node sends the private key when attempting to access a remote machine. With both pieces of information, the target machine can validate the login attempt. 2.2.1 Define a common account We’ve been speaking in general terms of one node accessing another; more precisely this access is from a user account on one node to another user account on the target machine. For Hadoop, the accounts should have the same username on all of the nodes (we use hadoop-user in this book), and for security purpose we recommend it being a user-level account. This account is only for managing your Hadoop cluster. Once the cluster daemons are up and running, you’ll be able to run your actual MapReduce jobs from other accounts. 2.2.2 Verify SSH installation The first step is to check whether SSH is installed on your nodes. We can easily do this by use of the "which" UNIX command: [hadoop-user@master]$ which ssh /usr/bin/ssh [hadoop-user@master]$ which sshd /usr/bin/sshd [hadoop-user@master]$ which ssh-keygen /usr/bin/ssh-keygen If you instead receive an error message such as this, /usr/bin/which: no ssh in (/usr/bin:/bin:/usr/sbin... install OpenSSH (www.openssh.com) via a Linux package manager or by downloading the source directly. (Better yet, have your system administrator do it for you.) 2.2.3 Generate SSH key pair Having verified that SSH is correctly installed on all nodes of the cluster, we use ssh- keygen on the master node to generate an RSA key pair . Be certain to avoid entering a passphrase, or you’ll have to manually enter that phrase every time the master node attempts to access another node. [hadoop-user@master]$ ssh-keygen -t rsa Generating public/private rsa key pair. Enter file in which to save the key (/home/hadoop-user/.ssh/id_rsa): Enter passphrase (empty for no passphrase): Enter same passphrase again: Running Hadoop 27 Your identification has been saved in /home/hadoop-user/.ssh/id_rsa. Your public key has been saved in /home/hadoop-user/.ssh/id_rsa.pub. After creating your key pair, your public key will be of the form [hadoop-user@master]$ more /home/hadoop-user/.ssh/id_rsa.pub ssh-rsa AAAAB3NzaC1yc2EAAAABIwAAAQEA1WS3RG8LrZH4zL2/1oYgkV1OmVclQ2OO5vRi0Nd K51Sy3wWpBVHx82F3x3ddoZQjBK3uvLMaDhXvncJG31JPfU7CTAfmtgINYv0kdUbDJq4TKG/fuO5q J9CqHV71thN2M310gcJ0Y9YCN6grmsiWb2iMcXpy2pqg8UM3ZKApyIPx99O1vREWm+4moFTg YwIl5be23ZCyxNjgZFWk5MRlT1p1TxB68jqNbPQtU7fIafS7Sasy7h4eyIy7cbLh8x0/V4/mcQsY 5dvReitNvFVte6onl8YdmnMpAh6nwCvog3UeWWJjVZTEBFkTZuV1i9HeYHxpm1wAzcnf7az78jT IRQ== hadoop-user@master and we next need to distribute this public key across your cluster. 2.2.4 Distribute public key and validate logins Albeit a bit tedious, you’ll next need to copy the public key to every slave node as well as the master node: [hadoop-user@master]$ scp ~/.ssh/id_rsa.pub hadoop-user@target:~/master_key Manually log in to the target node and set the master key as an authorized key (or append to the list of authorized keys if you have others defined). [hadoop-user@target]$ mkdir ~/.ssh [hadoop-user@target]$ chmod 700 ~/.ssh [hadoop-user@target]$ mv ~/master_key ~/.ssh/authorized_keys [hadoop-user@target]$ chmod 600 ~/.ssh/authorized_keys After generating the key, you can verify it’s correctly defined by attempting to log in to the target node from the master: [hadoop-user@master]$ ssh target The authenticity of host 'target (xxx.xxx.xxx.xxx)' can’t be established. RSA key fingerprint is 72:31:d8:1b:11:36:43:52:56:11:77:a4:ec:82:03:1d. Are you sure you want to continue connecting (yes/no)? yes Warning: Permanently added 'target' (RSA) to the list of known hosts. Last login: Sun Jan 4 15:32:22 2009 from master After confirming the authenticity of a target node to the master node, you won’t be prompted upon subsequent login attempts. [hadoop-user@master]$ ssh target Last login: Sun Jan 4 15:32:49 2009 from master We’ve now set the groundwork for running Hadoop on your own cluster. Let’s discuss the different Hadoop modes you might want to use for your projects. 2.3 Running Hadoop We need to configure a few things before running Hadoop. Let’s take a closer look at the Hadoop configuration directory : [hadoop-user@master]$ cd $HADOOP_HOME [hadoop-user@master]$ ls -l conf/ total 100 28 CHAPTER 2 Starting Hadoop -rw-rw-r-- 1 hadoop-user hadoop 2065 Dec 1 10:07 capacity-scheduler.xml -rw-rw-r-- 1 hadoop-user hadoop 535 Dec 1 10:07 configuration.xsl -rw-rw-r-- 1 hadoop-user hadoop 49456 Dec 1 10:07 hadoop-default.xml -rwxrwxr-x 1 hadoop-user hadoop 2314 Jan 8 17:01 hadoop-env.sh -rw-rw-r-- 1 hadoop-user hadoop 2234 Jan 2 15:29 hadoop-site.xml -rw-rw-r-- 1 hadoop-user hadoop 2815 Dec 1 10:07 log4j.properties -rw-rw-r-- 1 hadoop-user hadoop 28 Jan 2 15:29 masters -rw-rw-r-- 1 hadoop-user hadoop 84 Jan 2 15:29 slaves -rw-rw-r-- 1 hadoop-user hadoop 401 Dec 1 10:07 sslinfo.xml.example The first thing you need to do is to specify the location of Java on all the nodes includ- ing the master. In hadoop-env.sh define the JAVA_HOME environment variable to point to the Java installation directory. On our servers, we’ve it defined as export JAVA_HOME=/usr/share/jdk (If you followed the examples in chapter 1, you’ve already completed this step.) The hadoop-env.sh file contains other variables for defining your Hadoop environment, but JAVA_HOME is the only one requiring initial modification. The default settings on the other variables will probably work fine. As you become more familiar with Hadoop you can later modify this file to suit your individual needs (logging directory location, Java class path, and so on). The majority of Hadoop settings are contained in XML configuration files. Before version 0.20, these XML files are hadoop-default.xml and hadoop-site.xml . As the names imply, hadoop-default.xml contains the default Hadoop settings to be used unless they are explicitly overridden in hadoop-site.xml. In practice you only deal with hadoop-site.xml. In version 0.20 this file has been separated out into three XML files: core-site.xml , hdfs-site.xml , and mapred-site.xml . This refactoring better aligns the configuration settings to the subsystem of Hadoop that they control. In the rest of this chapter we’ll generally point out which of the three files used to adjust a configuration setting. If you use an earlier version of Hadoop, keep in mind that all such configuration settings are modified in hadoop-site.xml. In the following subsections we’ll provide further details about the different operational modes of Hadoop and example configuration files for each. 2.3.1 Local (standalone) mode The standalone mode is the default mode for Hadoop. When you first uncompress the Hadoop source package, it’s ignorant of your hardware setup. Hadoop chooses to be conservative and assumes a minimal configuration. All three XML files (or hadoop- site.xml before version 0.20) are empty under this default mode: Running Hadoop 29 With empty configuration files, Hadoop will run completely on the local machine. Because there’s no need to communicate with other nodes, the standalone mode doesn’t use HDFS, nor will it launch any of the Hadoop daemons. Its primary use is for developing and debugging the application logic of a MapReduce pro- gram without the additional complexity of interacting with the daemons. When you ran the example MapReduce program in chapter 1, you were running it in standalone mode. 2.3.2 Pseudo-distributed mode The pseudo-distributed mode is running Hadoop in a “cluster of one” with all daemons running on a single machine. This mode complements the standalone mode for debugging your code, allowing you to examine memory usage, HDFS input/out- put issues, and other daemon interactions. Listing 2.1 provides simple XML files to configure a single server in this mode. Listing 2.1 Example of the three configuration files for pseudo-distributed mode core-site.xml fs.default.name hdfs://localhost:9000 The name of the default file system. A URI whose scheme and authority determine the FileSystem implementation. mapred-site.xml mapred.job.tracker localhost:9001 The host and port that the MapReduce job tracker runs at. hdfs-site.xml 30 CHAPTER 2 Starting Hadoop dfs.replication 1 The actual number of replications can be specified when the file is created. In core-site.xml and mapred-site.xml we specify the hostname and port of the NameNode and the JobTracker, respectively. In hdfs-site.xml we specify the default replication factor for HDFS, which should only be one because we’re running on only one node. We must also specify the location of the Secondary NameNode in the mas- ters file and the slave nodes in the slaves file: [hadoop-user@master]$ cat masters localhost [hadoop-user@master]$ cat slaves localhost While all the daemons are running on the same machine, they still communicate with each other using the same SSH protocol as if they were distributed over a cluster. Section 2.2 has a more detailed discussion of setting up the SSH channels, but for single-node operation simply check to see if your machine already allows you to ssh back to itself. [hadoop-user@master]$ ssh localhost If it does, then you’re good. Otherwise setting up takes two lines. [hadoop-user@master]$ ssh-keygen -t dsa -P '' -f ~/.ssh/id_dsa [hadoop-user@master]$ cat ~/.ssh/id_dsa.pub >> ~/.ssh/authorized_keys You are almost ready to start Hadoop. But first you’ll need to format your HDFS by using the command [hadoop-user@master]$ bin/hadoop namenode -format We can now launch the daemons by use of the start-all.sh script. The Java jps command will list all daemons to verify the setup was successful. [hadoop-user@master]$ bin/start-all.sh [hadoop-user@master]$ jps 26893 Jps 26832 TaskTracker 26620 SecondaryNameNode 26333 NameNode 26484 DataNode 26703 JobTracker Running Hadoop 31 When you’ve finished with Hadoop you can shut down the Hadoop daemons by the command [hadoop-user@master]$ bin/stop-all.sh Both standalone and pseudo-distributed modes are for development and debug- ging purposes. An actual Hadoop cluster runs in the third mode, the fully distrib- uted mode. 2.3.3 Fully distributed mode After continually emphasizing the benefits of distributed storage and distributed computation, it’s time for us to set up a full cluster. In the discussion below we’ll use the following server names: ■ master—The master node of the cluster and host of the NameNode and Job- Tracker daemons ■ backup—The server that hosts the Secondary NameNode daemon ■ hadoop1, hadoop2, hadoop3, ...—The slave boxes of the cluster running both DataNode and TaskTracker daemons Using the preceding naming convention, listing 2.2 is a modified version of the pseudo-distributed configuration files (listing 2.1) that can be used as a skeleton for your cluster’s setup. Listing 2.2 Example configuration files for fully distributed mode core-site.xml fs.default.name q hdfs://master:9000 The name of the default file system. A URI whose scheme and authority determine the FileSystem implementation. mapred-site.xml Locate NameNode for filesystem 32 CHAPTER 2 Starting Hadoop mapred.job.tracker w master:9001 The host and port that the MapReduce job tracker runs at. hdfs-site.xml dfs.replication e 3 The actual number of replications can be specified when the file is created. The key differences are ■ We explicitly stated the hostname for location of the NameNode q and JobTracker w daemons. ■ We increased the HDFS replication factor to take advantage of distributed storage e. Recall that data is replicated across HDFS to increase availability and reliability. We also need to update the masters and slaves files to reflect the locations of the other daemons. [hadoop-user@master]$ cat masters backup [hadoop-user@master]$ cat slaves hadoop1 hadoop2 hadoop3 ... Once you have copied these files across all the nodes in your cluster, be sure to format HDFS to prepare it for storage: [hadoop-user@master]$ bin/hadoop namenode-format Now you can start the Hadoop daemons: [hadoop-user@master]$ bin/start-all.sh and verify the nodes are running their assigned jobs. [hadoop-user@master]$ jps 30879 JobTracker 30717 NameNode Locate JobTracker master Increase HDFS replication factor Running Hadoop 33 Switching between modes A practice that I found useful when starting with Hadoop was to use symbolic links to switch between Hadoop modes instead of constantly editing the XML files. To do so, create a separate configuration folder for each of the modes and place the appropriate version of the XML files in the corresponding folder. Below is an example directory listing: [hadoop@hadoop_master hadoop]$ ls -l total 4884 drwxr-xr-x 2 hadoop-user hadoop 4096 Nov 26 17:36 bin -rw-rw-r- - 1 hadoop-user hadoop 57430 Nov 13 19:09 build.xml drwxr-xr-x 4 hadoop-user hadoop 4096 Nov 13 19:14 c++ -rw-rw-r- - 1 hadoop-user hadoop 287046 Nov 13 19:09 CHANGES.txt lrwxrwxrwx 1 hadoop-user hadoop 12 Jan 5 16:06 conf -> conf.cluster drwxr-xr-x 2 hadoop-user hadoop 4096 Jan 8 17:05 conf.cluster drwxr-xr-x 2 hadoop-user hadoop 4096 Jan 2 15:07 conf.pseudo drwxr-xr-x 2 hadoop-user hadoop 4096 Dec 1 10:10 conf.standalone drwxr-xr-x 12 hadoop-user hadoop 4096 Nov 13 19:09 contrib drwxrwxr-x 5 hadoop-user hadoop 4096 Jan 2 09:28 datastore drwxr-xr-x 6 hadoop-user hadoop 4096 Nov 26 17:36 docs ... You can then switch between configurations by using the Linux ln command (e.g., ln -s conf.cluster conf). This practice is also useful to temporarily pull a node out of the cluster to debug a MapReduce program in pseudo-distributed mode, but be sure that the modes have different file locations for HDFS and stop all daemons on the node before changing configurations. 30965 Jps [hadoop-user@backup]$ jps 2099 Jps 1679 SecondaryNameNode [hadoop-user@hadoop1]$ jps 7101 TaskTracker 7617 Jps 6988 DataNode You have a functioning cluster! Now that we’ve gone through all the settings to successfully get a Hadoop cluster up and running, we’ll introduce the Web UI for basic monitoring of the cluster’s state. 34 CHAPTER 2 Starting Hadoop 2.4 Web-based cluster UI Having covered the operational modes of Hadoop, we can now introduce the web interfaces that Hadoop provides to monitor the health of your cluster. The browser in- terface allows you to access information you desire much faster than digging through logs and directories. The NameNode hosts a general report on port 50070. It gives you an overview of the state of your cluster’s HDFS. Figure 2.4 displays this report for a 2-node cluster example. From this interface, you can browse through the filesystem, check the status of each DataNode in your cluster, and peruse the Hadoop daemon logs to verify your cluster is functioning correctly. Hadoop provides a similar status overview of ongoing MapReduce jobs. Figure 2.5 depicts one hosted at port 50030 of the JobTracker. Again, a wealth of information is available through this reporting interface. You can access the status of ongoing MapReduce tasks as well as detailed reports about completed jobs. The latter is of particular importance—these logs describe which nodes performed which tasks and the time/resources required to complete each task. Finally, the Hadoop configuration for each job is also available, as shown in figure 2.6. With all of this information you can streamline your MapReduce programs to better utilize the resources of your cluster. Figure 2.4 A snapshot of the HDFS web interface. From this interface you can browse through the HDFS filesystem, determine the storage available on each individual node, and monitor the overall health of your cluster. Web-based cluster UI 35 Figure 2.5 A snapshot of the MapReduce web interface. This tool allows you to monitor active MapReduce jobs and access the logs of each map and reduce task. The logs of previously submitted jobs are also available and are useful for debugging your programs. Figure 2.6 Configuration details for a specific MapReduce job. This information is potentially useful when tuning parameters to optimize the performance of your programs. 36 CHAPTER 2 Starting Hadoop Though the usefulness of these tools may not be immediately apparent at this stage, they’ll come in handy as you begin to perform more sophisticated tasks on your cluster. You’ll realize their importance as we study Hadoop more in depth. 2.5 Summary In this chapter we’ve discussed the key nodes and the roles they play within the Hadoop architecture. You’ve learned how to configure your cluster, as well as manage some basic tools to monitor your cluster’s overall health. Overall, this chapter focuses on one-time tasks. Once you’ve formatted the NameNode for your cluster, you’ll (hopefully) never need to do so again. Likewise, you shouldn’t keep altering the hadoop-site.xml configuration file for your cluster or assigning daemons to nodes. In the next chapter, you’ll learn about the aspects of Hadoop you’ll be interacting with on a daily basis, such as managing files in HDFS. With this knowledge you’ll be able to begin writing your own MapReduce applications and realize the true potential that Hadoop has to offer. 3 37 Components of Hadoop This chapter covers ■ Managing files in HDFS ■ Reading and writing input and output data Analyzing components of the MapReduce framework ■ In the last chapter we looked at setting up and installing Hadoop. We covered what the different nodes do and how to configure them to work with each other. Now that you have Hadoop running, let’s look at the Hadoop framework from a pro- grammer’s perspective. If the previous chapter is like teaching you how to connect your turntable, your mixer, your amplifier, and your speakers together, then this chapter is about the techniques of mixing music. We first cover HDFS, where you’ll store data that your Hadoop applications will process. Next we explain the MapReduce framework in more detail. In chapter 1 we’ve already seen a MapReduce program, but we discussed the logic only at the conceptual level. In this chapter we get to know the Java classes and methods, as well as the underlying processing steps. We also learn how to read and write using different data formats. 38 CHAPTER 3 Components of Hadoop 3.1 Working with files in HDFS HDFS is a filesystem designed for large-scale distributed data processing under frame- works such as MapReduce. You can store a big data set of (say) 100 TB as a single file in HDFS , something that would overwhelm most other filesystems. We discussed in chapter 2 how to replicate the data for availability and distribute it over multiple ma- chines to enable parallel processing. HDFS abstracts these details away and gives you the illusion that you’re dealing with only a single file. As HDFS isn’t a native Unix filesystem, standard Unix file tools, such as ls and cp don’t work on it, 1 and neither do standard file read/write operations, such as fopen() and fread(). On the other hand, Hadoop does provide a set of command line utilities that work similarly to the Linux file commands. In the next section we’ll discuss those Hadoop file shell commands, which are your primary interface with the HDFS system. Section 3.1.2 covers Hadoop Java libraries for handling HDFS files programmatically. NOTE A typical Hadoop workflow creates data files (such as log files) elsewhere and copies them into HDFS using one of the command line utilities discussed in the next section. Your MapReduce programs then process this data, but they usually don’t read any HDFS files directly. Instead they rely on the MapReduce framework to read and parse the HDFS files into individual records (key/ value pairs ), which are the unit of data MapReduce programs do work on. You rarely will have to programmatically read or write HDFS files except for custom import and export of data. 3.1.1 Basic file commands Hadoop file commands take the form of hadoop fs -cmd where cmd is the specific file command and is a variable number of arguments. The command cmd is usually named after the corresponding Unix equivalent. For example, the command for listing files is 2 hadoop fs –ls Let’s look at the most common file management tasks in Hadoop, which include Adding files and directories■ Retrieving files■ Deleting files■ 1 There are several ongoing projects that try to make HDFS mountable as a Unix filesystem. More details are at http://wiki.apache.org/hadoop/MountableHDFS. As of this writing these projects aren’t officially part of Hadoop and they may not have the reliability needed for some production systems. 2 Some older documentation shows file utilities in the form of hadoop dfs -cmd . Both dfs and fs are equivalent, although fs is the preferred form now. Working with files in HDFS 39 ADDING FILES AND DIRECTORIES Before you can run Hadoop programs on data stored in HDFS, you’ll need to put the data into HDFS first. Let’s assume you’ve already formatted and started a HDFS filesys- tem. (For learning purposes, we recommend a pseudo-distributed configuration as a playground.) Let’s create a directory and put a file in it. HDFS has a default working directory of /user/$USER, where $USER is your login user name. This directory isn’t automatically created for you, though, so let’s create it with the mkdir command. For the purpose of illustration, we use chuck. You should substitute your user name in the example commands. hadoop fs –mkdir /user/chuck URI for specifying exact file and directory location Hadoop file commands can interact with both the HDFS filesystem and the local filesystem. (And as we’ll see in chapter 9, it can also interact with Amazon S3 as a filesystem.) A URI pinpoints the location of a specific file or directory. The full URI format is scheme://authority/path. The scheme is similar to a protocol. It can be hdfs or file, to specify the HDFS filesystem or the local filesystem, respectively. For HDFS, authority is the NameNode host and path is the path of the file or directory of interest. For example, for a standard pseudo-distributed configuration running HDFS on the local machine on port 9000, a URI to access the example.txt file under the directory user/chuck will look like hdfs://localhost:9000/user/chuck/example.txt. You can use the Hadoop cat command to show the content of that file: hadoop fs -cat hdfs://localhost:9000/user/chuck/example.txt As we’ll see shortly, most setups don’t need to specify the scheme://authority part of the URI. When dealing with the local filesystem, you’ll probably prefer your standard Unix commands rather than the Hadoop file commands. For copying files between the local filesystem and HDFS, Hadoop commands, such as put and get use the local filesystem as source and destination, respectively, without you specifying the file:// scheme. For other commands, if you leave out the scheme://authority part of the URI, the default from the Hadoop configuration is used. For example, if you have changed the conf/core-site.xml file to the pseudo-distributed configuration, your fs.default.name property in the file should be fs.default.name hdfs://localhost:9000 Under this configuration, shorten the URI hdfs://localhost:9000/user/chuck/example. txt to /user/chuck/example.txt. Furthermore, HDFS defaults to a current working directory of /user/$USER, where $USER is your login user name. If you’re logged in as chuck, then shorten the URI hdfs://localhost:9000/user/chuck/example.txt to example.txt. The Hadoop cat command to show the content of the file is hadoop fs -cat example.txt 40 CHAPTER 3 Components of Hadoop Hadoop’s mkdir command automatically creates parent directories if they don’t already exist, similar to the Unix mkdir command with the -p option. So the preced- ing command will create the /user directory too. Let’s check on the directories with the ls command. hadoop fs -ls / You’ll see this response showing the /user directory at the root / directory. Found 1 items drwxr-xr-x - chuck supergroup 0 2009-01-14 10:23 /user If you want to see all the subdirectories, in a way similar to Unix’s ls with the -r option, you can use Hadoop’s lsr command . hadoop fs -lsr / You’ll see all the files and directories recursively. drwxr-xr-x - chuck supergroup 0 2009-01-14 10:23 /user drwxr-xr-x - chuck supergroup 0 2009-01-14 10:23 /user/chuck Now that we have a working directory, we can put a file into it. Create some text file on your local filesystem called example.txt. The Hadoop command put is used to copy files from the local system into HDFS. hadoop fs -put example.txt . Note the period (.) as the last argument in the command above. It means that we’re put- ting the file into the default working directory. The command above is equivalent to hadoop fs -put example.txt /user/chuck We can re-execute the recursive file listing command to see that the new file is added to HDFS. $ hadoop fs -lsr / drwxr-xr-x - chuck supergroup 0 2009-01-14 10:23 /user drwxr-xr-x - chuck supergroup 0 2009-01-14 11:02 /user/chuck -rw-r--r-- 1 chuck supergroup 264 2009-01-14 11:02 ➥/user/chuck/example.txt In practice we don’t need to check on all files recursively, and we may restrict ourselves to what’s in our own working directory. We would use the Hadoop ls command in its simplest form: $ hadoop fs -ls Found 1 items -rw-r--r-- 1 chuck supergroup 264 2009-01-14 11:02 ➥/user/chuck/example.txt The output displays properties, such as permission, owner, group, file size, and last modification date, all of which are familiar Unix concepts. The column stating “1” re- ports the replication factor of the file. It should always be 1 for the pseudo-distributed configuration . For production clusters, the replication factor is typically 3 but can be any positive integer. Replication factor is not applicable to directories, so they will only show a dash (-) for that column. After you’ve put data into HDFS, you can run Hadoop programs to process it. The output of the processing will be a new set of files in HDFS, and you’ll want to read or retrieve the results. RETRIEVING FILES The Hadoop command get does the exact reverse of put. It copies files from HDFS to the local filesystem. Let’s say we no longer have the example.txt file locally and we want to retrieve it from HDFS; we can run the command hadoop fs -get example.txt . to copy it into our local current working directory. Another way to access the data is to display it. The Hadoop cat command allows us to do that. hadoop fs -cat example.txt We can use the Hadoop file command with Unix pipes to send its output for further processing by other Unix commands. For example, if the file is huge (as typical Hadoop files are) and you’re interested in a quick check of its content, you can pipe the output of Hadoop’s cat into a Unix head. hadoop fs -cat example.txt | head Hadoop natively supports a tail command for looking at the last kilobyte of a file. hadoop fs -tail example.txt After you finish working with files in HDFS, you may want to delete them to free up space. DELETING FILES You shouldn’t be too surprised by now that the Hadoop command for removing files is rm. hadoop fs –rm example.txt The rm command can also be used to delete empty directories. LOOKING UP HELP A list of Hadoop file commands, together with the usage and description of each com- mand, is given in the appendix. For the most part, the commands are modeled after their Unix equivalent. You can execute hadoop fs (with no parameters) to get a com- plete list of all commands available on your version of Hadoop. You can also use help to display the usage and a short description of each command. For example, to get a summary of ls, execute hadoop fs –help ls Working with files in HDFS 41 42 CHAPTER 3 Components of Hadoop and you should see the following description: -ls : List the contents that match the specified file pattern. If path is not specified, the contents of /user/ will be listed. Directory entries are of the form dirName (full path) and file entries are of the form fileName(full path) size where n is the number of replicas specified for the file and size is the size of the file, in bytes. Although the command line utilities are sufficient for most of your interaction with the HDFS filesystem, they’re not exhaustive and there’ll be situations where you may want deeper access into the HDFS API. Let’s see how to do so in the next section. 3.1.2 Reading and writing to HDFS programmatically To motivate an examination of the HDFS Java API, we’ll develop a PutMerge program for merging files while putting them into HDFS. The command line utilities don’t sup- port this operation; we’ll use the API. The motivation for this example came when we wanted to analyze Apache log files coming from many web servers. We can copy each log file into HDFS, but in general, Hadoop works more effectively with a single large file rather than a number of smaller ones. (“Smaller” is relative here as it can still be tens or hundreds of gigabytes.) Besides, for analytics purposes we think of the log data as one big file. That it’s spread over multiple files is an incidental result of the physical web server architecture. One solution is to merge all the files first and then copy the combined file into HDFS. Unfortunately, the file merging will require a lot of disk space in the local machine. It would be much easier if we could merge all the files on the fly as we copy them into HDFS. What we need is, therefore, a PutMerge-type of operation. Hadoop’s command line utilities include a getmerge command for merging a number of HDFS files before copying them onto the local machine. What we’re looking for is the exact opposite. This is not available in Hadoop’s file utilities. We’ll write our own program using the HDFS API. The main classes for file manipulation in Hadoop are in the package org.apache. hadoop.fs . Basic Hadoop file operations include the familiar open, read, write, and close. In fact, the Hadoop file API is generic and can be used for working with filesystems other than HDFS. For our PutMerge program, we’ll use the Hadoop file API to both read the local filesystem and write to HDFS. The starting point for the Hadoop file API is the FileSystem class . This is an abstract class for interfacing with the filesystem, and there are different concrete subclasses for handling HDFS and the local filesystem. You get the desired FileSystem instance by calling the factory method FileSystem.get(Configuration conf). The Configuration class is a special class for holding key/value configuration parameters. Its default instantiation is based on the resource configuration for your HDFS system. We can get the FileSystem object to interface with HDFS by Configuration conf = new Configuration(); FileSystem hdfs = FileSystem.get(conf); To get a FileSystem object specifically for the local filesystem, there’s the FileSystem. getLocal(Configuration conf) factory method. FileSystem local = FileSystem.getLocal(conf); Hadoop file API uses Path objects to encode file and directory names and FileStatus objects to store metadata for files and directories. Our PutMerge program will merge all files from a local directory. We use the FileSystem’s listStatus () method to get a list of files in a directory. Path inputDir = new Path(args[0]); FileStatus[] inputFiles = local.listStatus(inputDir); The length of the inputFiles array is the number of files in the specified directory. Each FileStatus object in inputFiles has metadata information such as file length, permissions, modification time, and others. Of interest to our PutMerge program is each file’s Path representation, inputFiles[i].getPath(). We can use this Path to request an FSDataInputStream object for reading in the file. FSDataInputStream in = local.open(inputFiles[i].getPath()); byte buffer[] = new byte[256]; int bytesRead = 0; while( (bytesRead = in.read(buffer)) > 0) { ... } in.close(); FSDataInputStream is a subclass of Java’s standard java.io.DataInputStream with ad- ditional support for random access. For writing to a HDFS file, there’s the analogous FSDataOutputStream object . Path hdfsFile = new Path(args[1]); FSDataOutputStream out = hdfs.create(hdfsFile); out.write(buffer, 0, bytesRead); out.close(); To complete the PutMerge program, we create a loop that goes through all the files in inputFiles as we read each one in and write it out to the destination HDFS file. You can see the complete program in listing 3.1. Listing 3.1 A PutMerge program import java.io.IOException; import org.apache.hadoop.conf.Configuration; import org.apache.hadoop.fs.FSDataInputStream; import org.apache.hadoop.fs.FSDataOutputStream; import org.apache.hadoop.fs.FileStatus; import org.apache.hadoop.fs.FileSystem; import org.apache.hadoop.fs.Path; public class PutMerge { Working with files in HDFS 43 44 CHAPTER 3 Components of Hadoop public static void main(String[] args) throws IOException { Configuration conf = new Configuration(); FileSystem hdfs = FileSystem.get(conf); FileSystem local = FileSystem.getLocal(conf); Path inputDir = new Path(args[0]); q Path hdfsFile = new Path(args[1]); try { FileStatus[] inputFiles = local.listStatus(inputDir); w FSDataOutputStream out = hdfs.create(hdfsFile); e for (int i=0; i 0) { out.write(buffer, 0, bytesRead); } in.close(); } out.close(); } catch (IOException e) { e.printStackTrace(); } } } The general flow of the program involves first setting the local directory and the HDFS destination file based on user-specified arguments q. In w we extract information about each file in the local input directory. We create an output stream to write to the HDFS file in e. We loop through each file in the local directory, and r opens an in- put stream to read that file. The rest of the code is standard Java file copy. The FileSystem class also has methods such as delete() , exists() , mkdirs() , and rename() for other standard file operations. You can find the most recent Javadoc for the Hadoop file API at http://hadoop.apache.org/core/docs/current/api/org/ apache/hadoop/fs/package-summary.html. We have covered how to work with files in HDFS. You now know a few ways to put data into and out of HDFS. But merely having data isn’t terribly interesting. You want to process it, analyze it, and do other things. Let’s conclude our discussion of HDFS and move on to the other major component of Hadoop, the MapReduce framework, and how to program under it. 3.2 Anatomy of a MapReduce program As we have mentioned before, a MapReduce program processes data by manipulating (key/value) pairs in the general form map: (K1,V1) ➞ list(K2,V2) reduce: (K2,list(V2)) ➞ list(K3,V3) Specify input directory and output file Get list of local files Create HDFS output stream Open local input stream Anatomy of a MapReduce program 45 Not surprisingly, this is an overly generic representation of the data flow. In this section we learn more details about each stage in a typical MapReduce program. Figure 3.1 displays a high-level diagram of the entire process, and we further dissect each compo- nent as we step through the flow. Input Data Node 1 Node 1 Node 2 Node 2 Input data is distributed to nodes Map Map Reduce Reduce Each map task works on a “split” of data Mapper outputs intermediate data Data exchange between nodes in a “shuffle” process Intermediate data of the same key goes to the same reducer Reducer output is stored Figure 3.1 The general MapReduce data flow. Note that after distributing input data to different nodes, the only time nodes communicate with each other is at the “shuffle” step . This restriction on communication greatly helps scalability. 46 CHAPTER 3 Components of Hadoop Before we analyze how data gets passed onto each individual stage, we should first fa- miliarize ourselves with the data types that Hadoop supports. 3.2.1 Hadoop data types Despite our many discussions regarding keys and values, we have yet to mention their types. The MapReduce framework won’t allow them to be any arbitrary class. For ex- ample, although we can and often do talk about certain keys and values as integers, strings, and so on, they aren’t exactly standard Java classes, such as Integer, String, and so forth. This is because the MapReduce framework has a certain defined way of serializing the key/value pairs to move them across the cluster’s network, and only classes that support this kind of serialization can function as keys or values in the framework. More specifically, classes that implement the Writable interface can be values, and classes that implement the WritableComparable interface can be either keys or values. Note that the WritableComparable interface is a combination of the Writable and java.lang.Comparable interfaces . We need the comparability requirement for keys because they will be sorted at the reduce stage, whereas values are simply passed through. Hadoop comes with a number of predefined classes that implement WritableComparable, including wrapper classes for all the basic data types, as seen in table 3.1. Table 3.1 List of frequently used types for the key/value pairs . These classes all implement the WritableComparable interface. Class Description BooleanWritable Wrapper for a standard Boolean variable ByteWritable Wrapper for a single byte DoubleWritable Wrapper for a Double FloatWritable Wrapper for a Float IntWritable Wrapper for a Integer LongWritable Wrapper for a Long Text Wrapper to store text using the UTF8 format NullWritable Placeholder when the key or value is not needed Keys and values can take on types beyond the basic ones which Hadoop natively sup- ports. You can create your own custom type as long as it implements the Writable (or WritableComparable) interface. For example, listing 3.2 shows a class that can represent edges in a network. This may represent a flight route between two cities. Listing 3.2 An example class that implements the WritableComparable interface public class Edge implements WritableComparable{ private String departureNode; private String arrivalNode; public String getDepartureNode() { return departureNode;} @Override public void readFields(DataInput in) throws IOException { q departureNode = in.readUTF(); arrivalNode = in.readUTF(); } @Override public void write(DataOutput out) throws IOException { w out.writeUTF(departureNode); out.writeUTF(arrivalNode); } @Override public int compareTo(Edge o) { e return (departureNode.compareTo(o.departureNode) != 0) ? departureNode.compareTo(o.departureNode) : arrivalNode.compareTo(o.arrivalNode); } } The Edge class implements the readFields () q and write() w methods of the Writable interface. They work with the Java DataInput and DataOutput classes to serialize the class contents. Implement the compareTo() method e for the Comparable interface. It returns -1, 0, or 1 if the called Edge is less than, equal to, or greater than the given Edge . With the data type interfaces now defined, we can proceed to the first stage of the data flow process as described in figure 3.1: the mapper. 3.2.2 Mapper To serve as the mapper , a class implements from the Mapper interface and inherits the MapReduceBase class . The MapReduceBase class, not surprisingly, serves as the base class for both mappers and reducers. It includes two methods that effectively act as the constructor and destructor for the class: ■ void configure( JobConf job) —In this function you can extract the parameters set either by the configuration XML files or in the main class of your application. Call this function before any data processing begins. ■ void close ()—As the last action before the map task terminates, this function should wrap up any loose ends—database connections, open files, and so on. The Mapper interface is responsible for the data processing step. It utilizes Java generics of the form Mapper where the key classes and value classes implement the WritableComparable and Writable interfaces, respectively. Its single method is to process an individual (key/value) pair: Specify how to read data in Specify how to write data out Define ordering of data Anatomy of a MapReduce program 47 48 CHAPTER 3 Components of Hadoop void map(K1 key, V1 value, OutputCollector output, Reporter reporter ) throws IOException The function generates a (possibly empty) list of (K2, V2) pairs for a given (K1, V1) input pair. The OutputCollector receives the output of the mapping process, and the Reporter provides the option to record extra information about the mapper as the task progresses. Hadoop provides a few useful mapper implementations. You can see some of them in the table 3.2. Table 3.2 Some useful Mapper implementations predefined by Hadoop Class Description IdentityMapper Implements Mapper and maps inputs directly to outputs InverseMapper Implements Mapper and reverses the key/value pair RegexMapper Implements Mapper and generates a (match, 1) pair for every regular expression match TokenCountMapper Implements Mapper and generates a (token, 1) pair when the input value is tokenized As the MapReduce name implies, the major data flow operation after map is the re- duce phase, shown in the bottom part of figure 3.1. 3.2.3 Reducer As with any mapper implementation, a reducer must first extend the MapReduce base class to allow for configuration and cleanup. In addition, it must also implement the Reducer interface which has the following single method: void reduce(K2 key, Iterator values, OutputCollector output, Reporter reporter ) throws IOException When the reducer task receives the output from the various mappers, it sorts the incoming data on the key of the (key/value) pair and groups together all values of the same key. The reduce() function is then called, and it generates a (possibly empty) list of (K3, V3) pairs by iterating over the values associated with a given key. The OutputCollector receives the output of the reduce process and writes it to an output file. The Reporter provides the option to record extra information about the reducer as the task progresses. Table 3.3 lists a couple of basic reducer implementations provided by Hadoop. Table 3.3 Some useful Reducer implementations predefined by Hadoop Class Description IdentityReducer Implements Reducer and maps inputs directly to outputs LongSumReducer Implements Reducer and determines the sum of all values corresponding to the given key Although we have referred to Hadoop programs as MapReduce applications, there is a vital step between the two stages: directing the result of the mappers to the different reducers. This is the responsibility of the partitioner. 3.2.4 Partitioner— redirecting output from Mapper A common misconception for first-time MapReduce programmers is to use only a single reducer. After all, a single reducer sorts all of your data before processing— and who doesn’t like sorted data? Our discussions regarding MapReduce expose the folly of such thinking. We would have ignored the benefits of parallel com- putation . With one reducer, our compute cloud has been demoted to a compute raindrop. With multiple reducers, we need some way to determine the appropriate one to send a (key/value) pair outputted by a mapper. The default behavior is to hash the key to determine the reducer. Hadoop enforces this strategy by use of the HashPartitioner class . Sometimes the HashPartitioner will steer you awry. Let’s return to the Edge class introduced in section 3.2.1. Suppose you used the Edge class to analyze flight information data to determine the number of passengers departing from each airport. Such data may be (San Francisco, Los Angeles) Chuck Lam (San Francisco, Dallas) James Warren ... If you used HashPartitioner, the two rows could be sent to different reducers. The number of departures would be processed twice and both times erroneously. How do we customize the partitioner for your applications? In this situation, we want all edges with a common departure point to be sent to the same reducer. This is done easily enough by hashing the departureNode member of the Edge : public class EdgePartitioner implements Partitioner { @Override public int getPartition(Edge key, Writable value, int numPartitions) { return key.getDepartureNode().hashCode() % numPartitions; } @Override public void configure(JobConf conf) { } } Anatomy of a MapReduce program 49 50 CHAPTER 3 Components of Hadoop A custom partitioner only needs to implement two functions: configure() and getPartition() . The former uses the Hadoop job configuration to configure the partitioner, and the latter returns an integer between 0 and the number of reduce tasks indexing to which reducer the (key/value) pair will be sent. The exact mechanics of the partitioner may be difficult to follow. Figure 3.2 illustrates this for better understanding. Between the map and reduce stages, a MapReduce application must take the output from the mapper tasks and distribute the results among the reducer tasks. This process is typically called shuffling , because the output of a mapper on a single node may be sent to reducers across multiple nodes in the cluster. 3.2.5 Combiner—local reduce In many situations with MapReduce applications, we may wish to perform a “local reduce ” before we distribute the mapper results. Consider the WordCount example of IN IN IN IN IN IN IN IN IN OUT OUT OUT Map Reduce Figure 3.2 The MapReduce data flow, with an emphasis on partitioning and shuffling. Each icon is a key/value pair. The shapes represents keys, whereas the inner patterns represent values. After shuffling, all icons of the same shape (key) are in the same reducer. Different keys can go to the same reducer, as seen in the rightmost reducer. The partitioner decides which key goes where. Note that the leftmost reducer has more load due to more data under the “ellipse” key. Reading and writing 51 chapter 1 once more. If the job processes a document containing the word “the” 574 times, it’s much more efficient to store and shuffle the pair (“the”, 574) once instead of the pair (“the”, 1) multiple times. This processing step is known as combining. We explain combiners in more depth in section 4.6. 3.2.6 Word counting with predefined mapper and reducer classes We have concluded our preliminary coverage of all the basic components of MapReduce. Now that you’ve seen more classes provided by Hadoop, it’ll be fun to revisit the Word- Count example (see listing 3.3), using some of the classes we’ve learned. Listing 3.3 Revised version of the WordCount example public class WordCount2 { public static void main(String[] args) { JobClient client = new JobClient(); JobConf conf = new JobConf(WordCount2.class); FileInputFormat.addInputPath(conf, new Path(args[0])); FileOutputFormat.setOutputPath(conf, new Path(args[1])); conf.setOutputKeyClass(Text.class); conf.setOutputValueClass(LongWritable.class); conf.setMapperClass(TokenCountMapper.class); q conf.setCombinerClass(LongSumReducer.class); conf.setReducerClass(LongSumReducer.class); w client.setConf(conf); try { JobClient.runJob(conf); } catch (Exception e) { e.printStackTrace(); } } } We have to write only the driver for this MapReduce program because we have used Hadoop’s predefined TokenCountMapper class q and LongSumReducer class w. Easy, isn’t it? Hadoop provides the ability to generate more sophisticated programs (this will be the focus of part 2 of the book), but we want to emphasize that Hadoop allows you to rapidly generate useful programs with a minimal amount of code. 3.3 Reading and writing Let’s see how MapReduce reads input data and writes output data and focus on the file formats it uses. To enable easy distributed processing, MapReduce makes certain assumptions about the data it’s processing. It also provides flexibility in dealing with a variety of data formats. Input data usually resides in large files, typically tens or hundreds of gigabytes or even more. One of the fundamental principles of MapReduce’s processing power is the splitting of the input data into chunks . You can process these chunks in parallel using multiple machines. In Hadoop terminology these chunks are called input splits . Hadoop’s own TokenCountMapper Hadoop’s own LongSumReducer 52 CHAPTER 3 Components of Hadoop The size of each split should be small enough for a more granular parallelization . (If all the input data is in one split, then there is no parallelization.) On the other hand, each split shouldn’t be so small that the overhead of starting and stopping the processing of a split becomes a large fraction of execution time. The principle of dividing input data (which often can be one single massive file) into splits for parallel processing explains some of the design decisions behind Hadoop’s generic FileSystem as well as HDFS in particular. For example, Hadoop’s FileSystem provides the class FSDataInputStream for file reading rather than using Java’s java. io.DataInputStream . FSDataInputStream extends DataInputStream with random read access, a feature that MapReduce requires because a machine may be assigned to process a split that sits right in the middle of an input file. Without random access, it would be extremely inefficient to have to read the file from the beginning until you reach the location of the split. You can also see how HDFS is designed for storing data that MapReduce will split and process in parallel. HDFS stores files in blocks spread over multiple machines. Roughly speaking, each file block is a split. As different machines will likely have different blocks, parallelization is automatic if each split/ block is processed by the machine that it’s residing at. Furthermore, as HDFS replicates blocks in multiple nodes for reliability, MapReduce can choose any of the nodes that have a copy of a split/block. Input splits and record boundaries Note that input splits are a logical division of your records whereas HDFS blocks are a physical division of the input data. It’s extremely efficient when they’re the same but in practice it’s never perfectly aligned. Records may cross block boundaries. Hadoop guarantees the processing of all records . A machine processing a particular split may fetch a fragment of a record from a block other than its “main” block and which may reside remotely. The communication cost for fetching a record fragment is inconsequential because it happens relatively rarely. You’ll recall that MapReduce works on key/value pairs. So far we’ve seen that Hadoop by default considers each line in the input file to be a record and the key/value pair is the byte offset (key) and content of the line (value), respectively. You may not have recorded all your data that way. Hadoop supports a few other data formats and allows you to define your own. 3.3.1 InputFormat The way an input file is split up and read by Hadoop is defined by one of the imple- mentations of the InputFormat interface . TextInputFormat is the default Input- Format implementation, and it’s the data format we’ve been implicitly using up to now. It’s often useful for input data that has no definite key value, when you want to get the content one line at a time. The key returned by TextInputFormat is the byte offset of each line, and we have yet to see any program that uses that key for its data processing. POPULAR INPUTFORMAT CLASSES Table 3.4 lists other popular implementations of InputFormat along with a descrip- tion of the key/value pair each one passes to the mapper. Table 3.4 Main InputFormat classes. TextInputFormat is the default unless an alternative is specified. The object type for key and value are also described. InputFormat Description TextInputFormat Each line in the text files is a record. Key is the byte offset of the line, and value is the content of the line. key: LongWritable value: Text KeyValueTextInputFormat Each line in the text files is a record. The first separator character divides each line. Everything before the separator is the key, and everything after is the value. The separator is set by the key.value.separator.in.input. line property, and the default is the tab (\t) character. key: Text value: Text SequenceFileInputFormat An InputFormat for reading in sequence files. Key and value are user defined. Sequence file is a Hadoop- specific compressed binary file format. It’s optimized for passing data between the output of one MapReduce job to the input of some other MapReduce job. key: K (user defined) value: V (user defined) NLineInputFormat Same as TextInputFormat, but each split is guaranteed to have exactly N lines. The mapred.line.input.format. linespermap property, which defaults to one, sets N. key: LongWritable value: Text KeyValueTextInputFormat is used in the more structured input files where a pre- defined character, usually a tab (\t), separates the key and value of each line (record). For example, you may have a tab-separated data file of timestamps and URLs: 17:16:18 http://hadoop.apache.org/core/docs/r0.19.0/api/index.html 17:16:19 http://hadoop.apache.org/core/docs/r0.19.0/mapred_tutorial.html 17:16:20 http://wiki.apache.org/hadoop/GettingStartedWithHadoop 17:16:20 http://www.maxim.com/hotties/2008/finalist_gallery.aspx 17:16:25 http://wiki.apache.org/hadoop/ ... Reading and writing 53 54 CHAPTER 3 Components of Hadoop You can set your JobConf object to use the KeyValueTextInputFormat class to read this file. conf.setInputFormat(KeyValueTextInputFormat.class); Given the preceding example file, the first record your mapper reads will have a key of “17:16:18” and a value of “http://hadoop.apache.org/core/docs/r0.19.0/api/ index.html”. The second record to your mapper will have a key of “17:16:19” and a value of “http://hadoop.apache.org/core/docs/r0.19.0/mapred_tutorial.html.” And so on. Recall that our previous mappers had used LongWritable and Text as the key and value types, respectively. LongWritable is a reasonable type for the key under TextInputFormat because the key is a numerical offset. When using KeyValueTextInputFormat, both the key and the value will be of type Text , and you’ll have to change your Mapper implementation and map() method to reflect the new key type. The input data to your MapReduce job does not necessarily have to be some external data. In fact it’s often the case that the input to one MapReduce job is the output of some other MapReduce job. As we’ll see, you can customize your output format too. The default output format writes the output in the same format that KeyValueTextInputFormat can read back in (i.e., each line is a record with key and value separated by a tab character). Hadoop provides a much more efficient binary compressed file format called sequence file . This sequence file is optimized for Hadoop processing and should be the preferred format when chaining multiple MapReduce jobs. The InputFormat class to read sequence files is SequenceFileInputFormat . The object type for key and value in a sequence file are definable by the user. The output and the input type have to match, and your Mapper implementation and map() method have to take in the right input type. CREATING A CUSTOM INPUTFORMAT—INPUTSPLIT AND RECORDREADER Sometimes you may want to read input data in a way different from the standard InputFormat classes. In that case you’ll have to write your own custom InputFormat class. Let’s look at what it involves. InputFormat is an interface consisting of only two methods. public interface InputFormat { InputSplit[] getSplits(JobConf job, int numSplits) throws IOException; RecordReader getRecordReader(InputSplit split, JobConf job, Reporter reporter) throws IOException; } The two methods sum up the functions that InputFormat has to perform: ■ Identify all the files used as input data and divide them into input splits. Each map task is assigned one split. ■ Provide an object (RecordReader) to iterate through records in a given split, and to parse each record into key and value of predefined types. Who wants to worry about how files are divided into splits ? In creating your own InputFormat class you should subclass the FileInputFormat class, which takes care of file splitting. In fact, all the InputFormat classes in table 3.4 subclass FileInputFormat. FileInputFormat implements the getSplits() method but leaves getRecordReader() abstract for the subclass to fill out. FileInputFormat’s getSplits() implementation tries to divide the input data into roughly the number of splits specified in numSplits , subject to the constraints that each split must have more than mapred.min.split.size number of bytes but also be smaller than the block size of the filesystem. In practice, a split usually ends up being the size of a block, which defaults to 64 MB in HDFS. FileInputFormat has a number of protected methods a subclass can overwrite to change its behavior, one of which is the isSplitable(FileSystem fs, Path filename) method. It checks whether you can split a given file. The default implementation always returns true, so all files larger than a block will be split. Sometimes you may want a file to be its own split, and you’ll overwrite isSplitable() to return false in those situations. For example, some file compression schemes don’t support splits. (You can’t start reading from the middle of those files.) Some data processing operations, such as file conversion, will need to treat each file as an atomic record and one should also not be able to split it. In using FileInputFormat you focus on customizing RecordReader, which is responsible for parsing an input split into records and then parsing each record into a key/value pair . Let’s look at the signature of this interface. public interface RecordReader { boolean next(K key, V value) throws IOException; K createKey(); V createValue(); long getPos() throws IOException; public void close() throws IOException; float getProgress() throws IOException; } Instead of writing our own RecordReader , we’ll again leverage existing classes pro- vided by Hadoop. For example, LineRecordReader implements RecordReader . It’s used in TextInputFormat and reads one line at a time, with byte offset as key and line content as value. KeyValueLineRecordReader uses KeyValueTextInputFormat . For the most part, your custom RecordReader will be a wrapper around an existing implementation, and most of the action will be in the next() method . One use case for writing your own custom InputFormat class is to read records in a specific type rather than the generic Text type. For example, we had previously used KeyValueTextInputFormat to read a tab-separated data file of timestamps Reading and writing 55 56 CHAPTER 3 Components of Hadoop and URLs. The class ends up treating both the timestamp and the URL as Text type. For our illustration, let’s create a TimeUrlTextInputFormat that works exactly the same but treats the URL as a URLWritable type 3. As mentioned earlier, we create our InputFormat class by extending FileInputFormat and implementing the factory method to return our RecordReader. public class TimeUrlTextInputFormat extends ➥ FileInputFormat { public RecordReader getRecordReader( ➥ InputSplit input, JobConf job, Reporter reporter) ➥ throws IOException { return new TimeUrlLineRecordReader(job, (FileSplit)input); } } Our URLWritable class is quite straightforward: public class URLWritable implements Writable { protected URL url; public URLWritable() { } public URLWritable(URL url) { this.url = url; } public void write(DataOutput out) throws IOException { out.writeUTF(url.toString()); } public void readFields(DataInput in) throws IOException { url = new URL(in.readUTF()); } public void set(String s) throws MalformedURLException { url = new URL(s); } } Our TimeUrlLineRecordReader will implement the six methods in the RecordReader interface, in addition to the class constructor. It’s mostly a wrapper around KeyValue- TextInputFormat, but converts the record value from Text to type URLWritable. class TimeUrlLineRecordReader implements RecordReader { private KeyValueLineRecordReader lineReader; private Text lineKey, lineValue; public TimeUrlLineRecordReader(JobConf job, FileSplit split) throws 3 We may also want the time key to be some type other than Text. For example, we can make up a type CalendarWritableComparable for it. We leave that as an exercise for the reader as we focus on a simpler illustration. ➥ IOException { lineReader = new KeyValueLineRecordReader(job, split); lineKey = lineReader.createKey(); lineValue = lineReader.createValue(); } public boolean next(Text key, URLWritable value) throws IOException { if (!lineReader.next(lineKey, lineValue)) { return false; } key.set(lineKey); value.set(lineValue.toString()); return true; } public Text createKey() { return new Text(""); } public URLWritable createValue() { return new URLWritable(); } public long getPos() throws IOException { return lineReader.getPos(); } public float getProgress() throws IOException { return lineReader.getProgress(); } public void close() throws IOException { lineReader.close(); } } Our TimeUrlLineRecordReader class creates a KeyValueLineRecordReader object and passes the getPos() , getProgress() , and close() method calls directly to it. The next () method casts the lineValue Text object into the URLWritable type. 3.3.2 OutputFormat MapReduce outputs data into files using the OutputFormat class , which is analogous to the InputFormat class. The output has no splits, as each reducer writes its output only to its own file. The output files reside in a common directory and are typically named part-nnnnn, where nnnnn is the partition ID of the reducer. RecordWriter objects format the output and RecordReaders parse the format of the input. Hadoop provides several standard implementations of OutputFormat, as shown in table 3.5. Not surprisingly, almost all the ones we deal with inherit from the File OutputFormat abstract class; InputFormat classes inherit from FileInputFormat. You specify the OutputFormat by calling setOutputFormat () of the JobConf object that holds the configuration of your MapReduce job. Reading and writing 57 58 CHAPTER 3 Components of Hadoop NOTE You may wonder why there’s a separation between OutputFormat (InputFormat) and FileOutputFormat (FileInputFormat) when it seems all OutputFormat (InputFormat) classes extend FileOutputFormat (FileInputFormat). Are there OutputFormat (InputFormat) classes that don’t work with files? Well, the NullOutputFormat implements OutputFormat in a trivial way and doesn’t need to subclass FileOutputFormat. More impor- tantly, there are OutputFormat (InputFormat) classes that work with databas- es rather than files, and these classes are in a separate branch in the class hierarchy from FileOutputFormat (FileInputFormat). These classes have specialized applications, and the interested reader can dig further in the online Java docu- mentation for DBInputFormat and DBOutputFormat. Table 3.5 Main OutputFormat classes. TextOutputFormat is the default. OutputFormat Description TextOutputFormat Writes each record as a line of text. Keys and values are written as strings and separated by a tab (\t) character, which can be changed in the mapred. textoutputformat.separator property. SequenceFileOutputFormat Writes the key/value pairs in Hadoop’s proprietary sequence file format. Works in conjunction with SequenceFileInputFormat. NullOutputFormat Outputs nothing. The default OutputFormat is TextOutputFormat, which writes each record as a line of text. Each record’s key and value are converted to strings through toString() , and a tab (\t) character separates them. The separator character can be changed in the mapred.textoutputformat.separator property. TextOutputFormat outputs data in a format readable by KeyValueTextInput- Format. It can also output in a format readable by TextInputFormat if you make the key type a NullWritable . In that case the key in the key/value pair is not written out, and neither is the separator character. If you want to suppress the output completely, then you should use the NullOutputFormat . Suppressing the Hadoop output is useful if your reducer writes its output in its own way and doesn’t need Hadoop to write any additional files. Finally, SequenceFileOutputFormat writes the output in a sequence file format that can be read back in using SequenceFileInputFormat. It’s useful for writing intermediate data results when chaining MapReduce jobs. 3.4 Summary Hadoop is a software framework that demands a different perspective on data process- ing. It has its own filesystem, HDFS, that stores data in a way optimized for data-intensive processing. You need specialized Hadoop tools to work with HDFS, but fortunately most of those tools follow familiar Unix or Java syntax. The data processing part of the Hadoop framework is better known as MapReduce. Although the highlight of a MapReduce program is, not surprisingly, the Map and the Reduce operations, other operations done by the framework, such as data splitting and shuffling, are crucial to how the framework works. You can customize the other operations, such as Partitioning and Combining. Hadoop provides options for reading data and also to output data of different formats. Now that we have a better understanding of how Hadoop works, let’s go on to part 2 of this book and look at various techniques for writing practical programs using Hadoop. Summary 59 Part 2 Hadoop in Action Part 2 teaches the practical skills required to write and run data processing programs in Hadoop. We explore various examples of using Hadoop to analyze a patent data set, including advanced algorithms such as the Bloom filter. We also cover programming and administration techniques that are uniquely useful to working with Hadoop in production. 4 63 Writing basic MapReduce programs This chapter covers ■ Patent data as an example data set to process with Hadoop ■ Skeleton of a MapReduce program ■ Basic MapReduce programs to count statistics ■ Hadoop’s Streaming API for writing MapReduce programs using scripting languages Combiner to improve performance■ The MapReduce programming model is unlike most programming models you may have learned. It’ll take some time and practice to gain familiarity. To help develop your proficiency, we go through many example programs in the next couple chap- ters. These examples will illustrate various MapReduce programming techniques. By applying MapReduce in multiple ways you’ll start to develop an intuition and a habit of “MapReduce thinking.” The examples will cover simple tasks to advanced uses. In one of the advanced applications we introduce the Bloom filter, a data struc- ture not normally taught in the standard computer science curriculum. You’ll see that processing large data sets, whether you’re using Hadoop or not, often requires a rethinking of the underlying algorithms. 64 CHAPTER 4 Writing basic MapReduce programs We assume you already have a basic grasp of Hadoop. You can set up Hadoop, and you have compiled and run an example program, such as word counting from chapter 1. Let’s use examples—from a real-world data set. 4.1 Getting the patent data set To do anything meaningful with Hadoop we need data. Many of our examples will use patent data sets, both of which are available from the National Bureau of Economic Research (NBER ) at http://www.nber.org/patents/. The data sets were originally compiled for the paper “The NBER Patent Citation Data File: Lessons, Insights and Methodological Tools.”1 We use the citation data set cite75_99.txt and the patent description data set apat63_99.txt. NOTE The data sets are approximately 250 MB each, which are small enough to make our examples runnable in Hadoop’s standalone or pseudo-distributed mode . You can practice writing MapReduce programs using them even when you don’t have access to a live cluster. The best part of Hadoop is that you can be fairly sure your MapReduce program will run on clusters of machines processing data sets 100 or 1,000 times larger with virtually no code changes. A popular development tactic is to create a smaller, sampled subset of your large production data and call it the development data set . This development data set may only have several hundred megabytes. You develop your program in standalone or pseudo-distributed mode with the development data set. This gives your development process a fast turnaround time, the convenience of running on your own machine, and an isolated environment for debugging. We have chosen these two data sets for our example programs because they’re similar to most data types you’ll encounter. First of all, the citation data encodes a graph in the same vein that web links and social networks are also graphs. Patents are published in chronological order; some of their properties resemble time series . Each patent is linked with a person (inventor) and a location (country of inventor). You can view them as personal or geographical data . Finally, you can look at the data as generic database relations with well-defined schemas, in a simple comma- separated format. 2 4.1.1 The patent citation data The patent citation data set contains citations from U.S. patents issued between 1975 and 1999. It has more than 16 million rows and the first few lines resemble the following: 1 NBER Working Paper 8498, by Hall, B. H., A. B. Jaffe, and M. Tratjenberg (2001). 2 There are more common data types than two data sets can possibly represent. An important one that’s missing here is text, but you’ve already seen text used in the word count example. Other missing types include XML, image, and geolocation (the lat-long variety). Math matrix is not represented in general, although the citation graph can be interpreted as a sparse 0/1 matrix. Getting the patent data set 65 "CITING","CITED" 3858241,956203 3858241,1324234 3858241,3398406 3858241,3557384 3858241,3634889 3858242,1515701 3858242,3319261 3858242,3668705 3858242,3707004 ... The data set is in the standard comma-separated values (CSV) format, with the first line a description of the columns. Each of the other lines record one particu- lar citation. For example, the second line shows that patent 3858241 cites patent 956203. The file is sorted by the citing patent. We can see that patent 3858241 cites five patents in total. Analyzing the data more quantitatively will give us deeper in- sights into it. If you’re only reading the data file, the citation data appears to be a bunch of numbers. You can “think” of this data in more interesting terms. One way is to visualize it as a graph. In figure 4.1 we’ve shown a portion of this citation graph . We can see that some patents are cited often whereas others aren’t cited at all. 3 Patents like 5936972 and 6009552 cite a similar set of patents (4354269, 4486882, 5598422) even though they don’t cite each other. We use Hadoop to derive descriptive statistics about this patent data, as well as look for interesting patterns that aren’t immediately obvious. 4.1.2 The patent description data The other data set we use is the patent description data. It has the patent number, the patent application year, the patent grant year, the number of claims, and other meta- data about patents. Look at the first few lines of this data set. It’s similar to a table in a relational database, but in CSV format. This data set has more than 2.9 million records. As in many real-world data sets, it has many missing values . "PATENT","GYEAR","GDATE","APPYEAR","COUNTRY","POSTATE","ASSIGNEE", ➥ "ASSCODE","CLAIMS","NCLASS","CAT","SUBCAT","CMADE","CRECEIVE", ➥ "RATIOCIT","GENERAL","ORIGINAL","FWDAPLAG","BCKGTLAG","SELFCTUB", ➥ "SELFCTLB","SECDUPBD","SECDLWBD" 3070801,1963,1096,,"BE","",,1,,269,6,69,,1,,0,,,,,,, 3070802,1963,1096,,"US","TX",,1,,2,6,63,,0,,,,,,,,, 3070803,1963,1096,,"US","IL",,1,,2,6,63,,9,,0.3704,,,,,,, 3070804,1963,1096,,"US","OH",,1,,2,6,63,,3,,0.6667,,,,,,, 3070805,1963,1096,,"US","CA",,1,,2,6,63,,1,,0,,,,,,, 3 As with any data analysis, we must be careful when interpreting with limited data. When a patent doesn’t seem to cite any other patents, it may be an older patent for which we have no citation information. On the other hand, more recent patents are cited less often because only newer patents can be aware of their existence. 66 CHAPTER 4 Writing basic MapReduce programs 3070806,1963,1096,,"US","PA",,1,,2,6,63,,0,,,,,,,,, 3070807,1963,1096,,"US","OH",,1,,623,3,39,,3,,0.4444,,,,,,, 3070808,1963,1096,,"US","IA",,1,,623,3,39,,4,,0.375,,,,,,, 3070809,1963,1096,,"US","AZ",,1,,4,6,65,,0,,,,,,,,, The first row contains the name of a couple dozen attributes, which are meaningful only to patent specialists. Even without understanding all the attributes, it’s still useful to have some idea of a few of them. Table 4.1 describes the first ten. 6009552 3909721 4354269 4486882 5598422 3697947 4395768 3566352 5771248 5931967 5936972 4201976 4205324 5193094 3891959 3728678 3868632 4052698 380028 3851306 3958220 3519988 3629824 3675200 3697948 437510052186895283791 4870645 6003144 5841795 46370214849976 4796261 5546409 5729647 Figure 4.1 A partial view of the patent citation data set as a graph. Each patent is shown as a vertex (node), and each citation is a directed edge (arrow). Constructing the basic template of a MapReduce program 67 Table 4.1 Definition of the first 10 attributes in the patent description data set Attribute name Content PATENT Patent number GYEAR Grant year GDATE Grant date, given as the number of days elapsed since January 1, 1960 APPYEAR Application year (available only for patents granted since 1967) COUNTRY Country of first inventor POSTATE State of first inventory (if country is U.S.) ASSIGNEE Numeric identifier for assignee (i.e., patent owner) ASSCODE One-digit (1-9) assignee type. (The assignee type includes U.S. individual, U.S. government, U.S. organization, non-U.S. individual, etc.) CLAIMS Number of claims (available only for patents granted since 1975) NCLASS 3-digit main patent class Now that we have two patent data sets, let’s write Hadoop programs to process the data. 4.2 Constructing the basic template of a MapReduce program We write most MapReduce programs in brief and as variations on a template. When writing a new MapReduce program, you generally take an existing MapReduce program and modify it until it does what you want. In this section, we write our first MapReduce program and explain its different parts. This program can serve as a template for future MapReduce programs . Our first program will take the patent citation data and invert it. For each patent, we want to find and group the patents that cite it. Our output should be similar to the following: 1 3964859,4647229 10000 4539112 100000 5031388 1000006 4714284 1000007 4766693 1000011 5033339 1000017 3908629 1000026 4043055 1000033 4190903,4975983 1000043 4091523 1000044 4082383,4055371 1000045 4290571 1000046 5918892,5525001 1000049 5996916 68 CHAPTER 4 Writing basic MapReduce programs 1000051 4541310 1000054 4946631 1000065 4748968 1000067 5312208,4944640,5071294 1000070 4928425,5009029 We have discovered that patents 5312208, 4944640, and 5071294 cited patent 1000067. For this section we won’t focus too much on the MapReduce data flow, which we’ve already covered in chapter 3. Instead we focus on the structure of a MapReduce program. We need only one file for the entire program as you can see in listing 4.1. Listing 4.1 Template for a typical Hadoop program public class MyJob extends Configured implements Tool { public static class MapClass extends MapReduceBase implements Mapper { public void map(Text key, Text value, OutputCollector output, Reporter reporter) throws IOException { output.collect(value, key); } } public static class Reduce extends MapReduceBase implements Reducer { public void reduce(Text key, Iterator values, OutputCollector output, Reporter reporter) throws IOException { String csv = ""; while (values.hasNext()) { if (csv.length() > 0) csv += ","; csv += values.next().toString(); } output.collect(key, new Text(csv)); } } public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, MyJob.class); Path in = new Path(args[0]); Path out = new Path(args[1]); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); job.setJobName("MyJob"); job.setMapperClass(MapClass.class); job.setReducerClass(Reduce.class); job.setInputFormat(KeyValueTextInputFormat.class); Constructing the basic template of a MapReduce program 69 job.setOutputFormat(TextOutputFormat.class); job.setOutputKeyClass(Text.class); job.setOutputValueClass(Text.class); job.set("key.value.separator.in.input.line", ","); JobClient.runJob(job); return 0; } public static void main(String[] args) throws Exception { int res = ToolRunner.run(new Configuration(), new MyJob(), args); System.exit(res); } } Our convention is that a single class, called MyJob in this case, completely defines each MapReduce job. Hadoop requires the Mapper and the Reducer to be their own static classes. These classes are quite small, and our template includes them as inner classes to the MyJob class. The advantage is that everything fits in one file, simplifying code management. But keep in mind that these inner classes are independent and don’t interact much with the MyJob class. Various nodes with different JVMs clone and run the Mapper and the Reducer during job execution , whereas the rest of the job class is executed only at the client machine. We investigate the Mapper and the Reducer classes in a while. Without those classes, the skeleton of the MyJob class is public class MyJob extends Configured implements Tool { public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, MyJob.class); Path in = new Path(args[0]); Path out = new Path(args[1]); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); job.setJobName("MyJob"); job.setMapperClass(MapClass.class); job.setReducerClass(Reduce.class); job.setInputFormat(KeyValueTextInputFormat.class); job.setOutputFormat(TextOutputFormat.class); job.setOutputKeyClass(Text.class); job.setOutputValueClass(Text.class); job.set("key.value.separator.in.input.line", ","); JobClient.runJob(job); return 0; } public static void main(String[] args) throws Exception { 70 CHAPTER 4 Writing basic MapReduce programs int res = ToolRunner.run(new Configuration(), new MyJob(), args); System.exit(res); } } The core of the skeleton is within the run() method, also known as the driver. The driver instantiates, configures, and passes a JobConf object named job to JobClient. runJob() to start the MapReduce job. (The JobClient class, in turn, will communicate with the JobTracker to start the job across the cluster.) The JobConf object will hold all configuration parameters necessary for the job to run. The driver needs to specify in job the input paths, the output paths, the Mapper class, and the Reducer class—the basic parameters for every job. In addition, each job can reset the default job proper- ties, such as InputFormat, OutputFormat, and so on. One can also call the set() method on the JobConf object to set up any configuration parameter. Once you pass the JobConf object to JobClient.runJob(), it’s treated as the master plan for the job. It becomes the blueprint for how the job will be run. The JobConf object has many parameters, but we don’t want to program the driver to set up all of them. The configuration files of the Hadoop installation are a good starting point. When starting a job from the command line, the user may also want to pass extra arguments to alter the job configuration. The driver can define its own set of commands and process the user arguments itself to enable the user to modify some of the configuration parameters. As this task is needed often, the Hadoop framework provides ToolRunner , Tool , and Configured to simplify it. When used together in the MyJob skeleton above, these classes enable our job to understand user-supplied options that are supported by GenericOptionsParser . For example, we have previously executed the MyJob class using this command line: bin/hadoop jar playground/MyJob.jar MyJob input/cite75_99.txt output Had we wanted to run the job only to see the mapper’s output (which you may want to do for debugging purposes), we could set the number of reducers to zero with the option -D mapred.reduce.tasks=0. bin/hadoop jar playground/MyJob.jar MyJob -D mapred.reduce.tasks=0 ➥ input/cite75_99.txt output It works even though our program doesn’t explicitly understand the -D option. By us- ing ToolRunner , MyJob will automatically support the options in table 4.2. Table 4.2 Options supported by GenericOptionsParser Option Description -conf Specify a configuration file. -D Set value for a JobConf property. -fs Specify a NameNode, can be “local”. Constructing the basic template of a MapReduce program 71 Table 4.2 Options supported by GenericOptionsParser (continued) Option Description -jt Specify a JobTracker. -files Specify a comma-separated list of files to be used with the MapReduce job. These files are automatically distributed to all task nodes to be locally available. -libjars Specify a comma-separated list of jar files to be included in the classpath of all task JVMs. -archives Specify a comma-separated list of archives to be unarchived on all task nodes. The convention for our template is to call the Mapper class MapClass and the Reducer class Reduce . The naming would seem more symmetrical if we call the Mapper class Map, but Java already has a class (interface) named Map. Both the Mapper and the Reducer extend MapReduceBase , which is a small class providing no-op implementations to the configure() and close() methods required by the two interfaces. We use the configure() and close() methods to set up and clean up the map (reduce) tasks. We won’t need to override them except for more advanced jobs. The signatures for the Mapper class and the Reducer class are public static class MapClass extends MapReduceBase implements Mapper { public void map(K1 key, V1 value, OutputCollector output, Reporter reporter) throws IOException { } } public static class Reduce extends MapReduceBase implements Reducer { public void reduce(K2 key, Iterator values, OutputCollector output, Reporter reporter) throws IOException { } } The center of action for the Mapper class is the map() method and for the Reducer class the reduce() method. Each invocation of the map() method is given a key/value pair of types K1 and V1, respectively . The key/value pairs generated by the mapper are outputted via the collect() method of the OutputCollector object. Somewhere in your map() method you need to call output.collect((K2) k, (V2) v); Each invocation of the reduce() method at the reducer is given a key of type K2 and a list of values of type V2. Note that it must be the same K2 and V2 types used in the Mapper. The reduce() method will likely have a loop to go through all the values of type V2. 72 CHAPTER 4 Writing basic MapReduce programs while (values.hasNext()) { V2? v = values.next(); ... } The reduce() method is also given an OutputCollector to gather its key/value out- put, which is of type K3/V3. Somewhere in the reduce() method you’ll call output.collect((K3) k, (V3) v); In addition to having consistent K2 and V2 types across Mapper and Reducer, you’ll also need to ensure that the key and value types used in Mapper and Reducer are con- sistent with the input format, output key class, and output value class set in the driver. The use of KeyValueTextInputFormat means that K1 and V1 must both be type Text . The driver must call setOutputKeyClass() and setOutputValueClass() with the classes of K2 and V2, respectively. Finally, all the key and value types must be subtypes of Writable , which ensures a serialization interface for Hadoop to send the data around in a distributed cluster. In fact, the key types implement WritableComparable , a subinterface of Writable. The key types need to additionally support the compareTo() method, as keys are used for sorting in various places in the MapReduce framework. 4.3 Counting things Much of what the layperson thinks of as statistics is counting, and many basic Hadoop jobs involve counting. We’ve already seen the word count example in chapter 1. For the patent citation data, we may want the number of citations a patent has received. This too is counting. The desired output would look like this: 1 2 10000 1 100000 1 1000006 1 1000007 1 1000011 1 1000017 1 1000026 1 1000033 2 1000043 1 1000044 2 1000045 1 1000046 2 1000049 1 1000051 1 1000054 1 1000065 1 1000067 3 In each record, a patent number is associated with the number of citations it has re- ceived. We can write a MapReduce program for this task. Like we said earlier, you hardly ever write a MapReduce program from scratch. You have an existing MapReduce Counting things 73 program that processes the data in a similar way. You copy that and modify it until it fits what you want. We already have a program for getting the inverted citation index. We can modify that program to output the count instead of the list of citing patents. We need the modification only at the Reducer. If we choose to output the count as an IntWritable, we need to specify IntWritable in three places in the Reducer code. We called them V3 in our previous notation. public static class Reduce extends MapReduceBase implements Reducer { public void reduce(Text key, Iterator values, OutputCollector output, Reporter reporter) throws IOException { int count = 0; while (values.hasNext()) { values.next(); count++; } output.collect(key, new IntWritable(count)); } } By changing a few lines and matching class types, we have a new MapReduce program. This program may seem a minor modification. Let’s go through another example that requires more changes, but you’ll see that the basic MapReduce program structure remains. After running the previous example, we now have a data set that counts the number of citations for each patent. A neat exercise would be to count the counts. Let’s build a histogram of the citation counts. We expect a large number of patents to have been only cited once, and a small number may have been cited hundreds of times. It would be interesting to see the distribution of the citation counts. NOTE As the patent citation data set only covers patents issued between 1975 and 1999, the citation count is necessarily an underestimate. (Citations from patents outside of that period aren’t counted.) We also don’t deal with patents that supposedly have been cited zero times. Despite these caveats, the analysis will be useful. The first step to writing a MapReduce program is to figure out the data flow. In this case, as a mapper reads a record, it ignores the patent number and outputs an interme- diate key/value pair of . The reducer will sum up the number of 1s for each citation count and output the total. After figuring out the data flow, decide on the types for the key/value pairs—K1, V1, K2, V2, K3, and V3 for the input, intermediate, and output key/value pairs. Let’s use the KeyValueTextInputFormat , which automatically breaks each input record into key/value pairs based on a separator character. The input format produces K1 and V1 74 CHAPTER 4 Writing basic MapReduce programs as Text . We choose to use IntWritable for K2, V2, K3, and V3 because we know those data must be integers and it’s more efficient to use IntWritable. Based on the data flow and the data types, you’ll be able to see the final program shown in listing 4.2 and understand what it’s doing. You can see that it’s structurally similar to the other MapReduce programs we’ve seen so far. We go into details about the program after the listing. Listing 4.2 CitationHistogram.java: count patents cited once, twice, and so on public class CitationHistogram extends Configured implements Tool { public static class MapClass extends MapReduceBase implements Mapper { private final static IntWritable uno = new IntWritable(1); private IntWritable citationCount = new IntWritable(); public void map(Text key, Text value, OutputCollector output, Reporter reporter) throws IOException { citationCount.set(Integer.parseInt(value.toString())); output.collect(citationCount, uno); } } public static class Reduce extends MapReduceBase implements Reducer { public void reduce(IntWritable key, Iterator values, OutputCollectoroutput, Reporter reporter) throws IOException { int count = 0; while (values.hasNext()) { count += values.next().get(); } output.collect(key, new IntWritable(count)); } } public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, CitationHistogram.class); Path in = new Path(args[0]); Path out = new Path(args[1]); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); job.setJobName("CitationHistogram"); job.setMapperClass(MapClass.class); job.setReducerClass(Reduce.class); job.setInputFormat(KeyValueTextInputFormat.class); job.setOutputFormat(TextOutputFormat.class); Counting things 75 job.setOutputKeyClass(IntWritable.class); job.setOutputValueClass(IntWritable.class); JobClient.runJob(job); return 0; } public static void main(String[] args) throws Exception { int res = ToolRunner.run(new Configuration(), new CitationHistogram(), args); System.exit(res); } } The class name is now CitationHistogram; all references to MyJob were changed to re- flect the new name. The main() method is almost always the same. The driver is mostly intact. The input format and output format are still KeyValueTextInputFormat and TextOutputFormat , respectively. The main change is that the output key class and the output value class are now IntWritable, to reflect the new type for K2 and V2. We’ve also removed this line: job.set("key.value.separator.in.input.line", ","); It sets the separator character used by KeyValueTextInputFormat to break each in- put line into a key/value pair. Previously it was a comma for processing the original patent citation data. By not setting this property it defaults to the tab character, which is appropriate for the citation count data. The data flow for this mapper is similar to that of the previous mappers, only here we’ve chosen to define and use a couple class variables—citationCount and uno. public static class MapClass extends MapReduceBase implements Mapper { private final static IntWritable uno = new IntWritable(1); private IntWritable citationCount = new IntWritable(); public void map(Text key, Text value, OutputCollector output, Reporter reporter) throws IOException { citationCount.set(Integer.parseInt(value.toString())); output.collect(citationCount, uno); } } The map() method has one extra line for setting citationCount, which is for type casting. The reason for defining citationCount and uno in the class rather than inside the method is purely one of efficiency. The map() method will be called as many times as there are records (in a split, for each JVM). Reducing the number of objects created inside the map() method can increase performance and reduce garbage collection . As we pass citationCount and uno to output.collect(), 76 CHAPTER 4 Writing basic MapReduce programs we’re relying on the output.collect() method’s contract to not modify those two objects.4 The reducer sums up the values for each key. It seems inefficient because we know all values are 1s (uno, to be exact). Why do we need to sum the count? We’ve chosen this route because it will be easier for us later if we choose to add a combiner to enhance performance. Unlike MapClass, the call to output.collect() in Reduce instantiates a new IntWritable rather than reuse an existing one. output.collect(key, new IntWritable(count)); We can improve performance by using an IntWritable class variable. But the number of times reduce() is called is much smaller in this particular program, probably no more than a thousand times (across all reducers). We don’t have much need to opti- mize this particular code. Running the MapReduce job on the citation count data will show the following result. As we suspect, a large number (900K+) of patents have only one citation, whereas some have hundreds of citations. The most popular patent has 779 citations. 1 921128 2 552246 3 380319 4 278438 5 210814 6 163149 7 127941 8 102155 9 82126 10 66634 ... 411 1 605 1 613 1 631 1 633 1 654 1 658 1 678 1 716 1 779 1 As this histogram output is only several hundred lines long, we can put it into a spread- sheet and plot it. Figure 4.2 shows the number of patents at various citation frequen- cies. The plot is on a log-log scale. When a distribution shows as a line in a log-log plot, it’s considered to be a power law distribution . The citation count histogram seems to fit the description, although its approximately parabolic curvature also suggests a log- normal distribution . As you’ve seen in our examples so far, a MapReduce program is often not very big, and you can keep a certain structure across them to simplify development. Most of the work is in thinking through the data flow. 4 We see in section 5.1.3 that this reliance will forbid the ChainMapper from using pass-by-reference. Adapting for Hadoop’s API changes 77 4.4 Adapting for Hadoop’s API changes One of the main design goals driving toward Hadoop’s major 1.0 release is a stable and extensible MapReduce API. As of this writing, version 0.20 is the latest release and is considered a bridge between the older API (that we use throughout this book) and this upcoming stable API. The 0.20 release supports the future API while maintaining backward-compatibility with the old one by marking it as deprecated. Future releases after 0.20 will stop supporting the older API. As of this writing, we don’t recommend jumping into the new API yet for a couple reasons: 1 Many of Hadoop’s own library classes in 0.20 aren’t rewritten under the new API yet. You won’t be able to use those classes if your MapReduce code uses the new API in 0.20. 2 Many still consider the most production-ready and stable version of Hadoop as of this writing to be 0.18.3. Some users are warming up to version 0.20, but we suggest you wait a little longer before going full production with it.5 By the time you read this the situation may be different. In this section we cover the changes the new API presents. Fortunately, almost all the changes affect only the basic MapReduce template. We rewrite the template under the new API to enable you to use it in the future. 5 You may be wondering about version 0.19. The general consensus is that its initial release was problematic and full of bugs. Some minor releases tried fixing the problems, but the community seems to want to skip straight to 0.20 instead. 1 10 100 1000 10000 100000 1000000 1 10 100 1000 Number of citations Number of patents Figure 4.2 Plotting the number of patents at different citation frequencies. Many patents have one citation (or not at all, which is not shown on this graph). Some patents have hundreds of citations. On a log-log graph, this looks close enough to a straight line to be considered a power-law distribution. 78 CHAPTER 4 Writing basic MapReduce programs The first thing you’ll notice in the new API is that many classes in org.apache. hadoop.mapred have been moved elsewhere. Many of them are now at org.apache. hadoop.mapreduce, and the library classes are under one of the packages in org. apache.hadoop.mapreduce.lib. After you’ve moved to the new API, you shouldn’t have any import statements (or full references) to any classes under org.apache. hadoop.mapred, all of which are to be deprecated. The most meaningful change in the new API is the introduction of context objects. Its most immediate impact is to replace the OutputCollector and Reporter objects used in the map() and reduce() methods. You now output key/value pairs by calling Context.write() instead of OutputCollector.collect(). The long-term consequences are to unify communication between your code and the MapReduce framework, and to stabilize the Mapper and Reducer API such that the basic method signatures will not change when new functionalities are added. New functionalities will only be additional methods on the context objects. Programs written before the introduction of those functionalities will be unaware of the new methods, and they will continue to compile and run against the newer releases. The new map() and reduce() methods are contained in new abstract classes Mapper and Reducer, respectively. They replace the Mapper and Reducer interfaces in the original API (org.apache.hadoop.mapred.Mapper and org.apache.hadoop.mapred. Reducer). The new abstract classes also replace the MapReduceBase class, which has been deprecated. The new map() and reduce() methods have a couple more minor changes. They can throw InterruptedException instead of only IOException. In addition, the reduce() method no longer accepts the list of values as an Iterator but as an Iterable, which is easier to iterate through using Java’s foreach syntax. We can summarize the changes we’ve discussed so far in the method signatures for MapClass and Reduce. Recall the signatures under the original API: public static class MapClass extends MapReduceBase implements Mapper { public void map(K1 key, V1 value, OutputCollector output, Reporter reporter) throws IOException { } } public static class Reduce extends MapReduceBase implements Reducer { public void reduce(K2 key, Iterator values, OutputCollector output, Reporter reporter) throws IOException { } } The new API has simplified them somewhat: public static class MapClass extends Mapper { public void map(K1 key, V1 value, Context context) throws IOException, InterruptedException { } Adapting for Hadoop’s API changes 79 } public static class Reduce extends Reducer { public void reduce(K2 key, Iterable values, Context context) throws IOException, InterruptedException { } } You also need to change a few things in the driver to support the new API. JobConf and JobClient classes have been replaced. Their functionalities have been pushed to the Configuration class (which was originally the parent class of JobConf) and a new class Job. The Configuration class purely configures a job, whereas the Job class defines and controls the execution of a job. Methods such as setOutputKeyClass() and setOutputValueClass() have moved from JobConf to Job. A job’s construction and submission for execution are now under Job. Originally you would construct a job using JobConf: JobConf job = new JobConf(conf, MyJob.class); job.setJobName("MyJob"); Now it’s done through Job: Job job = new Job(conf, "MyJob"); job.setJarByClass(MyJob.class); Previously JobClient submitted a job for execution: JobClient.runJob(job); Now it’s also done through Job: System.exit(job.waitForCompletion(true)?0:1); Listing 4.3 is the template program from listing 4.1 rewritten to use the new API in Hadoop 0.20. It incorporates all the changes we’ve mentioned in this section. Listing 4.3 Template for basic Hadoop program (listing 4.1) rewritten for new API v 0.20 public class MyJob extends Configured implements Tool { public static class MapClass extends Mapper { public void map(LongWritable key, Text value, Context context) throws IOException, InterruptedException { String[] citation = value.toString().split(","); context.write(new Text(citation[1]), new Text(citation[0])); } } public static class Reduce extends Reducer { public void reduce(Text key, Iterable values, Context context) throws IOException, InterruptedException { String csv = ""; 80 CHAPTER 4 Writing basic MapReduce programs for (Text val:values) { if (csv.length() > 0) csv += ","; csv += val.toString(); } context.write(key, new Text(csv)); } } public int run(String[] args) throws Exception { Configuration conf = getConf(); Job job = new Job(conf, "MyJob"); job.setJarByClass(MyJob.class); Path in = new Path(args[0]); Path out = new Path(args[1]); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); job.setMapperClass(MapClass.class); job.setReducerClass(Reduce.class); job.setInputFormatClass(TextInputFormat.class); q job.setOutputFormatClass(TextOutputFormat.class); job.setOutputKeyClass(Text.class); job.setOutputValueClass(Text.class); System.exit(job.waitForCompletion(true)?0:1); return 0; } public static void main(String[] args) throws Exception { int res = ToolRunner.run(new Configuration(), new MyJob(), args); System.exit(res); } } The code performs the same inverted indexing function as listing 4.1, but under the 0.20 API. Unfortunately, the KeyValueTextInputFormat class we had used in listing 4.1 hasn’t been ported to the new API as of version 0.20. We have to rewrite the tem- plate using TextInputFormat q. We expect all Hadoop classes to support the new API when version 0.21 is released. To keep presentation of examples in the rest of this book unified, we continue to use the API before 0.20. 4.5 Streaming in Hadoop We have been using Java to write all our Hadoop programs. Hadoop supports other languages via a generic API called Streaming . In practice, Streaming is most useful for writing simple, short MapReduce programs that are more rapidly developed in a scripting language that can take advantage of non-Java libraries. Hadoop Streaming interacts with programs using the Unix streaming paradigm. Inputs come in through STDIN and outputs go to STDOUT. Data has to be text based and each line is considered a record. Note that this is exactly how many Iterable allows foreach looping Compatible InputFormat class Streaming in Hadoop 81 Unix commands work, and Hadoop Streaming enables those commands to be used as mappers and reducers. If you’re familiar with using Unix commands, such as wc, cut, or uniq for data processing, you can apply them to large data sets using Hadoop Streaming. The overall data flow in Hadoop Streaming is like a pipe where data streams through the mapper, the output of which is sorted and streamed through the reducer. In pseudo-code using Unix’s command line notation, it’s cat [input_file] | [mapper] | sort | [reducer] >[output_file] The following examples will illustrate how to use Streaming with Unix commands. 4.5.1 Streaming with Unix commands In the first example, let’s get a list of cited patents in cite75_99.txt. bin/hadoop jar contrib/streaming/hadoop-0.19.1-streaming.jar ➥ -input input/cite75_99.txt ➥ -output output ➥ -mapper 'cut -f 2 -d ,' ➥ -reducer 'uniq' That’s it! It’s a one-line command. Let’s see what each part of the command does. The Streaming API is in a contrib package at contrib/streaming/hadoop-*- streaming.jar. The first part and the -input and the -output arguments specify that we’re running a Streaming program with the corresponding input and output file/directory. The mapper and reducer are specified as arguments in quotes. We see that for the mapper we use the Unix cut command to extract the second column, where columns are separated by commas. In the citation data set this column is the patent number of a cited patent. These patent numbers are then sorted and passed to the reducer. The uniq command at the reducer will remove all duplicates in the sorted data. The output of this command is "CITED" 1 10000 100000 1000006 ... 999973 999974 999977 999978 999983 The first row has the column descriptor “CITED” from the original file. Note that the rows are sorted lexicographically because Streaming processes everything as text and doesn’t know other data types. After getting the list of cited patents, we may want to know how many are there. Again we can use Streaming to quickly get a count, using the Unix command wc –l. 82 CHAPTER 4 Writing basic MapReduce programs bin/hadoop jar contrib/streaming/hadoop-0.19.1-streaming.jar ➥ -input output ➥ -output output_a ➥ -mapper 'wc -l' ➥ -D mapred.reduce.tasks=0 Here we use wc –l as the mapper to count the number of records in each split. Hadoop Streaming (since version 0.19.0) supports the GenericOptionsParser . The -D argument is used for specifying configuration properties. We want the mapper to directly output the record count without any reducer, so we set mapred.reduce.tasks to 0 and don’t specify the -reducer option at all. 6 The final count is 3258984. More than 3 million patents have been cited according to our data. 4.5.2 Streaming with scripts We can use any executable script that processes a line-oriented data stream from STDIN and outputs to STDOUT with Hadoop Streaming. For example, the Python script in listing 4.4 randomly samples data from STDIN. For those who don’t know Python, the program has a for loop that reads STDIN one line at a time. For each line, we choose a random integer between 1 and 100 and check against the user-given argu- ment (sys.argv[1]). The comparison determines whether to pass that line on to the output or ignore it. You can use the script in Unix to uniformly sample a line-oriented data file, for example: cat input.txt | RandomSample.py 10 >sampled_output.txt The preceding command calls the Python script with an argument of 10; sampled_ output.txt will have (approximately) 10 percent of the records in input.txt. We can in fact specify any integer between 1 and 100 to get the corresponding percentage of data in the output. Listing 4.4 RandomSample.py: a Python script printing random lines from STDIN #!/usr/bin/env python import sys, random for line in sys.stdin: if (random.randint(1,100) <= int(sys.argv[1])): print line.strip() We can apply the same script in Hadoop to get a smaller sample of a data set. A sampled data set is often useful for development purposes, as you can run your Hadoop program on the sampled data in standalone or pseudo-distributed mode to quickly debug and iterate. Also, when you’re looking for some “descriptive” 6 You may notice that this approach counts the number of records in each split, not the entire file. With a bigger file, or multiple files, the user will have to sum up the counts herself to get the overall total. To fully automate a complete counting, the user will have to write a script at the reducer to sum up all the partial counts. Streaming in Hadoop 83 information about your data, the speed and convenience in processing a smaller data set generally outweigh any loss of precision. Finding data clusters is one ex- ample of such descriptive information. Optimized implementations of a variety of clustering algorithms are readily available in R, MATLAB, and other packages. It makes a lot more sense to sample down the data and apply some standard software package, instead of trying to process all data using some distributed clustering al- gorithms in Hadoop. WARNING The loss of precision from computing on a sampled data set may or may not be important. It depends on what you’re trying to compute and the distribution of your data set. For example, it’s usually fine to compute an average from a sampled data set, but if the data set is highly skewed and the average is dominated by a few values, sampling can be problematic. Similarly, clustering on a sampled data set is fine if it’s used only to get a general understanding of the data. If you were looking for small, anomalous clusters, sampling may get rid of them. For functions such as maximum and minimum, it’s not a good idea to apply them to sampled data. Running RandomSample.py using Streaming is like running Unix commands using Streaming, the difference being that Unix commands are already available on all nodes in the cluster, whereas RandomSample.py is not. Hadoop Streaming supports a -file option to package your executable file as part of the job submission. 7 Our command to execute RandomSample.py is bin/hadoop jar contrib/streaming/hadoop-0.19.1-streaming.jar ➥ -input input/cite75_99.txt ➥ -output output ➥ -mapper 'RandomSample.py 10' ➥ -file RandomSample.py ➥ -D mapred.reduce.tasks=1 In specifying the mapper to be 'RandomSample.py 10' we’re sampling at 10 percent. Note that we’ve set the number of reducers (mapred.reduce.tasks) to 1. As we haven’t specified any particular reducer, it will use the default IdentityReducer. As its name implies, IdentityReducer passes its input straight to output. In this case we can set the number of reducers to any non-zero value to get an exact number of output files. Alternatively, we can set the number of reducers to 0, and let the number of out- put files be the number of mappers. This is probably not ideal for the sampling task as each mapper’s output is only a small fraction of the input, and we may end up with a number of small files. We can easily correct that later using the HDFS shell command getmerge or other file manipulations to arrive at the right number of output files. The approach to use is more or less a personal preference. 7 It’s also implicitly assumed that you have installed the Python language on all the nodes in your cluster. 84 CHAPTER 4 Writing basic MapReduce programs The random sampling script was implemented in Python, although any scripting language that works with STDIN and STDOUT would work. For illustration we’ve rewritten the same script in PHP 8 (listing 4.5). Execute this Stream script with bin/hadoop jar contrib/streaming/hadoop-0.19.1-streaming.jar ➥ -input input/cite75_99.txt ➥ -output output ➥ -mapper 'php RandomSample.php 10' ➥ -file RandomSample.php ➥ -D mapred.reduce.tasks=1 Listing 4.5 RandomSample.php.: a PHP script printing random lines from STDIN to close the opening bracket is considered static content to be outputted. When using PHP as a pure scripting language, you need to be careful that you leave no whitespaces outside the brackets. Otherwise they will be outputted and may cause unintended behavior that is hard to debug. (It would appear whitespaces were introduced in the output data out of nowhere.) It’s easy to ensure that there’s no whitespaces before the opening bracket , as ending whitespaces don’t grab attention. When using a file as a PHP script, it’s safer to omit the closing bracket ?>. The PHP interpreter will quietly read everything till the end-of-file as PHP commands rather than static content. Streaming in Hadoop 85 global maximum is the maximum over the maxima of the splits. That sounded like a mouthful, but a simple example will make it clear. If we have four records X1, X2, X3, and X4, and they’re divided into two splits (X1, X2) and (X3, X4), we can find the maximum over all four records by looking at the maximum of each split, or max(X1,X2,X3,X4) = max(max(X1,X2), max(X3,X4)) Our strategy is to have mapper calculate the maximum over its individual split. Each mapper will output a single value at the end. We have a single reducer that looks at all those values and outputs the global maximum. Listing 4.6 depicts the Python script for a mapper to compute the maximum over a split. Listing 4.6 AttributeMax.py: Python script to find maximum value of an attribute #!/usr/bin/env python import sys index = int(sys.argv[1]) max = 0 for line in sys.stdin: fields = line.strip().split(",") if fields[index].isdigit(): val = int(fields[index]) if (val > max): max = val else: print max The script is not complicated. It has a for loop to read one record at a time. It tokenizes the record into fields and updates the maximum if the user-specified field is bigger. Note that the mapper doesn’t output any value until the end, when it sends out the maximum value of the entire split. This is different from what we’ve seen before, where each record sends out one or more intermediate records to be processed by the reducers. Given the parsimonious output of the mapper, we can use the default IdentityReducer to record the (sorted) output of the mappers. bin/hadoop jar contrib/streaming/hadoop-0.19.1-streaming.jar ➥ -input input/apat63_99.txt ➥ -output output ➥ -mapper 'AttributeMax.py 8' ➥ -file playground/AttributeMax.py ➥ -D mapred.reduce.tasks=1 The mapper is 'AttributeMax.py 8'. It outputs the maximum of the ninth column in a split. The single reducer collects all the mapper outputs. Given seven mappers, the final output of the above command is this: 0 260 306 86 CHAPTER 4 Writing basic MapReduce programs 348 394 706 868 Each line records the maximum over a particular split. We see that one split has zero claims in all its records. This sounds suspicious until we recall that the claim count at- tribute is not available for patents before 1975. We see that our mapper is doing the right thing. We can use a reducer that outputs the maximum over the values outputted by the mappers. We have an interesting situation here, due to the distributive property of maximum, where we can also use AttributeMax.py as the reducer. Only now the reducer is trying to find the maximum in the “first” column. bin/hadoop jar contrib/streaming/hadoop-0.18.1-streaming.jar ➥ -input input/apat63_99.txt ➥ -output output ➥ -mapper 'AttributeMax.py 8' ➥ -reducer 'AttributeMax.py 0' ➥ -file AttributeMax.py ➥ -D mapred.reduce.tasks=1 The output of the above command should be a one-line file, and you’ll find the maxi- mum number of claims in a patent to be 868. Classes of aggregation functions We use aggregation functions to compute descriptive statistics. They’re generally grouped into three classes: distributive , algebraic , and holistic . The maximum function is an example of a distributive function. Other distributive functions include minimum , sum , and count . As the name implies, distributive functions have distributive properties. Similar to the maximum function, you can globally compute these functions by iteratively applying them to smaller chunks of data. Another class of aggregation functions is the algebraic functions. Examples of this class include average and variance . They don’t follow the distributive property, and their derivation will require some “algebraic” computation over simpler functions. We get into examples of this in the next section. Finally, functions such as median and K smallest/largest value belong to the holistic class of aggregation functions. Readers interested in a challenge should try to implement the median function in an efficient manner using Hadoop. 4.5.3 Streaming with key/value pairs At this point you may wonder what happened to the key/value pair way of encoding records. Our discussion on Streaming so far talks about each record as an atomic unit Streaming in Hadoop 87 rather than as composed of a key and a value. The truth is that Streaming works on key/value pairs just like the standard Java MapReduce model. By default, Streaming uses the tab character to separate the key from the value in a record. When there’s no tab character, the entire record is considered the key and the value is empty text. For our data sets, which have no tab character, this provides the illusion that we’re process- ing each individual record as a whole unit. Furthermore, even if the records do have tab characters in them, the Streaming API will only shuffle and sort the records in a different order. As long as our mapper and reducer work in a record-oriented way, we can maintain the record-oriented illusion. Working with key/value pairs allows us to take advantage of the key-based shuffling and sorting to create interesting data analyses. To illustrate key/value pair processing using Streaming, we can write a program to find the maximum number of claims in a patent for each country. This would differ from AttributeMax.py in that this is trying to find the maximum for each key, rather than a maximum across all records. Let’s make this exercise more interesting by computing the average rather than finding the maximum. (As we see, Hadoop already includes a package called Aggregate that contains classes that help find the maximum for each key.) First, let’s examine how key/value pairs work in the Streaming API for each step of the MapReduce data flow. 1 As we’ve seen, the mapper under Streaming reads a split through STDIN and extracts each line as a record. Your mapper can choose to interpret each input record as a key/value pair or a line of text. 2 The Streaming API will interpret each line of your mapper’s output as a key/ value pair separated by tab. Similar to the standard MapReduce model, we apply the partitioner to the key to find the right reducer to shuffle the record to. All key/value pairs with the same key will end up at the same reducer. 3 At each reducer, key/value pairs are sorted according to the key by the Streaming API. Recall that in the Java model, all key/value pairs of the same key are grouped together into one key and a list of values. This group is then presented to the reduce() method. Under the Streaming API your reducer is responsible for performing the grouping. This is not too bad as the key/value pairs are already sorted by key. All records of the same key are in one contiguous chunk. Your reducer will read one line at a time from STDIN and will keep track of the new keys. 4 For all practical purposes, the output (STDOUT) of your reducer is written to a file directly. Technically a no-op step is taken before the file write. In this step the Streaming API breaks each line of the reducer’s output by the tab character and feeds the key/value pair to the default TextOutputFormat , which by default re-inserts the tab character before writing the result to a file. Without tab characters in the reducer’s output it will show the same 88 CHAPTER 4 Writing basic MapReduce programs no-op behavior. You can reconfigure the default behavior to do something different, but it makes sense to leave it as a no-op and push the processing into your reducer. To understand the data flow better, we write a Streaming program to compute the average number of claims for each country. The mapper will extract the country and the claims count for each patent and package them as a key/value pair. In accord with the default Streaming convention, the mapper outputs this key/value pair with a tab character to separate them. The Streaming API will pick up the key and the shuffling will guarantee that all claim counts of a country will end up at the same reducer. We can see the Python code in listing 4.7. For each record, the mapper extracts the coun- try (fields[4][1:-1]) as key and the claims count (fields[8]) as value. An extra concern with our data set is that missing values do exist. We’ve added a conditional statement to skip over records with missing claim counts. Listing 4.7 AverageByAttributeMapper.py: output country and claim count of patents #!/usr/bin/env python import sys for line in sys.stdin: fields = line.split(",") if (fields[8] and fields[8].isdigit()): print fields[4][1:-1] + "\t" + fields[8] Before writing the reducer, let’s run the mapper in two situations: without any reducer, and with the default IdentityReducer. It’s a useful approach now for learning as we can see exactly what’s being outputted by the mapper (by using no reducer) and what’s being inputted into the reducer (by using IdentityReducer). You’ll find this handy later when debugging your MapReduce program. You can at least check if the mapper is outputting the proper data and if the proper data is being sent to the reducer. First let’s run the mapper without any reducer. bin/hadoop jar contrib/streaming/hadoop-0.19.1-streaming.jar ➥ -input input/apat63_99.txt ➥ -output output ➥ -file playground/AverageByAttributeMapper.py ➥ -mapper 'AverageByAttributeMapper.py' ➥ -D mapred.reduce.tasks=0 The output should consist of lines where a country code is followed by a tab followed by a numeric count. The order of the output records is not sorted by (the new) key. In fact, it’s in the same order as the order of the input records, although that’s not obvi- ous from looking at the output. The more interesting case is to use the IdentityReducer with a non-zero number of reducers. We see how the shuffled and sorted records are presented to the reducer. To keep it simple let’s try a single reducer by setting -D mapred.reduce.tasks=1 and see the first 32 records. Streaming in Hadoop 89 AD 9 AD 12 AD 7 AD 28 AD 14 AE 20 AE 7 AE 35 AE 11 AE 12 AE 24 AE 4 AE 16 AE 26 AE 11 AE 4 AE 23 AE 12 AE 16 AE 10 AG 18 AG 12 AG 8 AG 14 AG 24 AG 20 AG 7 AG 3 AI 10 AM 18 AN 5 AN 26 Under the Streaming API, the reducer will see these text data in STDIN. We have to code our reducer to recover the key/value pairs by breaking each line at the tab character. Sort- ing has “grouped” together records of the same key. As you read each line from STDIN, you’ll be responsible for keeping track of the boundary between records of different keys. Note that although the keys are sorted, the values don’t follow any particular order. Finally, the reducer must perform its stated computation, which in this case is calculating the average value across a key. Listing 4.8 gives the complete reducer in Python . Listing 4.8 AverageByAttributeReducer.py #!/usr/bin/env python import sys (last_key, sum, count) = (None, 0.0, 0) for line in sys.stdin: (key, val) = line.split("\t") if last_key and last_key != key: print last_key + "\t" + str(sum / count) (sum, count) = (0.0, 0) last_key = key sum += float(val) count += 1 print last_key + "\t" + str(sum / count) The program keeps a running sum and count for each key. When it detects a new key in the input stream or the end of the file, it computes the average for the previous key and sends it to STDOUT. After running the entire MapReduce job, we can easily check the correctness of the first few results. AD 14.0 AE 15.4 AG 13.25 AI 10.0 AM 18.0 AN 9.625 90 CHAPTER 4 Writing basic MapReduce programs NOTE For those interested, the NBER website from where we get the patent data also has a file (list_of_countries.txt) that shows the full country name for each country code. Looking at the output of our job and the country codes, we see that Andorra (AD) patents have an average 14 claims. Arab Emirates (AE) patents average 15.4 claims. Antigua and Barbuda (AG) patents average 13.25 claims, and so forth. 4.5.4 Streaming with the Aggregate package Hadoop includes a library package called Aggregate that simplifies obtaining aggre- gate statistics of a data set. This package can simplify the writing of Java statistics collec- tors, especially when used with Streaming, which is the focus of this section.9 The Aggregate package under Streaming functions as a reducer that computes aggregate statistics. You only have to provide a mapper that processes records and sends out a specially formatted output. Each line of the mapper’s output looks like function:key\tvalue The output string starts with the name of a value aggregator function (from the set of predefined functions available in the Aggregate package). A colon and a tab-separated key/value pair follows. The Aggregate reducer applies the function to the set of values for each key. For example, if the function is LongValueSum , then the output is the sum of values for each key. (As the function name implies, each value is treated as a Java long type.) If the function is LongValueMax, then the output is the maximum value for each key. You can see the list of aggregator functions supported in the Aggregate package in table 4.3. Table 4.3 List of value aggregator functions supported by the Aggregate package Value aggregator Description DoubleValueSum Sums up a sequence of double values. LongValueMax Finds the maximum of a sequence of long values. LongValueMin Finds the minimum of a sequence of long values. LongValueSum Sums up a sequence of long values. StringValueMax Finds the lexicographical maximum of a sequence of string values. StringValueMin Finds the lexicographical minimum of a sequence of string values. UniqValueCount Finds the number of unique values (for each key). ValueHistogram Finds the count, minimum, median, maximum, average, and standard deviation of each value. (See text for further explanation.) 9 Using the Aggregate package in Java is explained in http://hadoop.apache.org/core/docs/current/api/ org/apache/hadoop/mapred/lib/aggregate/package-summary.html. Streaming in Hadoop 91 Let’s go through an exercise using the Aggregate package to see how easy it is. We want to count the number of patents granted each year. We can approach this prob- lem in a way similar to the word counting example we saw in chapter 1. For each record, our mapper will output the grant year as the key and a “1” as the value. The reducer will sum up all the values (“1”s) to arrive at a count. Only now we’re using Streaming with the Aggregate package. Our result will be the simple mapper shown in listing 4.9. Listing 4.9 AttributeCount.py #!/usr/bin/env python import sys index = int(sys.argv[1]) for line in sys.stdin: fields = line.split(",") print "LongValueSum:" + fields[index] + "\t" + "1" AttributeCount.py works for any CSV-formatted input file. The user only has to specify the column index to count the number of records for each attribute in that column. The print statement has the main “action” of this short program. It tells the Aggregate package to sum up all the values (of 1) for each key, defined as the user- specified column (index). To count the number of patents granted each year, we run this Streaming program with the Aggregate package, telling the mapper to use the second column (index = 1) of the input file as the attribute of interest. bin/hadoop jar contrib/streaming/hadoop-0.19.1-streaming.jar ➥ -input input/apat63_99.txt ➥ -output output ➥ -file AttributeCount.py ➥ -mapper 'AttributeCount.py 1' ➥ -reducer aggregate You’ll find most of the options of running the Streaming program familiar. The main thing to point out is that we’ve specified the reducer to be 'aggregate'. This is the signal to the Streaming API that we’re using the Aggregate package. The output of the MapReduce job (after sorting) is "GYEAR" 1 1963 45679 1964 47375 1965 62857 ... 1996 109645 1997 111983 1998 147519 1999 153486 The first row is anomalous because the first row of the input data is a column descrip- tion. Otherwise the MapReduce job neatly outputs the patent count for each year. As 92 CHAPTER 4 Writing basic MapReduce programs shown in figure 4.3, we can plot the data to visualize it better. You’ll see that it has a mostly steady upward trend. Looking at the list of functions in the Aggregate package in table 4.3, you’ll find that most of them are combinations of maximum, minimum, and sum for atomic data type. (For some reason DoubleValueMax and DoubleValueMin aren’t supported. They would be trivial modifications of LongValueMax and LongValueMin and an added advantage.) UniqValueCount and ValueHistogram are slightly different and we look at some examples of how to use them. UniqValueCount gives the number of unique values for each key. For example, we may want to know whether more countries are participating in the U.S. patent system over time. We can examine this by looking at the number of countries with patents granted each year. We use a straightforward wrapper of UniqValueCount in listing 4.10 and apply it to the year and country columns of apat63_99.txt (column index of 1 and 4, respectively). bin/hadoop jar contrib/streaming/hadoop-0.19.1-streaming.jar ➥ -input input/apat63_99.txt ➥ -output output ➥ -file UniqueCount.py ➥ -mapper 'UniqueCount.py 1 4' ➥ -reducer aggregate In the output we get one record for each year. Plotting it gives us figure 4.4. We can see that the increasing number of patents granted from 1960 to 1990 (from figure 4.3) didn’t come from more countries (figure 4.4). The same number of countries had filed more. 0 20000 40000 60000 80000 100000 120000 140000 160000 180000 1960 1965 1970 1975 1980 1985 1990 1995 2000 Year Figure 4.3 Using Hadoop to count patents published each year and Excel to plot the result. This analysis using Hadoop quickly shows the annual patent output to have almost quadrupled in 40 years. Streaming in Hadoop 93 Listing 4.10 UniqueCount.py: a wrapper around the UniqValueCount function #!/usr/bin/env python import sys index1 = int(sys.argv[1]) index2 = int(sys.argv[2]) for line in sys.stdin: fields = line.split(",") print "UniqValueCount:" + fields[index1] + "\t" + fields[index2] The aggregate function ValueHistogram is the most ambitious function in the Aggre- gate package. For each key, it outputs the following: 1 The number of unique values 2 The minimum count 3 The median count 4 The maximum count 5 The average count 6 The standard deviation In its most general form, it expects the output of the mapper to have the form ValueHistogram:key\tvalue\tcount We specify the function ValueHistogram followed by a colon, followed by a tab- separated key, value, and count triplet. The Aggregate reducer outputs the six statistics above for each key. Note that for everything except the first statistics (number of unique values) the counts are summed over each key/value pair. Outputting two records from your mapper as ValueHistogram:key_a\tvalue_a\t10 ValueHistogram:key_a\tvalue_a\t20 0 20 40 60 80 100 120 1960 1965 1970 1975 1980 1985 1990 1995 2000 Year Figure 4.4 The number of countries with U.S. patents granted in each year. We performed the computation with a MapReduce job and graphed the result with Excel. 94 CHAPTER 4 Writing basic MapReduce programs is no different than outputting a single record with the sum ValueHistogram:key_a\tvalue_a\t30 A useful variation is for the mapper to only output the key and value, without the count and the tab character that goes with it. ValueHistogram automatically assumes a count of 1 in this case. Listing 4.11 shows a trivial wrapper around ValueHistogram. Listing 4.11 ValueHistogram.py: wrapper around Aggregate package’s ValueHistogram #!/usr/bin/env python import sys index1 = int(sys.argv[1]) index2 = int(sys.argv[2]) for line in sys.stdin: fields = line.split(",") print "ValueHistogram:" + fields[index1] + "\t" + fields[index2] We run this program to find the distribution of countries with patents granted for each year. bin/hadoop jar contrib/streaming/hadoop-0.19.1-streaming.jar ➥ -input input/apat63_99.txt ➥ -output output ➥ -file ValueHist.py ➥ -mapper 'ValueHist.py 1 4' ➥ -reducer aggregate The output is a tab-separated value (TSV) file with seven columns. The first column, the year of patent granted, is the key. The other six columns are the six statistics the ValueHistogram is set to compute. A partial view of the output is here (we skip the first two rows for formatting reasons): 1964 58 1 7 38410 816.8103448275862 4997.413601595352 1965 67 1 5 50331 938.1641791044776 6104.779230296307 1966 71 1 5 54634 963.4507042253521 6443.625995189338 1967 68 1 8 51274 965.4705882352941 6177.445623039149 1968 71 1 7 45781 832.4507042253521 5401.229955880634 1969 68 1 8 50394 993.5147058823529 6080.713518728092 1970 72 1 7 47073 894.8472222222222 5527.883233761672 1971 74 1 9 55976 1058.337837837838 6492.837390992137 The first column after the year is the number of unique values. This is exactly the same as the output of UniqValueCount. The second, third, and fourth columns are the mini- mum , median , and maximum , respectively. For the patent data set we used, we see that (for every year) the country receiving the fewest granted patents (other than 0) received 1. Looking specifically at the output for 1964, the country receiving the most patents received 38410 patents, whereas half the countries received less than 7 patents. The average number of patents a country received in 1964 is 816.8 with a standard devi- ation of 4997.4. Needless to say, the number of patents granted to each country is highly skewed , given the discrepancy between the median (7) and the average (816.8). Improving performance with combiners 95 We’ve seen how using the Aggregate package under Streaming is a simple way to get some popular metrics. It’s a great demonstration of Hadoop’s power in simplifying the analysis of large data sets. 4.6 Improving performance with combiners We saw in AverageByAttributeMapper.py and AverageByAttributeReducer.py (listings 4.7 and 4.8) how to compute the average for each attribute. The mapper reads each record and outputs a key/value pair for the record’s attribute and count. It shuffles the key/value pairs across the network, and the reducer computes the aver- age for each key. In our example of computing the average number of claims for each country’s patents, we see at least two efficiency bottlenecks: 1 If we have 1 billion input records, the mappers will generate 1 billion key/ value pairs that will be shuffled across the network. If we were computing a function such as maximum, it’s obvious that the mapper only has to output the maximum for each key it has seen. Doing so would reduce network traffic and increase performance. For a function such as average , it’s a bit more complicated, but we can still redefine the algorithm such that for each mapper only one record is shuffled for each key. 2 Using country from the patent data set as key illustrates data skew . The data is far from uniformly distributed, as a significant majority of the records would have U.S. as the key. Not only does every key/value pair in the input map to a key/value pair in the intermediate data, most of the intermediate key/value pairs will end up at a single reducer, overwhelming it. Hadoop solves these bottlenecks by extending the MapReduce framework with a com- biner step in between the mapper and reducer. You can think of the combiner as a helper for the reducer. It’s supposed to whittle down the output of the mapper to lessen the load on the network and on the reducer. If we specify a combiner, the MapReduce framework may apply it zero, one, or more times to the intermediate data. In order for a combiner to work, it must be an equivalent transformation of the data with respect to the reducer. If we take out the combiner, the reducer’s output will remain the same. Furthermore, the equivalent transformation property must hold when the combiner is applied to arbitrary subsets of the intermediate data. If the reducer only performs a distributive function, such as maximum, minimum, and summation (counting), then we can use the reducer itself as the combiner. But many useful functions aren’t distributive. We can rewrite some of them, such as averaging to take advantage of a combiner. The averaging approach taken by AverageByAttributeMapper.py is to output only each key/value pair. AverageByAttributeReducer.py will count the number of key/value pairs it receives and sum up their values, in order for a single final division to compute the average. The main obstacle to using a combiner is the counting operation, as the reducer assumes the number of key/value pairs it receives is the 96 CHAPTER 4 Writing basic MapReduce programs number of key/value pairs in the input data. We can refactor the MapReduce program to track the count explicitly. The combiner becomes a simple summation function with the distributive property. Let’s first refactor the mapper and reducer before writing the combiner, as the operation of the MapReduce job must be correct even without a combiner. We write the new averaging program in Java as the combiner must be a Java class. NOTE The Streaming API allows you to specify a combiner using the -combiner option. For versions up to at least 0.20, the combiner must still be a Java class. It’s best to write your mapper and reducer in a Java language. Fortunately, the Hadoop roadmap supports native Streaming scripts as combiners. In practice, one can get the equivalent of a combiner by setting the mapper to a Unix pipe ‘mapper.py | sort | combiner.py’. In addition, if you’re using the Aggregate package, each value aggregator already has a built-in (Java) combiner. The Aggregate package will automatically use these combiners. Let’s write a Java mapper (listing 4.12) that’s analogous to AverageByAttributeMapper. py of listing 4.7. Listing 4.12 Java equivalent of AverageByAttributeMapper.py public static class MapClass extends MapReduceBase implements Mapper { public void map(LongWritable key, Text value, OutputCollector output, Reporter reporter) throws IOException { String fields[] = value.toString().split(",", -20); String country = fields[4]; String numClaims = fields[8]; if (numClaims.length() > 0 && !numClaims.startsWith("\"")) { output.collect(new Text(country), new Text(numClaims + ",1")); q } } } The crucial difference in this new Java mapper is that the output is now appended with a count of 1 q. We could’ve defined a new Writable data type that holds both the value and count, but things are simple enough that we’re just keeping a comma- separated string in Text. At the reducer, the list of values for each key are parsed. The total sum and count are then computed by summation and divided at the end to get the average . public static class Reduce extends MapReduceBase implements Reducer { public void reduce(Text key, Iterator values, Include output count of 1 OutputCollector output, Reporter reporter) throws IOException { double sum = 0; int count = 0; while (values.hasNext()) { String fields[] = values.next().toString().split(","); sum += Double.parseDouble(fields[0]); count += Integer.parseInt(fields[1]); } output.collect(key, new DoubleWritable(sum/count)); } } The logic of the refactored MapReduce job was not too hard to follow, was it? We add- ed an explicit count for each key/value pair. This refactoring allows the intermediate data to be combined at each mapper before it’s sent across the network. Programmatically, the combiner must implement the Reducer interface. The combiner’s reduce() method performs the combining operation. This may seem like a bad naming scheme, but recall that for the important class of distributive functions , the combiner and the reducer perform the same operations. Therefore, the combiner has adopted the reducer’s signature to simplify its reuse. You don’t have to rename your Reduce class to use it as a combiner class. In addition, because the combiner is performing an equivalent transformation, the type for the key/value pair in its output must match that of its input. In the end, we’ve created a Combine class that looks similar to the Reduce class, except it only outputs the (partial) sum and count at the end, whereas the reducer computes the final average. public static class Combine extends MapReduceBase implements Reducer { public void reduce(Text key, Iterator values, OutputCollector output, Reporter reporter) throws IOException { double sum = 0; int count = 0; while (values.hasNext()) { String fields[] = values.next().toString().split(","); sum += Double.parseDouble(fields[0]); count += Integer.parseInt(fields[1]); } output.collect(key, new Text(sum + "," + count)); } } To enable the combiner, the driver must specify the combiner’s class to the JobConf object. You can do this through the setCombinerClass() method. The driver sets the mapper, combiner, and the reducer: Improving performance with combiners 97 98 CHAPTER 4 Writing basic MapReduce programs Job Counters Counter Map Reduce Total Map Reduce Framework File Systems Data-local map tasks Launched reduce tasks Launched map tasks Reduce input records Reduce input groups Map output records Map input records Combine output records Reduce output records Combine input records Map output bytes Map input bytes HDFS bytes written HDFS bytes read Local bytes written Local bytes read 0 0 0 0 0 1,984,055 2,923,923 1,063 0 1,984,625 18,862,764 236,903,179 0 236,915,470 20,554 21,112 0 0 0 151 151 0 0 151 151 493 0 0 2.658 0 2,510 2,510 4 2 4 151 151 1,984,055 2,923,923 1,214 151 1,985,118 18,862,764 236,903,179 2,658 236,915,470 23,064 23,622 Figure 4.5 Monitoring the effectiveness of the combiner in the AveragingWithCombiner job job.setMapperClass(MapClass.class); job.setCombinerClass(Combine.class); job.setReducerClass(Reduce.class); A combiner doesn’t necessarily improve performance. You should monitor the job’s behavior to see if the number of records outputted by the combiner is meaningfully less than the number of records going in. The reduction must justify the extra execution time of running a combiner. You can easily check this through the JobTracker’s Web UI , which we’ll see in chapter 6. Looking at figure 4.5, note that in the map phase, combine has 1,984,625 input records and only 1,063 output records. Clearly the combiner has reduced the amount of intermediate data. Note that the reduce side executes the combiner, though the benefit of this is negligible in this case. 4.7 Exercising what you’ve learned Practice is the path to proficiency. You can try the following exercises to hone your ability to think in the MapReduce paradigm. 1 Top K records—Change AttributeMax.py (or AttributeMax.php) to output the entire record rather than only the maximum value. Rewrite it such that the MapReduce job outputs the records with the top K values rather than only the maximum. Exercising what you’ve learned 99 2 Web traffic measurement—Take a web server log file and write a Streaming program with the Aggregate package to find the hourly traffic to that site. 3 Inner product of two sparse vectors—A vector is a list of values. Given two vectors, X = [x1, x2, ...] and Y = [y1, y2, ...], their inner product is Z = x1 * y1 + x2 * y2 + ... . When X and Y are long but have many elements with zero value, they’re usually given in a sparse representation: 1,0.46 9,0.21 17,0.92 ... where the key (first column) is the index into the vector. All elements not explicitly specified are considered to have a value of zero. Note that the keys don’t need to be in a sorted order. In fact, the keys may not even be numerical. (For natural language processing , the keys can be words in a document, and the inner product is a measure of document similarity .) Write a Streaming job to compute the inner product of two sparse vectors . You can add a post- processing step after the MapReduce job to complete the computation. 4 Time series processing—Consider time-series data, where each record has a timestamp as key and a measurement (on that time period) as value. We want an output that is a linear function of the time series in a form: y(t) = a0 * x(t) + a1 * x(t-1) + a2 * x(t-2) + ... + aN * x(t-N) where t stands for time and a0,…,aN are known constants. In signal processing , this is known as an FIR filter . A particularly popular instance is the moving average , where a0 = a1 = … = aN = 1/N. Each point in y is the average of the previous N points in x. It’s a simple way to smooth out time series. Implement this linear filter in MapReduce. Be sure to use a combiner. If you order the time series data chronologically (as they usually are) and N is relatively small, what’s the reduction in network traffic for shuffling when a combiner is used? For extra credit, write your own partitioner so the output stays ordered chronologically. For the more advanced practitioners, this example illustrates the difference between scalability and performance. Implementing an FIR filter in Hadoop makes it scalable to process terabytes or more of data. Students of signal processing will recognize that a high performance implementation of an FIR filter often calls for a technique known as Fast Fourier Transform (FFT ). A solution that is scalable and high performing would call for a MapReduce implementation of FFT, which is beyond the scope of this book. 5 Commutative property—Recall from basic math that the commutative property means the order of operation is irrelevant. For example, addition obeys the commutative property, as a+b=b+a and a+b+c=b+a+c=b+c+a=c+a+b=c+b+a. Is the 100 CHAPTER 4 Writing basic MapReduce programs MapReduce framework fundamentally designed for implementing commutative functions? Why or why not? 6 Multiplication (product)—Many machine-learning and statistical-classification algorithms involve the multiplication of a large number of probability values. Usually we compare the product of one set of probabilities to the product of a different set, and choose a classification corresponding to the bigger product. We’ve seen that maximum is a distributive function. Is the product also distributive ? Write a MapReduce program that multiplies all values in a data set. For full credit, apply the program to a reasonably large data set. Does implementing the program in MapReduce solve all scalability issues? What should you do to fix it? (Writing your own floating-point library is a popular answer, but not a good one.) 7 Translation into fictional dialect—A popular assignment in introductory computer science classes is to write a program that converts English to “pirate-speak .” Many variations of the exercise exist for other semi-fictional dialects, such as “Snoop Dogg ” and “E-40.” Usually the solution involves a dictionary look-up for exact word matches (“for” becomes “fo,” “sure” becomes “sho,” “the” becomes “da,” etc.), simple text rules (words ending in “ing” now ends in “in’,” replace the last vowel of a word and everything after it with “izzle,” etc.), and random injections (“kno’ wha’ im sayin’?”). Write such translations and use Hadoop to apply it to a large corpus such as Wikipedia. 4.8 Summary MapReduce programs follow a template. Often the whole program is defined within a single Java class. Within the class, a driver sets up a MapReduce job’s configuration object, which is used as the blueprint for how the job is set up and run. You’ll find the map and reduce functions in subclasses of Mapper and Reducer, respectively. Those classes are often no more than a couple dozen lines long, so they’re usually written as inner classes for convenience. Hadoop provides a Streaming API for writing MapReduce programs in a language other than Java. Many MapReduce programs are much easier to develop in a scripting language using the Streaming API, especially for ad hoc data analysis. The Aggregate package, when used with Streaming, enables one to rapidly write programs for counting and getting basic statistics. MapReduce programs are largely about the map and the reduce functions, but Hadoop allows for a combiner function to improve performance by “pre-reducing” the intermediate data at the mapper before the reduce phase. In standard programming (outside of the MapReduce paradigm), counting, summing, averaging, and so on are usually done through a simple, single pass of the data. Refactoring those programs to run in MapReduce, as we’ve done in this chapter, is relatively straightforward conceptually. More complex data analysis algorithms call for deeper reworking of the algorithms, which we cover in the next chapter. Further resources 101 4.9 Further resources Although we’ve focused on the patent data sets in this chapter, there are other large publicly accessible data sets that you can download and play around with. Below are a few examples. http://www.netflixprize.com/index—Netflix is an online movie rental site. A crucial part of its business is a recommendation engine that suggests new movies to a user based on the user’s ratings of previous movies. As part of a competition, it released a data set of user ratings to challenge people to develop better recommendation algorithms. The uncompressed data comes at 2 GB+. It contains 100 M+ movie ratings from 480 K users on 17 K movies. http://aws.amazon.com/publicdatasets/—Amazon has hosted for free several large public data sets for its EC2 users. As of this writing, the data sets belong to the three categories of biology, chemistry, and economics. For example, one of the biological data sets is an annotated human genome data of roughly 550 GB. Under economics you can find data sets, such as the 2000 U.S. Census (approximately 200 GB). http://boston.lti.cs.cmu.edu/Data/clueweb09/—Carnegie Mellon University’s Language Technologies Institute has released the ClueWeb09 data set to aid large-scale web research. It’s a crawl of a billion web pages in 10 languages. The uncompressed data set takes up 25 TB. Given the size of the data set, the most efficient way to get it is in compressed form (which takes up 5 TB) shipped in hard disk drives. (At a certain scale, shipping hard drives through FedEx has a high “bandwidth.”) As of this writing, CMU charges US$790 to ship four 1.5 TB drives with the compressed data. 5 102 Advanced MapReduce This chapter covers ■ Chaining multiple MapReduce jobs ■ Creating Bloom filters Performing joins of multiple data sets ■ As your data processing becomes more complex you’ll want to exploit different Hadoop features. This chapter will focus on some of these more advanced techniques. When handling advanced data processing, you’ll often find that you can’t program the process into a single MapReduce job. Hadoop supports chaining MapReduce programs together to form a bigger job. You’ll also find that advanced data processing often involves more than one data set. We’ll explore various joining techniques in Hadoop for simultaneously processing multiple data sets. You can code certain data processing tasks more efficiently when processing a group of records at a time. We’ve seen how Streaming natively supports the ability to process a whole split at a time, and the Streaming implementation of the maximum function takes advantage of this ability. We’ll see that the same is true for Java programs. We’ll discover the Bloom filter and implement it with a mapper that keeps state information across records. Chaining MapReduce jobs 103 5.1 Chaining MapReduce jobs You’ve been doing data processing tasks which a single MapReduce job can accom- plish. As you get more comfortable writing MapReduce programs and take on more ambitious data processing tasks, you’ll find that many complex tasks need to be broken down into simpler subtasks, each accomplished by an individual MapReduce job. For example, from the citation data set you may be interested in finding the ten most- cited patents. A sequence of two MapReduce jobs can do this. The first one creates the “inverted” citation data set and counts the number of citations for each patent, and the second job finds the top ten in that “inverted” data. 5.1.1 Chaining MapReduce jobs in a sequence Though you can execute the two jobs manually one after the other, it’s more convenient to automate the execution sequence. You can chain MapReduce jobs to run sequen- tially, with the output of one MapReduce job being the input to the next. Chaining MapReduce jobs is analogous to Unix pipes . mapreduce-1 | mapreduce-2 | mapreduce-3 | ... Chaining MapReduce jobs sequentially is quite straightforward. Recall that a driver sets up a JobConf object with the configuration parameters for a MapReduce job and passes the JobConf object to JobClient.runJob() to start the job. As Job- Client.runJob() blocks until the end of a job, chaining MapReduce jobs involves calling the driver of one MapReduce job after another. The driver at each job will have to create a new JobConf object and set its input path to be the output path of the previous job. You can delete the intermediate data generated at each step of the chain at the end. 5.1.2 Chaining MapReduce jobs with complex dependency Sometimes the subtasks of a complex data processing task don’t run sequentially, and their MapReduce jobs are therefore not chained in a linear fashion. For example, mapreduce1 may process one data set while mapreduce2 independently processes an- other data set. The third job, mapreduce3, performs an inner join of the first two jobs’ output. (We’ll discuss data joining in the next sections.) It’s dependent on the other two and can execute only after both mapreduce1 and mapreduce2 are completed. But mapreduce1 and mapreduce2 aren’t dependent on each other. Hadoop has a mechanism to simplify the management of such (nonlinear) job dependencies via the Job and JobControl classes. A Job object is a representation of a MapReduce job. You instantiate a Job object by passing a JobConf object to its constructor. In addition to holding job configuration information, Job also holds dependency information, specified through the addDependingJob() method. For Job objects x and y, x.addDependingJob(y) 104 CHAPTER 5 Advanced MapReduce means x will not start until y has finished. Whereas Job objects store the configuration and dependency information, JobControl objects do the managing and monitoring of the job execution. You can add jobs to a JobControl object via the addJob() method. After adding all the jobs and dependencies, call JobControl’s run() method to spawn a thread to submit and monitor jobs for execution. JobControl has methods like all- Finished() and getFailedJobs() to track the execution of various jobs within the batch. 5.1.3 Chaining preprocessing and postprocessing steps A lot of data processing tasks involve record-oriented preprocessing and postprocess- ing. For example, in processing documents for information retrieval , you may have one step to remove stop words (words like a, the, and is that occur frequently but aren’t too meaningful), and another step for stemming (converting different forms of a word into the same form, such as finishing and finished into finish.) You can write a separate MapReduce job for each of these pre- and postprocessing steps and chain them to- gether, using IdentityReducer (or no reducer at all) for these steps. This approach is inefficient as each step in the chain takes up I/O and storage to process the intermedi- ate results. Another approach is for you to write your mapper such that it calls all the preprocessing steps beforehand and the reducer to call all the postprocessing steps afterward. This forces you to architect the pre- and postprocessing steps in a modular and composable manner. Hadoop introduced the ChainMapper and the ChainReducer classes in version 0.19.0 to simplify the composition of pre- and postprocessing. You can think of chaining MapReduce jobs, as explained in section 5.1.1, symbolically using the pseudo-regular expression: [MAP | REDUCE]+ where a reducer REDUCE comes after a mapper MAP, and this [MAP | REDUCE] se- quence can repeat itself one or more times, one right after another. The analogous expression for a job using ChainMapper and ChainReducer would be MAP+ | REDUCE | MAP* The job runs multiple mappers in sequence to preprocess the data, and after running reduce it can optionally run multiple mappers in sequence to postprocess the data. The beauty of this mechanism is that you write the pre- and postprocessing steps as standard mappers. You can run each one of them individually if you want. (This is useful when you want to debug them individually.) You call the addMapper() method in ChainMapper and ChainReducer to compose the pre- and postprocessing steps, respectively. Running all the pre- and postprocessing steps in a single job leaves no intermediate file and there’s a dramatic reduction in I/O. Consider the example where there are four mappers (Map1, Map2, Map3, and Map4) and one reducer (Reduce), and they’re chained into a single MapReduce job in this sequence: Map1 | Map2 | Reduce | Map3 | Map4 Chaining MapReduce jobs 105 In this setup, you should think of Map2 and Reduce as the core of the MapReduce job, with the standard partitioning and shuffling applied between the mapper and reducer. You should consider Map1 as a preprocessing step and Map3 and Map4 as postprocess- ing steps. The number of processing steps can vary. This is only an example. You can specify the composition of this sequence of mappers and reducer with the driver. See listing 5.1. You need to make sure the key and value outputs of one task have matching types (classes) with the inputs of the next task. Listing 5.1 Driver for chaining mappers within a MapReduce job Configuration conf = getConf(); JobConf job = new JobConf(conf); job.setJobName("ChainJob"); job.setInputFormat(TextInputFormat.class); job.setOutputFormat(TextOutputFormat.class); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); JobConf map1Conf = new JobConf(false); ChainMapper.addMapper(job, Map1.class, LongWritable.class, Text.class, Text.class, Text.class, true, map1Conf); JobConf map2Conf = new JobConf(false); ChainMapper.addMapper(job, Map2.class, Text.class, Text.class, LongWritable.class, Text.class, true, map2Conf); JobConf reduceConf = new JobConf(false); ChainReducer.setReducer(job, Reduce.class, LongWritable.class, Text.class, Text.class, Text.class, true, reduceConf); JobConf map3Conf = new JobConf(false); ChainReducer.addMapper(job, Map3.class, Text.class, Text.class, LongWritable.class, Add Map1 step to job Add Map2 step to job Add Reduce step to job Add Map3 step to job 106 CHAPTER 5 Advanced MapReduce Text.class, true, map3Conf); JobConf map4Conf = new JobConf(false); ChainReducer.addMapper(job, Map4.class, LongWritable.class, Text.class, LongWritable.class, Text.class, true, map4Conf); JobClient.runJob(job); The driver first sets up the “global” JobConf object with the job’s name, input path, output path, and so forth. It adds the five steps of the chained job one at a time, in the sequence of the steps’ execution. It adds all the steps before Reduce using the static ChainMapper.addMapper() method. It sets the reducer with the static ChainReducer.setReducer() method. Using the ChainReducer.addMapper() method, it adds the last steps. The global JobConf object (job) is passed through all five add* methods. In addition, each mapper and the reducer have a local JobConf object (map1Conf, map2Conf, map3Conf, map4Conf, and reduceConf) that takes pre- cedence over the global one in configuring the individual mapper/reducer. The recommended local JobConf object is a new JobConf object initiated without defaults — new JobConf(false). Let’s look at the signature of the ChainMapper.addMapper() method to understand in detail how to add each step to the chained job. The signature and function of ChainReducer.setReducer() and ChainReducer.addMapper() are analogous and we’ll skip them. public static void addMapper(JobConf job, Class> klass, Class inputKeyClass, Class inputValueClass, Class outputKeyClass, Class outputValueClass, boolean byValue, JobConf mapperConf) This method has eight arguments. The first and last are the global and local JobConf objects, respectively. The second argument (klass ) is the Mapper class that will do the data processing. The four arguments inputValueClass, inputKey- Class, outputKeyClass, and outputValueClass are the input/output class types of the Mapper class. The argument byValue will need a little explanation. In the standard Mapper model , the output key/value pairs are serialized and written to disk,1 prepared to be shuffled Add Map3 step to job Add Map4 step to job 1 The key and value’s ability to be cloned and serialized is provided by them being implemented as Writables. Joining data from different sources 107 to a reducer that may be at a completely different node. Formally this is considered to be passed by value , as a copy of the key/value pair is sent over. In the current case where we can chain one Mapper to another, we can execute the two in the same JVM thread. Therefore, it’s possible for the key/value pairs to be passed by reference , where the output of the initial Mapper stays in place in memory and the following Mapper refers to it directly in the same memory location. When Map1 calls OutputCollector.collect (K k, V v) , the objects k and v pass directly to Map2’s map() method. This improves performance by not having to clone a potentially large volume of data between the mappers. But doing this can violate one of the more subtle “contracts” in Hadoop’s MapReduce API. The call to OutputCollector.collect(K k, V v) is guaranteed to not alter the content of k and v. Map1 can call OutputCollector.collect(K k, V v) and then use the objects k and v afterward, fully expecting their values to stay the same. But if we pass those objects by reference to Map2, then Map2 may alter them and violate the API’s guarantee. If you’re sure that Map1’s map() method doesn’t use the content of k and v after calling OutputCollector.collect(K k, V v), or that Map2 doesn’t change the value of its k and v input, you can achieve some performance gains by setting byValue to false. If you’re not sure of the Mapper’s internal code, it’s best to play safe and let byValue be true, maintaining the pass-by-value model, and be certain that the Mappers will work as expected. 5.2 Joining data from different sources It’s inevitable that you’ll come across data analyses where you need to pull in data from different sources. For example, given our patent data sets, you may want to find out if certain countries cite patents from another country. You’ll have to look at citation data (cite75_99.txt) as well as patent data for country information (apat63_99. txt). In the database world it would just be a matter of joining two tables, and most databases automagically take care of the join processing for you. Unfortunately, join- ing data in Hadoop is more involved, and there are several possible approaches with different trade-offs. We use a couple toy data sets to better illustrate joining in Hadoop. Let’s take a comma-separated Customers file where each record has three fields: Customer ID, Name, and Phone Number. We put four records in the file for illustration: 1,Stephanie Leung,555-555-5555 2,Edward Kim,123-456-7890 3,Jose Madriz,281-330-8004 4,David Stork,408-555-0000 We store Customer orders in a separate file, called Orders. It’s in CSV format, with four fields: Customer ID, Order ID, Price, and Purchase Date. 3,A,12.95,02-Jun-2008 1,B,88.25,20-May-2008 108 CHAPTER 5 Advanced MapReduce 2,C,32.00,30-Nov-2007 3,D,25.02,22-Jan-2009 If we want an inner join of the two data sets above, the desired output would look a listing 5.2. Listing 5.2 Desired output of an inner join between Customers and Orders data 1,Stephanie Leung,555-555-5555,B,88.25,20-May-2008 2,Edward Kim,123-456-7890,C,32.00,30-Nov-2007 3,Jose Madriz,281-330-8004,A,12.95,02-Jun-2008 3,Jose Madriz,281-330-8004,D,25.02,22-Jan-2009 Hadoop can also perform outer joins, although to simplify explanation we focus on inner joins. 5.2.1 Reduce-side joining Hadoop has a contrib package called datajoin that works as a generic framework for data joining in Hadoop. Its jar file is at contrib/datajoin/hadoop-*-datajoin. jar. To distinguish it from other joining techniques, it’s called the reduce-side join, as we do most of the processing on the reduce side. It’s also known as the reparti- tioned join (or the repartitioned sort-merge join), as it’s the same as the database tech- nique of the same name. Although it’s not the most efficient joining technique, it’s the most general and forms the basis of some more advanced techniques (such as the semijoin ). Reduce-side join introduces some new terminologies and concepts, namely, data source , tag , and group key . A data source is analogous to a table in relational databases. We have two data sources in our toy example: Customers and Orders. A data source can be a single file or multiple files. The important point is that all the records in a data source have the same structure, analogous to a schema . The MapReduce paradigm calls for processing each record one at a time in a stateless manner. If we want some state information to persist, we have to tag the record with such state. For example, given our two files, a record may look to a mapper like this: 3,Jose Madriz,281-330-8004 or: 3,A,12.95,02-Jun-2008 where the record type (Customers or Orders) is dissociated from the record itself. Tag- ging the record will ensure that specific metadata will always go along with the record. For the purpose of data joining, we want to tag each record with its data source. Joining data from different sources 109 The group key functions like a join key in a relational database. For our example, the group key is the Customer ID. As the datajoin package allows the group key to be any user-defined function, group key is more general than a join key in a relational database. Before explaining how to use the contrib package, let’s go through all the major steps in a repartitioned sort-merge join of our toy datasets. After seeing how those steps fit together, we’ll see which steps are done by the datajoin package, and which ones we program. We’ll have code to see the hooks for integrating our code with the datajoin package. DATA FLOW OF A REDUCE-SIDE JOIN Figure 5.1 illustrates the data flow of a repartitioned join on the toy data sets Custom- ers and Orders, up to the reduce stage. We’ll go into more details later to see what happens in the reduce stage. First we see that mappers receive data from two files, Customers and Orders. Each mapper knows the filename of the data stream it’s processing. The map() function is called with each record, and the main goal of map() is to package each record such that joining on the reduce side is possible. Recall that in the MapReduce framework, map() outputs records as key/value pairs that are partitioned on the key, and all records of the same key will end up in a single reducer and be processed together. For joining, we would want the map() function to output a record package where the key is the group key for joining—the Customer ID in this case. The value in this key/value package will be the original record, tagged with the data source (i.e., filename). For example, for the record 3,A,12.95,02-Jun-2008 from the Orders file, map() will output a key/value pair where the key is "3", the Customer ID that will be used to join with records from the Customers file. The value output by map() is the entire record wrapped by a tag "Orders". After map()packages each record of the inputs, MapReduce’s standard partition, shuffle, and sort takes place. Note that as the group key is set to the join key, reduce() will process all records of the same join key together. The function reduce() will unwrap the package to get the original record and the data source of the record by its tag. We see that for group keys (Customer IDs) "1" and "2", the reduce() function gets two values. One value is tagged with "Customers" and the other value is tagged with "Orders". For the map output with (group) key "4", reduce() will only see one value, which is tagged with "Customers". This is expected as there is no record in Orders with a Customer ID of "4". On the other hand, reduce() will see three values for the (group) key "3". This is due to one record from Customers and two more from Orders. 110 CHAPTER 5 Advanced MapReduce Customers 1,Stephanie Leung,555-555-5555 2,Edward Kim,123-456-7890 3,Jose Madriz,281-330-8004 4,David Stork,408-555-0000 ... Orders 3,A,12.95,02-Jun-2008 1,B,88.25,20-May-2008 2,C,32.00,30-Nov-2007 3,D,25.02,22-Jan-2009 ... Customers 1,Stephanie Leung,555-555-5555 1 Customers 2,Edward Kim,123-456-7890 2 Customers 3,Jose Madriz,281-330-8004 3 Customers 4,David Stork,408-555-0000 4 Orders 3,A,12.95,02-Jun-2008 3 Orders 1,B,88.25,20-May-2008 1 Orders 2,C,32.00,30-Nov-2007 2 Orders 3,D,25.02,22-Jan-2009 3 Group key Tag Customers 4,David Stork,408-555-0000 4 Customers 1,Stephanie Leung,555-555-5555 1 Orders 1,B,88.25,20-May-2008 Customers 2,Edward Kim,123-456-7890 2 Orders 2,C,32.00,30-Nov-2007 Customers 3,Jose Madriz,281-330-8004 3 Orders 3,A,12.95,02-Jun-2008 Orders 3,D,25.02,22-Jan-2009 map() map() shuffle() reduce() reduce() Figure 5.1 In repartitioned join , the mapper first wraps each record with a group key and a tag. The group key is the joining attribute, and the tag is the data source (table in SQL parlance) of the record. The partition and shuffle step will group all the records with the same group key together. The reducer is called on the set of records with the same group key. The function reduce() will take its input and do a full cross-product on the values. Reduce() creates all combinations of the values with the constraint that a combination will not be tagged more than once. In cases where reduce() sees values of distinct tags, the cross-product is the original set of values. In our example, this is the case for group Joining data from different sources 111 keys 1, 2, and 4. Figure 5.2 illustrates the cross product for group key 3. We have three values, one tagged with Customers and two tagged with Orders. The cross-product cre- ates two combinations. Each combination consists of the Customers value and one of the Orders value. NOTE Our toy example has an implicit schema that Customer ID identifies a unique record in Customers, making the number of combinations in a cross- product always the number of Orders records with the Customer ID (except Customers 3,Jose Madriz,281-330-8004 3 Orders 3,A,12.95,02-Jun-2008 Orders 3,D,25.02,22-Jan-2009 Customers 3,Jose Madriz,281-330-8004 Orders 3,A,12.95,02-Jun-2008 Customers 3,Jose Madriz,281-330-8004 Orders 3,D,25.02,22-Jan-2009 reduce() combine() combine() 3,Jose Madriz,281-330-8004,A,12.95,02-Jun-2008 3,Jose Madriz,281-330-8004,D,25.02,22-Jan-2009 Figure 5.2 The reduce side of a repartitioned join . For a given join key, the reduce task performs a full cross-product of values from different sources. It sends each combination to combine() to create an output record. The combine() function can choose to not output any particular combination. 112 CHAPTER 5 Advanced MapReduce when it’s zero, in which case the cross-product is the Customers record itself). For more complicated settings, the number of combinations generated by the cross-product is the product of the number of records under each tag. If reduce() sees two Customers records and three Orders records together, then the cross-product will create six (2 * 3) combinations. If there’s a third data source (Accounts) with two records, then the cross-product will create twelve (2 * 2 * 3) combinations. It feeds each combination from the cross-product into a function called combine(). (Don’t confuse with combiners as explained in section 4.5.) Due to the nature of the cross-product, combine() is guaranteed to see at most one record from each of the data sources (tags), and all the records it sees have the same join key. It’s the combine() function that determines whether the whole operation is an inner join , outer join , or another type of join . In an inner join , combine() drops all combinations where not all tags are present, such as our case with group key "4". Otherwise combine() merges the records from different sources into a single output record. Now you see why we call this joining process the repartitioned sort-merge join. The records in the original input sources can be in random order. They are repartitioned onto the reducers in the right grouping. The reducer can then merge records of the same join key together to create the desired join output. (The sort happens but it’s not critical to understanding the operation.) IMPLEMENTING JOIN WITH THE DATAJOIN PACKAGE Hadoop’s datajoin package implements the dataflow of a join as described previously. We have certain hooks to handle the details of our particular data structure and a spe- cial hook for us to define the exact function of combine(). Hadoop ’s datajoin package has three abstract classes that we inherit and make concrete: DataJoinMapperBase , DataJoinReducerBase , and TaggedMapOutput . As the names suggest, our MapClass will extend DataJoinMapperBase, and our Reduce class will extend DataJoinReducerBase. The datajoin package has already implemented the map() and reduce() methods in these respective base classes to perform the join dataflow describe in the last section. Our subclass will only have to implement a few new methods to configure the details. Before explaining how to use DataJoinMapperBase and DataJoinReducerBase, you need to understand a new abstract data type TaggedMapOutput that is used throughout the code. Recall from the dataflow description that the mapper outputs a package with a (group) key and a value that is a tagged record. The datajoin package specifies the (group) key to be of type Text and the value (i.e., the tagged record) to be of type TaggedMapOutput. TaggedMapOutput is a data type for wrapping our records with a Text tag. It trivially implements a getTag() and a setTag(Text tag) method. It specifies an abstract method getData(). Our subclass will implement that method to handle the type of the record. There’s no explicit requirement for the subclass to have a setData() method but we must pass in the record data. The subclass can implement such a setData() method for the sake of symmetry or take in a record in Joining data from different sources 113 the constructor. In addition, as the output of a mapper, TaggedMapOutput needs to be Writable . Therefore, our subclass has to implement the readFields() and write() methods. We created TaggedWritable , a simple subclass for handling any Writable record type. public static class TaggedWritable extends TaggedMapOutput { private Writable data; public TaggedWritable(Writable data) { this.tag = new Text(""); this.data = data; } public Writable getData() { return data; } ... } Recall from the join dataflow that the mapper’s main function is to package a record such that it goes to the same reducer as other records with the same join key. Data- JoinMapperBase performs all the packaging, but the class specifies three abstract methods for our subclass to fill in: protected abstract Text generateInputTag(String inputFile); protected abstract TaggedMapOutput generateTaggedMapOutput(Object value); protected abstract Text generateGroupKey(TaggedMapOutput aRecord); The generateInputTag() is called at the start of a map task to globally specify the tag for all the records this map task will process. The tag is defined to be of type Text. Note that we call the generateInputTag()with the filename of the records. The mapper working on the Customers file will receive the string "Customers" as the argument to generateInputTag(). As we’re using the tag to signify the data source, and our file- name is set up to reflect the data source, generateInputTag() is protected Text generateInputTag(String inputFile) { return new Text(inputFile); } If a data source is spread out over several files (part-0000, part-0001, etc.), you would not want the tag to be the complete filename, rather some prefix of it. For example, the tag (data source) can be the filename before the dash (-) sign. protected Text generateInputTag(String inputFile) { String datasource = inputFile.split('-')[0]; return new Text(datasource); } We store the result of generateInputTag()in the DataJoinMapperBase object’s inputTag variable for later use. We can also store the filename in DataJoinMapper- Base’s inputFile variable if we want to refer to it again. After the map task’s initialization, DataJoinMapperBase’s map() is called for each record. It calls the two abstract methods that we have yet to implement. 114 CHAPTER 5 Advanced MapReduce public void map(Object key, Object value, OutputCollector output, Reporter reporter) throws IOException { TaggedMapOutput aRecord = generateTaggedMapOutput(value); Text groupKey = generateGroupKey(aRecord); output.collect(groupKey, aRecord); } The generateTaggedMapOutput() method wraps the record value into a Tagged- MapOutput type. Recall the concrete implementation of TaggedMapOutput that we’re using is called TaggedWritable. The method generateTaggedMapOutput() can re- turn a TaggedWritable with any Text tag that we want. In principle, the tag can even be different for different records in the same file. In the standard case, we want the tag to stand for the data source that our generateInputTag() had computed earlier and stored in this.inputTag. protected TaggedMapOutput generateTaggedMapOutput(Object value) { TaggedWritable retv = new TaggedWritable((Text) value); retv.setTag(this.inputTag); return retv; } The generateGroupKey() method takes a tagged record (of type TaggedMapOutput) and returns the group key for joining. For our current purpose, we unwrap the tagged record and take the first field in the CSV-formatted value as the join key. protected Text generateGroupKey(TaggedMapOutput aRecord) { String line = ((Text) aRecord.getData()).toString(); String[] tokens = line.split(","); String groupKey = tokens[0]; return new Text(groupKey); } In a more general implementation, the user will be able to specify which field should be the joining key and if the record separator may be some character other than a comma. DataJoinMapperBase is a simple class, and much of the mapper code is in our subclass. DataJoinReducerBase, on the other hand, is the workhorse of the datajoin package, and it simplifies our programming by performing a full outer join for us. Our reducer subclass only has to implement the combine() method to filter out unwanted combinations to get the desired join operation (inner join, left outer join, etc.). It’s also in the combine() method that we format the combination into the appropriate output format. We give the combine() method one combination of the cross product of the tagged records with the same join (group) key. This may sound complicated, but recall from the dataflow diagrams in figures 5.1 and 5.2 that the cross-product is simple for the canonical case of two data sources. Each combination will have either two records (meaning there’s at least one record in each data source with the join key) or one (meaning only one data source has that join key). Let’s look at the signature of combine(): Joining data from different sources 115 protected abstract TaggedMapOutput ➥ combine(Object[] tags, Object[] values); An array of tags and an array of values represent the combination. The size of those two arrays is guaranteed to be the same and equal to the number of tagged records in the combination. The first tagged record in the combination is represented by tags[0] and values[0], the second one is tags[1] and values[1], and so forth. Further- more, the tags are always in sorted order. As tags correspond to the data sources, in the canonical case of joining two data sources, the tags array to combine() won’t be longer than two. Figure 5.2 shows combine() being called twice. For the left side, the tags and values arrays are like this: 2 tags = {"Customers", "Orders"}; values = {"3,Jose Madriz,281-330-8004", "A,12.95,02-Jun-2008"}; For an inner join , combine() will ignore combinations where not all tags are present. It does so by returning null. Given a legal combination, the role of combine() is to concat- enate all the values into one single record for output. The order of concatenation is fully determined by combine(). In the case of an inner join, the length of values[] is always the number of data sources available (two in the canonical case), and the tags are always in sorted order. It’s a sensible choice to loop through the values[] array to get the default alphabetical ordering based on data source names. DataJoinReducerBase, like any reducer, outputs key/value pairs. For each legal combination, the key is always the join key and the value is the output of combine(). Note that the join key is still present in each element of the values[] array. The combine() method should strip out the join key in those elements before concatenating them. Otherwise the join key will be shown multiple times in one output record. Finally, DataJoinReducerBase expects combine() to return a TaggedMapOutput . It’s unclear why as DataJoinReducerBase ignores the tag in the TaggedMapOutput object. Listing 5.3 shows the complete code, including our reduce subclass. Listing 5.3 Inner join of data from two files using reduce-side join public class DataJoin extends Configured implements Tool { public static class MapClass extends DataJoinMapperBase { protected Text generateInputTag(String inputFile) { String datasource = inputFile.split("-")[0]; return new Text(datasource); } protected Text generateGroupKey(TaggedMapOutput aRecord) { String line = ((Text) aRecord.getData()).toString(); String[] tokens = line.split(","); String groupKey = tokens[0]; return new Text(groupKey); 2 The tags array is of type Text[] and values is of type TaggedWritable[]. We ignore those details to focus on the their contents. 116 CHAPTER 5 Advanced MapReduce } protected TaggedMapOutput generateTaggedMapOutput(Object value) { TaggedWritable retv = new TaggedWritable((Text) value); retv.setTag(this.inputTag); return retv; } } public static class Reduce extends DataJoinReducerBase { protected TaggedMapOutput combine(Object[] tags, Object[] values) { if (tags.length < 2) return null; String joinedStr = ""; for (int i=0; i 0) joinedStr += ","; TaggedWritable tw = (TaggedWritable) values[i]; String line = ((Text) tw.getData()).toString(); String[] tokens = line.split(",", 2); joinedStr += tokens[1]; } TaggedWritable retv = new TaggedWritable(new Text(joinedStr)); retv.setTag((Text) tags[0]); return retv; } } public static class TaggedWritable extends TaggedMapOutput { private Writable data; public TaggedWritable(Writable data) { this.tag = new Text(""); this.data = data; } public Writable getData() { return data; } public void write(DataOutput out) throws IOException { this.tag.write(out); this.data.write(out); } public void readFields(DataInput in) throws IOException { this.tag.readFields(in); this.data.readFields(in); } } public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, DataJoin.class); Path in = new Path(args[0]); Path out = new Path(args[1]); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); Joining data from different sources 117 job.setJobName("DataJoin"); job.setMapperClass(MapClass.class); job.setReducerClass(Reduce.class); job.setInputFormat(TextInputFormat.class); job.setOutputFormat(TextOutputFormat.class); job.setOutputKeyClass(Text.class); job.setOutputValueClass(TaggedWritable.class); job.set("mapred.textoutputformat.separator", ","); JobClient.runJob(job); return 0; } public static void main(String[] args) throws Exception { int res = ToolRunner.run(new Configuration(), new DataJoin(), args); System.exit(res); } } Next we’ll look at another way of doing joins that is more efficient in some common applications. 5.2.2 Replicated joins using DistributedCache The reduce-side join technique discussed in the last section is flexible, but it can also be quite inefficient. Joining doesn’t take place until the reduce phase. We shuffle all data across the network first, and in many situations we drop the majority of this data during the joining process. It would be more efficient if we eliminate the unnecessary data right in the map phase. Even better would be to perform the entire joining opera- tion in the map phase. The main obstacle to performing joins in the map phase is that a record being processed by a mapper may be joined with a record not easily accessible (or even located) by that mapper. If we can guarantee the accessibility of all the necessary data when joining a record, joining on the map side can work. For example, if we know that the two sources of data are partitioned into the same number of partitions and the partitions are all sorted on the key and the key is the desired join key, then each mapper (with the proper InputFormat and RecordReader ) can deterministically locate and retrieve all the data necessary to perform joining. In fact, Hadoop’s org. apache.hadoop.mapred.join package contains helper classes to facilitate this map- side join . Unfortunately, situations where we can naturally apply this are limited, and running extra MapReduce jobs to repartition the data sources to be usable by this package seems to defeat the efficiency gain. Therefore, we’ll not pursue this package further. All hope is not lost though. There’s another data pattern that occurs quite frequently that we can take advantage of. When joining big data, often only one of the sources is big; the second source may be orders of magnitude smaller. For example, a local 118 CHAPTER 5 Advanced MapReduce phone company’s Customers data may have only tens of millions of records (each record containing basic information for one customer), but its transaction log can have billions of records containing detailed call history. When the smaller source can fit in memory of a mapper, we can achieve a tremendous gain in efficiency by copying the smaller source to all mappers and performing joining in the map phase. This is called replicated join in the database literature as one of the data tables is replicated across all nodes in the cluster. (The next section will cover the case when the smaller source doesn’t fit in memory.) Hadoop has a mechanism called distributed cache that’s designed to distribute files to all nodes in a cluster. It’s normally used for distributing files containing “background” data needed by all mappers. For example, if you’re using Hadoop to classify documents , you may have a list of keywords for each class. (Or better yet, a probabilistic model for each class, but we digress…) You would use distributed cache to ensure all mappers have access to the lists of keywords, the “background” data. For executing replicated joins, we consider the smaller data source as background data. Distributed cache is handled by the appropriately named class DistributedCache . There are two steps to using this class. First, when configuring a job, you call the static method DistributedCache.addCacheFile() to specify the files to be disseminated to all nodes. These files are specified as URI objects, and they default to HDFS unless a different filesystem is specified. The JobTracker will take this list of URIs and create a local copy of the files in all the TaskTrackers when it starts the job. In the second step, your mappers on each individual TaskTracker will call the static method DistributedCache.getLocalCacheFiles() to get an array of local file Paths where the local copy is located. At this point the mapper can use standard Java file I/O techniques to read the local copy. Replicated joins using DistributedCache are simpler than reduce-side joins. Let’s begin with our standard Hadoop template . public class DataJoinDC extends Configured implements Tool { public static class MapClass extends MapReduceBase implements Mapper { ... } public int run(String[] args) throws Exception { ... } public static void main(String[] args) throws Exception { int res = ToolRunner.run(new Configuration(), new DataJoinDC(), args); System.exit(res); } } Joining data from different sources 119 Note that we’ve taken out the Reduce class. We plan on performing the joining in the map phase and will configure this job to have no reducers. You’ll find our driver method familiar too. public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, DataJoinDC.class); DistributedCache.addCacheFile(new Path(args[0]).toUri(), conf); Path in = new Path(args[1]); Path out = new Path(args[2]); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); job.setJobName("DataJoin with DistributedCache"); job.setMapperClass(MapClass.class); job.setNumReduceTasks(0); job.setInputFormat(KeyValueTextInputFormat.class); job.setOutputFormat(TextOutputFormat.class); job.set("key.value.separator.in.input.line", ","); JobClient.runJob(job); return 0; } The crucial addition here is where we take the file specified by the first argument and add it to DistributedCache. When we run the job, each node will create a local copy of that file from HDFS. The second and third arguments denote the input and output paths of the standard Hadoop job. Note that we’ve limited the number of data sources to two. This is not an inherent limitation of the technique, but a simplification that makes our code easier to follow. Up to now our MapClass has only had to define one method, map(). In fact, the Mapper interface (and also the Reducer interface) has two more abstract methods, configure() and close(). We call the method configure()when we first instantiate the MapClass, and the method close()when the mapper finishes processing its split. The MapReduceBase class provides default no-op implementations for these methods. Here we want to override configure() to load our join data into memory when a mapper is first initialized. This way we can have the data available each time we call map()to process a new record. public static class MapClass extends MapReduceBase implements Mapper { private Hashtable joinData = new Hashtable(); @Override public void configure(JobConf conf) { try { 120 CHAPTER 5 Advanced MapReduce Path [] cacheFiles = DistributedCache.getLocalCacheFiles(conf); if (cacheFiles != null && cacheFiles.length > 0) { String line; String[] tokens; BufferedReader joinReader = new BufferedReader( new FileReader(cacheFiles[0].toString())); try { while ((line = joinReader.readLine()) != null) { tokens = line.split(",", 2); joinData.put(tokens[0], tokens[1]); } } finally { joinReader.close(); } } } catch (IOException e) { System.err.println("Exception reading DistributedCache: " + e); } } public void map(Text key, Text value, OutputCollector output, Reporter reporter) throws IOException { String joinValue = joinData.get(key); if (joinValue != null) { output.collect(key, new Text(value.toString() + "," + joinValue)); } } } When we call configure(), we get an array of file paths to the local copy of files pushed by DistributedCache. As our driver method has only pushed one file (given by our first argument) into DistributedCache, this should be an array of size one. We read that file using standard Java file I/O. For our purpose, the program assumes each line is a record, the key/value pair is comma separated, and the key is unique and will be used for joining. The program reads this source file into a Java Hashtable called joinData that’s available throughout the mapper’s lifespan. The joining takes place in the map() method and is straightforward now that one of the sources resides in memory in the form of joinData. If we don’t find the join key in joinData, we drop the record. Otherwise, we match the (join) key to the value in joinData and concatenate the values. The result is outputted directly into HDFS as we don’t have any reducer for further processing. A not-infrequent situation in using DistributedCache is that the background data (the smaller data source in our data join case) is in the local filesystem of the client rather than stored in HDFS. One way to handle this is to add code to upload the local file on the client to HDFS before calling DistributedCache. addCacheFile(). Fortunately, this process is natively supported as one of the generic Hadoop command line arguments in GenericOptionsParser . The option is -files Joining data from different sources 121 and it automatically copies a comma-separated list of files to all the task nodes. Our command line statement is bin/hadoop jar -files small_in.txt DataJoinDC.jar big_in.txt output Now that we don’t need to call DistributedCache.addCacheFile() ourselves any- more, we no longer have to take in the filename of the smaller data source as one of the arguments. The index to the arguments has shifted. Path in = new Path(args[0]); Path out = new Path(args[1]); With these minor changes our DistributedCache join program will take a local file on the client machine as one of the input sources. 5.2.3 Semijoin : reduce-side join with map-side filtering One of the limitations in using replicated join is that one of the join tables has to be small enough to fit in memory. Even with the usual asymmetry of size in the input sources, the smaller one may still not be small enough. You can solve this problem by rearranging the processing steps to make them more efficient. For example, if you’re looking for the order history of all customers in the 415 area code, it’s correct but in- efficient to join the Orders and the Customers tables first before filtering out records where the customer is in the 415 area code. Both the Orders and Customers tables may be too big for replicated join and you’ll have to resort to the inefficient reduce-side join. A better approach is to first filter out customers living in the 415 area code. We store this in a temporary file called Customers415. We can arrive at the same end re- sult by joining Orders with Customers415, but now Customers415 is small enough that a replicated join is feasible. There is some overhead in creating and distributing the Customers415 file, but it’s often compensated by the overall gain in efficiency. Sometimes you may have a lot of data to analyze. You can’t use replicated join no matter how you rearrange your processing steps. Don’t worry. We still have ways to make reduce-side joining more efficient. Recall that the main problem with reduce- side joining is that the mapper only tags the data, all of which is shuffled across the network but most of which is ignored in the reducer. The inefficiency is ameliorated if the mapper has an extra prefiltering function to eliminate most or even all the unnecessary data before it is shuffled across the network. We need to build this filtering mechanism. Continuing our example of joining Customers415 with Orders, the join key is Customer ID and we would like our mappers to filter out any customer not from the 415 area code rather than send those records to reducers. We create a data set CustomerID415 to store all the Customer IDs of customers in the 415 area code. CustomerID415 is smaller than Customers415 because it only has one data field. Assuming CustomerID415 can now fit in memory, we can improve reduce-side join by using distributed cache to disseminate CustomerID415 across all the mappers. When processing records from Customers and Orders, the mapper will drop any record 122 CHAPTER 5 Advanced MapReduce whose key is not in the set CustomerID415. This is sometimes called a semijoin , taking the terminology from the database world. Last but not least, what if the file CustomerID415 is still too big to fit in memory? Or maybe CustomerID415 does fit in memory but it’s size makes replicating it across all the mappers inefficient. This situation calls for a data structure called a Bloom filter. A Bloom filter is a compact representation of a set that supports only the contain query. (“Does this set contain this element?”) Furthermore, the query answer is not completely accurate, but it’s guaranteed to have no false negatives and a small probability of false positives. The slight inaccuracy is the trade-off for the data structure’s compactness. By using a Bloom filter representation of CustomerID415, the mappers will pass through all customers in the 415 area code. It still guarantees the correctness of the data join algorithm. The Bloom filter will also pass a small portion of customers not in the 415 area code to the reduce phase. This is fine because those will be ignored in the reduce phase. We’ll still have improved performance by reducing dramatically the amount of traffic shuffled across the network. The use of Bloom filters is in fact a standard technique for joining in distributed databases, and it’s used in commercial products such as Oracle 11g. We’ll describe Bloom filter and its other applications in more details in the next section. 5.3 Creating a Bloom filter If you use Hadoop for batch processing of large data sets, your data-intensive com- puting needs probably include transaction-style processing as well. We won’t cover all the techniques for running real-time distributed data processing (caching, sharding, etc.). They aren’t necessarily Hadoop-related and are well beyond the scope of this book. One lesser-known tool for real-time data processing is the Bloom filter, which is a summary of a data set whose usage makes other data processing techniques more efficient. When that data set is big, Hadoop is often called in to generate the Bloom filter representation. As we mentioned earlier, a Bloom filter is also sometimes used for data joining within Hadoop itself. As a data processing expert, you’ll be well rewarded to have the Bloom filter in your bag of tricks. In this section we’ll explain this data structure in more detail and we’ll go through an online ad network example that will build a Bloom filter using Hadoop. 5.3.1 What does a Bloom filter do? At its most basic, a Bloom filter object supports two methods: add() and contains(). These two methods work in a similar way as in the Java Set interface. The method add() adds an object to the set, and the method contains() returns a Boolean true/false value denoting whether an object is in the set or not. But, for a Bloom filter, contains() doesn’t always give an accurate answer. It has no false negatives . If contains() returns false, you can be sure that the set doesn’t have the object queried. It does have a small probability of false positives though. contains() can return true for some objects not in the set. The probability of false positives depends on the number of elements in the set and some configuration parameters of the Bloom filter itself. Creating a Bloom filter 123 The major benefit of a Bloom filter is that its size, in number of bits, is constant and is set upon initialization. Adding more elements to a Bloom filter doesn’t increase its size. It only increases the false positive rate. A Bloom filter also has another configuration parameter to denote the number of hash functions it uses. We’ll discuss the reason for this parameter and how the hash functions are used later when we discuss the Bloom filter’s implementation. For now, its main implication is that it affects the false positive rate. The false positive rate is approximated by the equation (1 – exp(-kn/m))k where k is the number of hash functions used, m is the number of bits used to store the Bloom filter, and n is the number of elements to be added to the Bloom filter. In practice, m and n are determined by the requirement of the system, and therefore, k is chosen to minimize the false positive rate given m and n, which (after a little calculus) is k = ln(2) * (m/n) ≈ 0.7 * (m/n) The false positive rate with the given k is 0.6185m/n, and k has to be an integer. The false positive rate will only be an approximation. From a design point of view, one should think in terms of (m/n), number of bits per element, rather than m alone. For example, we have to store a set containing ten million URLs (n=10,000,000). Allocat- ing 8 bits per URL (m/n=8) will require a 10 MB Bloom filter (m = 80,000,000 bits). This Bloom filter will have a false positive rate of (0.6185)8, or about 2 percent. If we were to implement the Set class by storing the raw URLs, and let’s say the average URL length was 100 bytes, we would have to use 1 GB. Bloom filter has shrunk the storage requirement by 2 orders of magnitude at the expense of only a 2 percent false positive rate! A slight increase in storage allocated to the Bloom filter will reduce the false posi- tive rate further. At 10 bits per URL, the Bloom filter will take up 12.5 MB and have a false positive rate of only 0.8 percent. In summary, the signature of our Bloom filter class will look like the following: class BloomFilter { public BloomFilter(int m, int k) { ... } public void add(E obj) { ... } public boolean contains(E obj) { ... } } More applications of the Bloom filter The Bloom filter found its first successful applications back when memory was scarce. One of its first uses was in spellchecking. Not being able to store a whole dictionary in memory, spellcheckers used a Bloom filter representation (of the dictionary) to catch most misspellings. As memory size grew and became cheaper, such space consideration waned. Bloom filter usage is finding a resurgence in large-scale data- intensive operations as data is fast outgrowing memory and bandwidth. We’ve already seen commercial products, such as Oracle 11g, using Bloom filters to join data across distributed databases. In the networking world, one successful 124 CHAPTER 5 Advanced MapReduce (continued ) product using Bloom filters is the open source distributed web proxy called Squid . Squid caches frequently accessed web content to save bandwidth and give users a faster web experience. In a cluster of Squid servers, each one can cache a different set of content. An incoming request should be routed to the Squid server holding a copy of the requested content, or in case of a cache miss, the request is passed on to the originating server. The routing mechanism needs to know what each of the Squid servers contains. As sending a list of URLs for each Squid server and storing it in memory is expensive, Bloom filters are used. A false positive means a request is forwarded to the wrong Squid server, but that server would recognize it as a cache miss and pass it on to the originating server, ensuring the correctness of the overall operation. The small performance hit from a false positive is far outweighed by the overall improvement. Sharding systems are a similar application but more advanced. In a nutshell, database sharding is the partitioning of a database across multiple machines such that each machine only has to deal with a subset of records. Each record has some ID that determines which machine it’s assigned to. In more basic designs, the ID is hashed statically to one of a fixed number of database machines. This approach is inflexible to adding more shards or rebalancing existing ones. To add flexibility, it uses a dynamic look-up for each record ID, but unfortunately that adds processing delay if the look-up is done through a database (i.e., using disk). Like Squid, more advanced shard systems use in-memory Bloom filters as a fast look-up. It needs some mechanism to handle false positives, but the occurrence is small enough to not impact the overall performance improvement. For online display ad networks, it’s important to be able to target an ad from the right category to a visitor. Given the volume of traffic a typical ad network receives and the latency requirements, one can end up spending a lot of money on hardware to have the capability of retrieving the category in real time. A design based on Bloom filters can dramatically decrease that cost. Use an offline process to tag web pages (or visitors) on a limited number of categories (sports, family-oriented, music, etc.). Build a Bloom filter for each category and store it in memory at the ad servers. When an ad request arrives, the ad servers can quickly and cheaply determine which category of ads to show. The amount of false positives is negligible. 5.3.2 Implementing a Bloom filter Conceptually the implementation of a Bloom filter is quite straightforward. We describe its implementation in a single system first before implementing it using Hadoop in a distributed way. The internal representation of a Bloom filter is a bit array of size m. We have k independent hash functions, where each hash function takes an object as input and outputs an integer between 0 and m-1. We use the integer output as an index into the bit array. When we “add” an element to the Bloom filter, we use the hash functions to generate k indexes into the bit array. We set the k bits to 1. Figure 5.3 shows what happens when we add several objects (x, y, and z) over time, in a Bloom filter that uses three hash functions. Note that a bit will be set to 1 regardless of its previous state. The number of 1s in the bit array can only grow. Creating a Bloom filter 125 When an object comes in and we want to check whether it has been previously added to the Bloom filter, we use the same k hash functions to generate the bit array indexes as we would do in adding the object. Now we check whether all those k bits in the bit array are 1s. If all k bits are 1, we return true and claim that the Bloom filter contains the object. Otherwise we return false. We see that if the object has in fact been added before, then the Bloom filter will necessarily return true. There are no false negatives (returning false when the object is truly in the set). The k bits corresponding to the queried object can all be set to 1 even though the object has never been added to the set. It may happen that adding other objects set those bits leading to false positives.3 Our implementation of a Bloom filter in Java would use the Java BitSet class as its internal representation. We have a function getHashIndexes(obj) that takes an object and returns an integer array of size k, containing indexes into the BitSet. The main functions of the Bloom filter, add() and contains(), are quite straightforward: class BloomFilter { private BitSet bf; public void add(E obj) { int[] indexes = getHashIndexes(obj); for (int index : indexes) { bf.set(index); } } public boolean contains(E obj) { int[] indexes = getHashIndexes(obj); for (int index : indexes) { if (bf.get(index) == false) { return false; } } 3 For an accessible introduction to Bloom Filters, see http://en.wikipedia.org/wiki/Bloom_filter. 1000000010000100 1011000010000100 1011100010100101 1011100010100101 x y z ? Figure 5.3 A Bloom filter is a bit array that represents a set with some probability of false positives. Objects (such as x, y, and z) are deterministically hashed into positions in the array, and those bits are set to 1. You can check whether an object is in the set or not by hashing and checking the values of those bit positions. 126 CHAPTER 5 Advanced MapReduce return true; } protected int[] getHashIndexes(E obj) { ... } } To implement getHashIndexes() such that it works truly as k independent hash func- tions is nontrivial. Instead, in our Bloom filter implementation in listing 5.4, we use a hack to generate k indexes that are roughly independent and uniformly distributed. The getHashIndexes() method seeds the Java Random number generator with an MD5 hash of the object and then takes k “random” numbers as indexes. The Bloom fil- ter class would benefit from a more rigorous implementation of getHashIndexes(), but our hack suffices for illustration purposes. An ingenious way of creating a Bloom filter for the union of two sets is by OR’ing the (bit array of the) Bloom filters of each individual set. As adding an object is setting certain bits in a bit array to 1, it’s easy to see why this union rule is true: public void union(BloomFilter other) { bf.or(other.bf); } We’ll be exploiting this union trick to build Bloom filters in a distributed fashion. Each mapper will build a Bloom filter based on its own data split. We’ll send the Bloom fil- ters to a single reducer, which will take a union of them and record the final output. As the Bloom filter will be shuffled around as the mappers’ output, the BloomFilter class will have to implement the Writable interface, which consists of methods write() and readFields(). For our purpose these methods transform between the internal BitSet representation and a byte array such that the data can be serialized to DataInput/DataOutput. The final code is in listing 5.4. Listing 5.4 Basic Bloom filter implementation class BloomFilter implements Writable { private BitSet bf; private int bitArraySize = 100000000; private int numHashFunc = 6; public BloomFilter() { bf = new BitSet(bitArraySize); } public void add(E obj) { int[] indexes = getHashIndexes(obj); for (int index : indexes) { bf.set(index); } } public boolean contains(E obj) { int[] indexes = getHashIndexes(obj); for (int index : indexes) { Creating a Bloom filter 127 if (bf.get(index) == false) { return false; } } return true; } public void union(BloomFilter other) { bf.or(other.bf); } protected int[] getHashIndexes(E obj) { int[] indexes = new int[numHashFunc]; long seed = 0; byte[] digest; try { MessageDigest md = MessageDigest.getInstance("MD5"); md.update(obj.toString().getBytes()); digest = md.digest(); for (int i = 0; i < 6; i++) { seed = seed ^ (((long)digest[i] & 0xFF))<<(8*i); } } catch (NoSuchAlgorithmException e) {} Random gen = new Random(seed); for (int i = 0; i < numHashFunc; i++) { indexes[i] = gen.nextInt(bitArraySize); } return indexes; } public void write(DataOutput out) throws IOException { int byteArraySize = (int)(bitArraySize / 8); byte[] byteArray = new byte[byteArraySize]; for (int i = 0; i < byteArraySize; i++) { byte nextElement = 0; for (int j = 0; j < 8; j++) { if (bf.get(8 * i + j)) { nextElement |= 1< { OutputCollector oc = null; public void map(K1 key, V1 value, OutputCollector output, Reporter reporter) throws IOException { if (oc == null) oc = output; ... } public void close() throws IOException { oc.collect(k, v); } } The BloomFilters generated by all the mappers are sent to a single reducer. The reduce() method in the Reduce class will do a Bloom filter union of all of them. while (values.hasNext()) { bf.union((BloomFilter)values.next()); } As we mentioned earlier, we want the final BloomFilter to be written out in a file of our own format rather than one of Hadoop’s OutputFormats. We had already set the reducer’s OutputFormat to NullOutputFormat in the driver to turn off that output mechanism. Now the close() method will have to handle the file output itself. It will have to know the output path as specified by the user, which can be found in the mapred.output.dir property of the JobConf object. But the close() is not given the job configuration object. We handle this oversight the same way we handled Output- Collector in the mapper. We keep a reference to the JobConf object in the Reduce class to be used by the close() method. The rest of the close() method will use Hadoop’s file I/O to write out our BloomFilter in binary to a file in HDFS . The complete code is in listing 5.5. Listing 5.5 A MapReduce program to create a Bloom filter public class BloomFilterMR extends Configured implements Tool { public static class MapClass extends MapReduceBase implements Mapper> { BloomFilter bf = new BloomFilter(); OutputCollector> oc = null; public void map(Text key, Text value, OutputCollector> output, Reporter reporter) throws IOException { 130 CHAPTER 5 Advanced MapReduce if (oc == null) oc = output; bf.add(key.toString()); } public void close() throws IOException { oc.collect(new Text("testkey"), bf); } } public static class Reduce extends MapReduceBase implements Reducer, Text, Text> { JobConf job = null; BloomFilter bf = new BloomFilter(); public void configure(JobConf job) { this.job = job; } public void reduce(Text key, Iterator> values, OutputCollector output, Reporter reporter) throws IOException { while (values.hasNext()) { bf.union((BloomFilter)values.next()); } } public void close() throws IOException { Path file = new Path(job.get("mapred.output.dir") + "/bloomfilter"); FSDataOutputStream out = file.getFileSystem(job).create(file); bf.write(out); out.close(); } } public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, BloomFilterMR.class); Path in = new Path(args[0]); Path out = new Path(args[1]); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); job.setJobName("Bloom Filter"); job.setMapperClass(MapClass.class); job.setReducerClass(Reduce.class); job.setNumReduceTasks(1); job.setInputFormat(KeyValueTextInputFormat.class); job.setOutputFormat(NullOutputFormat.class); job.setOutputKeyClass(Text.class); job.setOutputValueClass(BloomFilter.class); job.set("key.value.separator.in.input.line", ","); JobClient.runJob(job); return 0; } Exercising what you’ve learned 131 public static void main(String[] args) throws Exception { int res = ToolRunner.run(new Configuration(), new BloomFilterMR(), args); System.exit(res); } } 5.3.3 Bloom filter in Hadoop version 0.20+ Hadoop version 0.20 has a Bloom filter class in it. It plays a support role to some of the new classes introduced in version 0.20, and it will likely stay around for future ver- sions as well. It functions much like our BloomFilter class in listing 5.4, although it’s much more rigorous in its implementation of the hashing functions. As a built-in class, it can be a good choice for semijoin within Hadoop. But it’s not easy to separate this class from the Hadoop framework to use it as a standalone class. If you’re building a Bloom filter for non-Hadoop applications, Hadoop’s built-in BloomFilter may not be appropriate. 5.4 Exercising what you’ve learned You can test your understanding of more advanced MapReduce techniques through these exercises: 1 Anomaly detection —Take a web server log file . Write a MapReduce program to aggregate the number of visits for each IP address. Write another MapReduce program to find the top K IP addresses in terms of visits. These frequent visitors may be legitimate ISP proxies (shared among many users) or they may be scrapers and fraudsters (if the server log is from an ad network). Chain these two MapReduce jobs together such that they can be easily run on a daily basis. 2 Filter out records in input—In both patent data sets we’ve used (cite75_99.txt and apat63_99.txt), the first row is metadata (column names). So far we’ve had to explicitly or implicitly filter out that row in our mappers, or interpret our results knowing that the metadata record has some deterministic influence. A more permanent solution is to remove the metadata row from the input data and keep track of it elsewhere. Another solution is to write a mapper as a preprocessor that filters all records that look like metadata. (For example, records that don’t start with a numeric patent number.) Write such a mapper and use ChainMapper /ChainReducer to incorporate it into your MapReduce programs. 3 Disjoint selection—Using the same Customers and Orders example for the datajoin package , how will you change the code to output customers not in the Orders data source? Perhaps the Orders data only contains orders made within the last N months, and these customers haven’t purchased anything in that time period. 132 CHAPTER 5 Advanced MapReduce A business may choose to re-engage these customers with discounts or other incentives. 4 Calculating ratios—Ratios are often a better unit of analysis than raw numbers. For example, say you have a data set of today’s stock prices and another data set for stock prices from yesterday. You may be more interested in each stock’s growth rate than its absolute price. Use the datajoin framework to write a program that takes two data sources and output the ratio. 5 Product of a vector with a matrix—Look up your favorite linear algebra text on the definition of matrix multiplication. Implement a MapReduce job to take the product of a vector and a matrix. You should use DistributedCache to hold the value of the vector. You may assume the matrix is in sparse representation . 6 Spatial join —Let’s get more adventurous. Consider a two-dimensional space where both the x and y coordinates range from -1,000,000,000 to +1,000,000,000. You have one file with the location of foos, and another file with the location of bars. Each record in those files is a comma-separated (x,y) coordinate. For example, a couple lines may look like 145999.32455,888888880.001 834478899.2,5656.87660922 Write a MapReduce job to find all foos that are less than 1 unit distance from a bar. Distance is measured using the familiar Euclidean distance, sqrt[(x1- x2)2 + (y1-y2)2]. Although both foos and bars are relatively sparse in this 2D space, their respective files are too big to be stored in memory. You can’t use DistributedCache for this spatial join. Hint: The datajoin package as it’s currently implemented doesn’t work that well for this problem either, but you can solve it with your own mapper and reducer that have a similar data flow as the datajoin package. Hint #2: In all the MapReduce programs we’ve discussed up till now, the keys are only extracted and passed around, whereas the values go through various computations. You should consider computing the key for the mapper’s output. 7 Spatial join, enhanced with Bloom filter—After you’ve answered the last question, figure out how you can use a Bloom filter to speed up the join operation. Assume bars are much fewer in number than foos, but still too many to fit all their locations in memory. 5.5 Summary We can often write the basic MapReduce programs as one job operating on one data set. We may need to write the more advanced programs as multiple jobs or we may operate them on multiple data sets. Hadoop has several different ways of coordinating multiple jobs together, including sequential chaining and executing them according to predefined dependencies. For the frequent case of chaining map-only jobs around a full MapReduce job, Hadoop has special classes to do it efficiently. Further resources 133 Joining is the canonical example for processing data from multiple data sources. Though Hadoop has a powerful datajoin package for doing arbitrary joins, its generality comes at the expense of efficiency. A couple other joining methods can provide faster joins by exploiting the relative asymmetry in data source sizes typical of most data joins. One of these methods leverages the Bloom filter, a data structure that’s useful in many data processing tasks. At this point, your knowledge of MapReduce programming should enable you to start writing your own programs. As all programmers know, programming is more than writing code. You have various techniques and processes—from development to deployment and testing and debugging. The nature of MapReduce programming and distributed computing adds complexity and nuance to these processes, which we’ll cover in the next chapter. 5.6 Further resources http://portal.acm.org/citation.cfm?doid=1247480.1247602—MapReduce’s lack of simple support for joining datasets is well-known. Many of the tools to enhance Hadoop (such as Pig, Hive, and CloudBase) offer data joins as a first-class operation. For a more formal treatment, Hung-chih Yang and coauthors have published a paper “Map-reduce-merge: simplified relational data processing on large clusters” that proposes a modified form of MapReduce with an extra “merge” step that supports data joining natively. http://umiacs.umd.edu/~jimmylin/publications/Lin_etal_TR2009.pdf—Section 5.2.2 describes the use of distributed cache to provide side data to tasks. The limitation of this technique is that the side data is replicated to every TaskTracker, and the side data must fit into memory. Jimmy Lin and colleagues explore the use of memcached , a distributed in-memory object caching system, to provide global access to side data. Their experience is summarized in the paper “Low-Latency, High-Throughput Access to Static Global Resources within the Hadoop Framework.” 6 134 Programming Practices This chapter covers ■ Best practices unique to developing Hadoop programs ■ Debugging programs in local, pseudo-distributed, and fully distributed modes ■ Logging and monitoring Sanity checking and regression testing program outputs ■ Performance tuning■ Now that you’ve gone through various programming techniques in MapReduce, this chapter will step back and cover programming practices. Programming on Hadoop differs from traditional programming mainly in two ways. First, Hadoop programs are primarily about processing data. Second, Hadoop programs are run over a distributed set of computers. These two differences will change some aspects of your development and debugging processes, which we cover in sections 6.1 and 6.2. Performance tuning techniques tend to be specific to the programming platform, and Hadoop is no different. We cover tools and approaches to optimizing Hadoop programs in section 6.3. Developing MapReduce programs 135 Let’s start with the development techniques applicable to Hadoop. Presumably you’re already familiar with standard Java software engineering techniques. We focus on practices unique to data-centric programming within Hadoop. 6.1 Developing MapReduce programs Chapter 2 discussed the three modes of Hadoop: local (standalone ), pseudo-distributed , and fully distributed . They correspond roughly to development, staging, and produc- tion setups. Your development process will go through each of the three modes. You’ll have to be able to switch between configurations easily. In practice you may even have more than one fully distributed cluster. Larger shops may, for example, have a “devel- opment” cluster to further harden MapReduce programs before running them on the real production cluster. You may have multiple clusters for different workloads. For example, there can be an in-house cluster for running many small- to medium-sized jobs and a cluster in the cloud that’s more cost effective for running large, infre- quent jobs. Section 2.3 discussed how you can have different versions of the hadoop-site. xml configuration file for different setups, and you switch a symlink to point to the configuration you want to work with at the moment. You can also specify the exact configuration file you want at each Hadoop command with the -conf option. For example, bin/hadoop fs -conf conf.cluster/hadoop-site.xml -lsr will list all files in your fully distributed cluster, even though you may be currently work- ing on a different mode or different cluster (assuming conf.cluster/hadoop-site. xml is where your fully distributed cluster’s configuration file is). Before you run and test your Hadoop program, you’ll need to make data available for the configuration you’re running. Section 3.1 describes various ways to get data into and out of HDFS. For local and pseudo-distributed modes, you’ll only want a subset of your full data. Section 4.4 presents a Streaming program (RandomSample. py) that can randomly sample a percentage of records from a data set in HDFS . As it’s a Python script, you can also use it to sample down a local file with a Unix pipe: cat datafile | RandomSample.py 10 will give you a 10 percent sample of the file datafile. Now that you have all the different configurations set up and know how to put data into each configuration, let’s look at how to develop and debug in local and pseudo- distributed modes. The techniques build on top of each other as you get closer to the production environment. We defer the discussion of debugging on the fully distributed cluster ‘till the next section. 6.1.1 Local mode Hadoop in local mode runs everything within one single Java Virtual Machine (JVM) and uses the local filesystem (i.e., no HDFS). Running within one JVM allows you to 136 CHAPTER 6 Programming Practices use all the familiar Java development tools, such as a debugger . Using files from the local filesystem means you can quickly apply Unix commands or simple scripts on the input and output data. Examining files in HDFS, on the other hand, is limited to commands provided by the Hadoop command line. For example, to count how many records are in an output file, you can use wc -l if the file is in the local filesystem. If the file is in HDFS, then you’ll either have to write a MapReduce program or download the file to local storage before applying the Unix commands. As you’ll see, being able to access input and output files easily will be important to our development practices under local mode. NOTE Local mode closely adheres to Hadoop’s MapReduce programming model, but it doesn’t support every feature. For example, it doesn’t support distributed cache , and it only allows a maximum of one reducer. A program running in local mode will output all log and error messages to the con- sole. It will also summarize the amount of data processed at the end. For example, running our skeleton MapReduce job (MyJob.java) to invert the patent citation data , the output is quite verbose, and figure 6.1 is a snapshot in the middle of the job. At the end of the job, Hadoop will print out the values of various internal counters . They’re the number of records and bytes going through the different stages of MapReduce: 09/05/27 03:34:37 INFO mapred.TaskRunner: Task ➥ attempt_local_0001_r_000000_0' done. 09/05/27 03:34:37 INFO mapred.TaskRunner: Saved output of task ➥ attempt_local_0001_r_000000_0' to ➥ file:/Users/chuck/Projects/Hadoop/hadoop-0.18.1/output/test Figure 6.1 Running a Hadoop program in local mode outputs all the log messages to the console. Developing MapReduce programs 137 09/05/27 03:34:37 INFO mapred.LocalJobRunner: reduce > reduce 09/05/27 03:34:37 INFO mapred.JobClient: Job complete: job_local_0001 09/05/27 03:34:37 INFO mapred.JobClient: Counters: 11 09/05/27 03:34:37 INFO mapred.JobClient: Map-Reduce Framework 09/05/27 03:34:37 INFO mapred.JobClient: Map output records=16522439 09/05/27 03:34:37 INFO mapred.JobClient: Reduce input records=33044878 09/05/27 03:34:37 INFO mapred.JobClient: Map output bytes=264075431 09/05/27 03:34:37 INFO mapred.JobClient: Map input records=16522439 09/05/27 03:34:37 INFO mapred.JobClient: Combine output records=0 09/05/27 03:34:37 INFO mapred.JobClient: Map input bytes=264075431 09/05/27 03:34:37 INFO mapred.JobClient: Combine input records=0 09/05/27 03:34:37 INFO mapred.JobClient: Reduce input groups=6517968 09/05/27 03:34:37 INFO mapred.JobClient: Reduce output records=6517968 09/05/27 03:34:37 INFO mapred.JobClient: File Systems 09/05/27 03:34:37 INFO mapred.JobClient: Local bytes written=4246405780 09/05/27 03:34:37 INFO mapred.JobClient: Local bytes read=4276658154 The input and output of the MapReduce job are both in the local filesystem. We can examine them using standard Unix commands such as wc -l or head. As we are delib- erately using smaller data sets during development, we can even load them into a text editor or a spreadsheet. We can use the many features of those applications to sanity check the correctness of our program. SANITY CHECKING Most MapReduce programs involve at least some counting or arithmetic, and bugs (especially typos) in mathematical programming don’t call attention to themselves in the form of thrown exceptions or threatening error messages. Your math can be wrong even though your program is technically “correct,” and everything will run smoothly, but the end result will be useless. There’s no simple way to uncover arithmetic mis- takes, but some sanity checking will go a long way. At a high level you can look at the overall count, maximum, average, and so on, of various metrics and see if they match expectation. At a low level you can pick a particular output record and verify that it was produced correctly. For example, when we created the inverted citation graph, the first few lines were "CITED" "CITING" 1 3964859,4647229 10000 4539112 100000 5031388 1000006 4714284 1000007 4766693 1000011 5033339 1000017 3908629 1000026 4043055 1000033 4190903,4975983 Our job concludes that patent number 1 is cited twice, by 3964859 and 4647229. We can verify this claim by grepping over the sampled input data to look for records where patent number 1 is cited. grep ",1$" input/cite75_99.txt 138 CHAPTER 6 Programming Practices We indeed get the two records as expected. You can verify a few more records to gain confidence in the correctness of your program’s math and logic.1 An eyesore about the output of this inverted citation graph is that the first line is not real data. "CITED" "CITING" It’s an artifact from the first line of the input data being used as data definition. Let’s add some code to our mapper to filter out non-numeric keys and values, and in the process demonstrate regression testing. REGRESSION TESTING Our data-centric approach to regression testing revolves around “diff ’ing” various out- put files from before and after code changes . For our particular change, we should only be taking out one line from the job’s output. To verify that this indeed is the case, let’s first save the output of our current job. In local mode, we have a maximum of only one reducer, so the job’s output is only one file, which we call job_1_output. For regression testing, it’s also useful to save the output of the map phase. This will help us isolate bugs to either the map phase or the reduce phase. We can save the output of the map phase by running the MapReduce job with zero reducers. We can do this easily using the -D mapred.reduce.tasks=0 option . In this mapper-only job, there will be multiple files as each map task will write its output to its own file. Let’s copy all of them into a directory called job_1_intermediate. Having stored away the output files, we can make the desired code changes to the map() method in MapClass. The code itself is trivial. We focus on testing it. public void map(Text key, Text value, OutputCollector output, Reporter reporter) throws IOException { try { if (Integer.parseInt(key.toString()) > 0 && Integer.parseInt(value.toString()) > 0) { output.collect(value, key); } } catch (NumberFormatException e) { } } Compile and execute the new code against the same input data. Let’s run it as a map- only job first and compare the intermediate data. As we’ve only changed the mapper, any bug should first manifest in differences in the intermediate data. diff output/job_1_intermediate/ output/test/ 1 In this case, you may suspect whether patent number 1 is really cited by those two patents. The number 1 feels wrong, an outlier in the range of patent numbers being cited. There can be mistakes in the original input data. We have to track down the patents themselves if we want to verify this. In any case, ensuring data quality is an important topic but is beyond our discussion of Hadoop. Developing MapReduce programs 139 We get the following output from the diff utility: Binary files output/job_1_intermediate/.part-00000.crc and ➥ output/test/.part-00000.crc differ diff output/job_1_intermediate/part-00000 output/test/part-00000 1d0 < "CITED" "CITING" We found differences in the binary file .part-00000.crc. This is an internal file for HDFS to keep checksums for the file part-00000. A difference in checksum means that part-00000 has changed, and diff prints out the exact differences later. The new intermediate file, under output/test, is missing the quoted field descriptors. More importantly, we find no other changes. So far so good. If we run the whole job with one reducer, we expect the final output to differ by one line too. Well, it turns out not to be the case. If you run the whole job with one reducer and compare the final output with job_1_output from the original run, you’ll find many differences. What do you think happened? Let’s look at the first few lines of the diff to find out. $ diff output/job_1_output output/test/part-00000 | head -n 15 1,2c1 < "CITED" "CITING" < 1 3964859,4647229 --- > 1 4647229,3964859 19c18 < 1000067 5312208,4944640,5071294 --- > 1000067 4944640,5071294,5312208 22,23c21,22 < 1000076 4867716,5845593 < 1000083 5566726,5322091 --- > 1000076 5845593,4867716 > 1000083 5322091,5566726 We see that the line with the field descriptors (“CITED” and “CITING”) are taken out as expected. As to the rest of the differences, there’s a definite pattern. In our reduce() method, we have concatenated the list of values for each key in the order Hadoop has given them to us. Hadoop doesn’t provide any guarantee as to the order of those values. We see that taking out one line in the intermediate data impacts the order of the values for many keys at the reducer . As we know that this job’s correctness is invariant to the ordering, we can ignore the differences. Regression testing is inherently conservative and tends to set off false alarms. You should use it with that in mind. We have advocated the use of a sampled data set for development, because it is more representative of the structure and properties of the data set we use in production. We have used the same sampled data set for regression testing, but you can also manually construct a separate input data set with edge cases that are atypical of the production data. For example, you may put in empty values or extra tab characters or other unusual 140 CHAPTER 6 Programming Practices records in this constructed data set. This test data set is for ensuring that your program continues to handle the edge cases even as it evolves. This test data set doesn’t need to have more than several dozen records. You can visually inspect the entire output to see if your program still functions as expected. CONSIDER USING LONG INSTEAD OF INT Most Java programmers instinctively default to the int type (or Integer or IntWritable ) to represent integer values. In Java the int type can hold any integer between 231-1 and -231, or between 2,147,483,647 and -2,147,483,648. This is adequate for most applica- tions. Rarely do programmers put too much thought into it. When you’re processing Hadoop-scale data, it’s not unusual for some counter variables to need a bigger range. You won’t see this requirement under your development data set , which by design is small. It may not even matter to your current production data, but as your business operation grows, your data set will get bigger. It may get to a point where some vari- ables will outgrow the int range and cause arithmetic errors. Take the canonical word counting example. When processing millions of documents, you won’t have any word count that goes beyond 2 billion,2 and an int is adequate. But as you grow to process tens of millions or hundreds of millions of documents, counting frequent words like the can cross the limit of an int type. Rather than wait for this kind of bug to creep up on you in production, which is much harder to debug and costlier to fix, now is the time you should go through your code and carefully consider whether your numeric variables should be long or LongWritable to handle future scale.3 6.1.2 Pseudo-distributed mode Local mode has none of the distributed characteristics of a production Hadoop cluster. We may not see many bugs when running in local mode. Hadoop provides a pseudo- distributed mode that has all the functionality and “nodes” of a production cluster—Name Node , SecondaryNameNode , DataNode , JobTracker , and TaskTracker , each running on a separate JVM. All the software components are distributed, and pseudo-distributed mode differs from a production cluster at only the system-and-hardware level. It uses only one physical machine—your own local computer. We should make sure our jobs can run in pseudo-distributed mode before deploying them to a full production cluster. Chapter 2 describes the configuration and commands to start pseudo-distributed mode. You’ll start all the daemons on your computer to make it function like a cluster. You interface with it as if it is a distinct Hadoop cluster. You put data into its own HDFS filesystem. You submit your jobs to it for running rather than run them in the same user space. Most importantly, you now monitor it “remotely” through log files and the web interface. You’ll use the same tools later to monitor a production cluster . 2 This is not absolutely true and will depend on your documents’ size and content. 3 The problem of exceeding a numeric range is not unique to Hadoop. You’ll remember the famous Y2K problem where older programs only allocated two digits to represent year. More recently, almost all web operations that have experienced explosive growth (such as Facebook, Twitter, and RockYou) have had to retool their systems to handle a bigger range of user IDs or document IDs than they originally expected. Developing MapReduce programs 141 LOGGING Let’s run in the pseudo-distributed cluster the same job we had in local mode. You put the input file into HDFS using the hadoop fs command. Submit the job for running using the same hadoop jar command as in local mode. The first thing you’ll notice is that you no longer have the torrent of messages on your console. You only get a measure of progress in the map phase and reduce phase, and the same summary of counters at the end as in local mode. You can see this in figure 6.2. Figure 6.2 In pseudo-distributed mode, the console only outputs a job’s progress and its counters at the end. 142 CHAPTER 6 Programming Practices Hadoop hasn’t stopped outputting debugging messages. In fact, it’s outputting much more now. These messages don’t go to the console screen. Instead, they’re saved into log files. You can find the log files under the /logs directory. Different services (NameNode, JobTracker, etc.) create separate log files. The filename should distinguish the service logging a file. Hadoop rotates log files daily. The most recent one ends in .log. It further appends the older ones with their date. Under the default setting, Hadoop doesn’t delete old log files automatically. You should proactively archive and delete them to make sure they’re not taking up too much space. Log files for the NameNode , SecondaryNameNode , DataNode , and JobTracker are used for debugging the respective services. They’re not too important in pseudo- distributed mode. In production clusters, you as a system administrator can look at them to debug problems in those corresponding nodes. As a programmer, you are always interested in the TaskTracker log though, as it records exceptions thrown. Your MapReduce program can output to STDOUT and STDERR (System.out and System.err in Java) its own logging messages. Hadoop records those under files named stdout and stderr, respectively. There will be a distinct file for each task attempt. (A task can have more than one attempt if the first one fails.) These user log files are under the /logs/userlogs subdirectory. Besides logging to STDOUT and STDERR, your program can also send out live status messages using the setStatus() method on the Reporter object being passed to the map() and reduce() methods. (For Streaming programs, the status information is updated by sending a string of the form reporter:status:message to STDERR.) This is useful for long-running jobs where you can monitor them as they run. The status message is shown on the JobTracker Web UI, to be described next. JOBTRACKER WEB UI By definition events occur in many different places in a distributed program. This makes monitoring more difficult. The system becomes more like a black box, and we need specialized monitoring tools to peek into the various states within it. The Job- Tracker provides a web interface for tracking the progress and various states of your jobs. Under the default configuration, you can set your browser to http://localhost:50030/jobtracker.jsp to view the starting page of the administration tool for your pseudo-distributed cluster.4 It shows a summary of the Hadoop cluster, as well as lists of jobs that are running, completed, and failed. See figure 6.3. Hadoop tracks jobs internally by their job ID . A job ID is a string prefixed with job_, followed by the cluster ID (which is a timestamp of when the cluster was started), followed by an auto-incremented job number. The web UI lists each job with the user name and job name . In pseudo-distributed mode, it’s relatively easy to identify the job 4 In fully distributed mode, replace “localhost” with the domain of the JobTracker master. Developing MapReduce programs 143 Figure 6.3 The JobTracker Web UI main page you’re currently working with, as you’ll run one job at a time. When you get to a multi- user production environment, you’ll have to narrow down your jobs by looking for your Hadoop user name and the name of your current job. The name of your job is set using the setJobName() method in the JobConf object. The name for a Streaming job is set through a configuration property shown in table 6.1. Table 6.1 Configuration property for setting a job’s name Property Description mapred.job.name String property denoting the name of a job 144 CHAPTER 6 Programming Practices In the administration page, you can see each job with the completion percentage of its map phase. It shows the number of map tasks for the job and the number of completed ones. You can see the same metrics for the reduce side. This gives you a rough summa- ry of your job’s progress. To drill down more on a particular job, you can click on the job ID, which is a link that’ll take you to the job’s administration page. See figure 6.4. The job page shows the volume of various input/output due to the running of the job. The page refreshes itself periodically but you can also refresh the page manually to get the updated numbers. You can start exploring the various aspects of your job from the many links on this page. For example, clicking on the map link will take you to a list of all map tasks for the job. See figure 6.5. Figure 6.4 The JobTracker ’s administration page for a single job Monitoring and debugging on a production cluster 145 Figure 6.5 List of tasks in the TaskTracker Web UI. This figure shows all the map tasks for a single job. Each task can update its own status message. Tasks are identified by a task ID . To construct the task ID, you start with the job ID the task runs under but replace the job_ prefix with task_. You then append it with _m for a map task or _r for a reduce task. You further append it with an auto-incremented number within each group. In the TaskTracker Web UI, you’ll see each task with its status , which you can programmatically set through the setStatus() method described earlier. Clicking on a task ID will bring you to a page that further describes different attempts of a task. Hadoop makes several retry attempts at a failed task before failing the entire job. The JobTracker and TaskTracker UIs provide many other links and metrics. Most should be self-explanatory. KILLING JOBS Unfortunately, sometimes a job goes awry after you’ve started it but it doesn’t actually fail. It may take a long time to run or may even be stuck in an infinite loop. In (pseudo-) distributed mode you can manually kill a job using the command bin/hadoop job -kill job_id where job_id is the job’s ID as given in JobTracker’s Web UI. 6.2 Monitoring and debugging on a production cluster After successfully running your job in a pseudo-distributed cluster, you’re ready to run it on a production cluster using real data. We can apply all the techniques we’ve used for development and debugging on the production cluster, although the exact usage may be slightly different. Your cluster should still have a JobTracker Web UI, but the domain is no longer localhost. It’s now the address of the cluster’s JobTracker. The port number will still be 50030 unless it’s been configured differently. 146 CHAPTER 6 Programming Practices In pseudo-distributed mode, when there’s only one node, all the log files are in a single /logs directory that you can access locally. In a fully distributed cluster, each node has its own /logs directory to keep its log files. You can diagnose problems on a node through the log files of that particular node. In addition to the development and testing techniques we’ve mentioned so far, you also have monitoring and debugging techniques that are more useful in a production cluster on real data, which we explore in this section. 6.2.1 Counters You can instrument your Hadoop job with counters to profile its overall operation. Your program defines various counters and increments their counts in response to specific events. Hadoop automatically sums the same counter from all tasks (of the same job) so that it reflects the profile of the overall job. It displays the value of your counters in the JobTracker’s Web UI along with Hadoop’s internal counters. The canonical application of counters is for tracking different input record types, particularly for tracking “bad” records. Recall from section 4.4 our example for finding the average number of claims for patents from each country. We know the number of claims is not available for many records. Our program skips over those records, and it’s useful to know the number of records we’re skipping. Beyond satisfying our curiosity, such instrumentation allows us to understand the program’s operation and do some “reality checks” for its correctness. We use counters through the Reporter.incrCounter() method. The Reporter object is passed to the map() and reduce() methods. You call incrCounter() with the name of the counter and the amount to increment. You use uniquely named counters for each different event. When you call incrCounter() with a new counter name, that counter is initialized and takes on the increment value. The Reporter.incrCounter() method has two signatures, depending on how you want to specify a counter’s name: public void incrCounter(String group, String counter, long amount) public void incrCounter(Enum key, long amount) The first form is more general in that it allows you to specify the counter name with dynamic strings at run time. The combination of two strings, group and counter, uniquely defines a counter. When counters are reported (in the Web UI or as text at the end of a job run), counters of the same group are reported together. The second form uses a Java enum to specify counter names, which forces you to have them defined at compile time, but it also allows for type checking. The enum’s name is used as the group string, whereas the enum’s field is used as the counter string. Listing 6.1 is the MapClass from listing 4.12 rewritten with counters to track the number of missing values and “quoted” values. (Only the first row of column description should be a “quoted” value.) An enum called ClaimsCounters is defined with values MISSING and QUOTED. Logic in the code increments the counters to reflect the record it’s processing. Monitoring and debugging on a production cluster 147 Listing 6.1 A MapClass with Counters to count the number of missing values public static class MapClass extends MapReduceBase implements Mapper { static enum ClaimsCounters { MISSING, QUOTED }; public void map(LongWritable key, Text value, OutputCollector output, Reporter reporter) throws IOException { String fields[] = value.toString().split(",", -20); String country = fields[4]; String numClaims = fields[8]; if (numClaims.length() == 0) { reporter.incrCounter(ClaimsCounters.MISSING, 1); } else if (numClaims.startsWith("\"")) { reporter.incrCounter(ClaimsCounters.QUOTED, 1); } else { output.collect(new Text(country), new Text(numClaims + ",1")); } } } After running the program, we can see the defined counters along with Hadoop’s internal counters in the JobTracker’s Web UI . See figure 6.6. Figure 6.6 JobTracker ’s Web UI collects and shows the counter information. 148 CHAPTER 6 Programming Practices We see that the enum’s fully qualified Java name (with $ to separate out the inner class hierarchy) is used as the group name. The fields MISSING and QUOTED are used to define separate counters. As expected, it increments the QUOTED counter only once and the MISSING counter 939,867 times. Does the data set have that many rows with missing claim counts? The originator of the data set stated that claim counts are missing for patents granted before 1975. Merely eyeballing figure 4.3, we guess that about a third of all the patents in our data set are granted before 1975. Looking at the map input records count (from figure 6.6) we see there’s a total of 2.9M+ records. The numbers seem consistent and we can feel more confident about the correctness of the processing. A Streaming process can also use counters. It needs to send a specially formatted line to STDERR in the form of reporter:counter:group,counter,amount where group, counter, and amount are the corresponding arguments one would’ve passed to incrCounter() in Java. For example, in Python one can increment the ClaimsCounters. MISSING counter with sys.stderr.write("reporter:counter:ClaimsCounters,MISSING,1\n") Be sure to include the newline character (“ \n”) at the end. Hadoop Streaming will not properly interpret the string without that. 6.2.2 Skipping bad records When dealing with large data sets, it is inevitable that some records will have errors. It’s not unusual to focus several iterations of your development cycle on making the pro- gram robust to unexpected data.5 Your program may never be completely foolproof, though. Your program will process new data, and new data will think of new ways to misbehave. You may even be using a parser that depends on third-party libraries you have no control over. While you should make your program as robust as possible to malformed records, you should also have a recovery mechanism to handle the cases you couldn’t plan for. You don’t want your whole job to fail only because it fails to handle one bad record. Hadoop’s mechanism for recovering from hardware failures doesn’t work for recovering from deterministic software failures caused by bad records. Instead it provides a feature for skipping over records that it believes to be crashing a task. If this skipping feature is on, a task will enter into skipping mode after the task has been retried several times. Once in skipping mode, the TaskTracker will track and determine which record range is causing failure. The TaskTracker will then restart the task but skip over the bad record range. 5 Unexpected data are not always mistakes. Someone once told me he had a program that was crashing in processing users’ geographical information. Further digging revealed that one user was from a real city named Null. Monitoring and debugging on a production cluster 149 CONFIGURING RECORD SKIPPING IN JAVA The skipping feature is available starting with version 0.19, but it’s disabled by default. In Java, the feature is controlled through the class SkipBadRecords, which consists entirely of static methods. The job driver needs to call one or both methods: public static void setMapperMaxSkipRecords(Configuration conf, long maxSkipRecs) public static void setReducerMaxSkipGroups(Configuration conf, long maxSkipGrps) to turn on record skipping for map tasks and reduce tasks, respectively. The driver calls the methods with the configuration object and the maximum number of records in a skip range. If the maximum skip range size is set to 0 (default), then record skipping is disabled. Hadoop finds the skip range using a divide-and-conquer approach. It executes the task with the skip range halved each time, and determines the half with the bad record(s). The process iterates until the skip range is within the acceptable size. This is a rather expensive operation, particularly if the maximum skip range size is small. You may need to increase the maximum number of task attempts in Hadoop’s normal task recovery mechanism to accommodate the extra attempts. You can do this using the methods JobConf.setMaxMapAttempts() and JobConf.setMaxReduceAttempts() , or set the equivalent properties mapred.map.max.attempts and mapred.reduce. max.attempts . If skipping is enabled, Hadoop enters skipping mode after the task has failed twice. You can set the number of task failures needed to trigger skipping mode in SkipBadRecords’s setAttemptsToStartSkipping() method: public static void setAttemptsToStartSkipping(Configuration conf, int attemptsToStartSkipping) Hadoop will log skipped records to HDFS for later analysis. They’re written as se- quence files in the _log/skip directory. We cover sequence files in more detail in section 6.3.3. For now you can think of it as a Hadoop-specific compressed format. It can be uncompressed and read using the command: bin/hadoop fs -text You can change the log directory for skipped records from _log/skip using the method SkipBadRecords.setSkipOutputPath(JobConf conf, Path path). If path is set to null or to a Path with a string value of "none", Hadoop will not record the skipped records. CONFIGURING RECORD SKIPPING OUTSIDE OF JAVA Although you can set the record-skipping feature in Java by calling methods in Skip- BadRecords in your driver, sometimes you may want to set this feature using the ge- neric options available in GenericOptionsParser instead. This is because the person running the program can have a better idea about the range of bad records to expect and set the parameters more appropriately than the original developer. Furthermore, Streaming programs can’t access SkipBadRecords; the record skipping features must 150 CHAPTER 6 Programming Practices Table 6.2 Equivalent JobConf properties to method calls in SkipBadRecords SkipBadRecords method JobConf property setAttemptsToStartSkipping() mapred.skip.attempts.to.start. skipping setMapperMaxSkipRecords() mapred.skip.map.max.skip.records setReducerMaxSkipGroups() mapred.skip.reduce.max.skip.groups setSkipOutputPath() mapred.skip.out.dir setAutoIncrMapperProcCount() mapred.skip.map.auto.incr.proc.count setAutoIncrReducerProcCount() mapred.skip.reduce.auto.incr.proc. count be configured using Streaming’s -D property (-jobconf in version 0.18). Table 6.2 shows the JobConf properties being set by the SkipBadRecords method calls. We haven’t explained the last two properties yet. Their default values are fine for most Java programs but we need to change them for Streaming ones. In determining the record range to skip, Hadoop needs an accurate count of the number of records a task has processed. Hadoop uses an internal counter and by default it’s incremented after each call to the map (reduce) function. For Java programs this is a good approach to track the number of records processed. It can break down in some cases, such as programs that process records asynchronously (say, by spawning threads) or buffer them to process in chunks, but it usually works. In Streaming programs, this default behavior wouldn’t work at all because there’s no equivalent of the map (reduce) function that gets called to process each record. In those situations you have to disable the default behavior by setting the Boolean properties to false, and your task has to update the record counters itself. In Python, the map task can update the counter with sys.stderr.write( ➥ "reporter:counter:SkippingTaskCounters,MapProcessedRecords,1\n") and the reduce task can use sys.stderr.write( ➥ "reporter:counter:SkippingTaskCounters,ReduceProcessedGroups,1\n") Java programs that cannot depend on the default record counting should use reporter.incrCounter(SkipBadRecords.COUNTER_GROUP, SkipBadRecords.COUNTER_MAP_PROCESSED_RECORDS, 1); and reporter.incrCounter(SkipBadRecords.COUNTER_GROUP, SkipBadRecords.COUNTER_REDUCE_PROCESSED_GROUPS, 1); when it has processed a key/value pair in its Mapper and Reducer, respectively. Monitoring and debugging on a production cluster 151 6.2.3 Rerunning failed tasks with IsolationRunner Debugging through log files is about reconstructing events using generic historical records. Sometimes there’s not enough information in the logs to trace back the cause of failure. Hadoop has an IsolationRunner utility that functions like a time machine for debugging. This utility can isolate and rerun the failed task with the exact same input on the same node. You can attach a debugger to monitor the task as it runs and focus on gathering evidence specific to the failure. To use the IsolationRunner feature, you must run your job with the configuration property keep.failed.tasks.files set to true. This tells every TaskTracker to keep all the data necessary to rerun the failed tasks. When a job fails, you use the JobTracker Web UI to locate the node, the job ID, and the task attempt ID of the failed task. You log into the node where the task failed and go to the work directory under the directory for the task attempt. Go to local_dir/taskTracker/jobcache/job_id/attempt_id/work where job_id and attempt_id are the job ID and task attempt ID of the failed task. (The job ID should start with “job_” and the task attempt ID should start with “attempt_”.) The root directory local_dir is what is set in the configuration property mapred.local. dir . Note that Hadoop allows a node to use multiple local directories (by setting mapred.local.dir to a comma-separated list of directories) to spread out disk I/O among multiple drives. If the node is configured that way, you’ll have to look in all the local directories to find the one with the right attempt_id subdirectory. Within the work directory you can execute IsolationRunner to rerun the failed task with the same input that it had before. In the rerun, we want the JVM to be enabled for remote debugging. As we’re not running the JVM directly but through the bin/ hadoop script, we specify the JVM debugging options through HADOOP_OPTS: export HADOOP_OPTS="-agentlib:jdwp=transport=dt_socket, ➥ server=y,address=8000" It tells the JVM to listen for the debugger at port 8000 and to wait for the debugger getting attached before running any code.6 We now use IsolationRunner to rerun the task: bin/hadoop org.apache.hadoop.mapred.IsolationRunner ../job.xml The job.xml file contains all the configuration information IsolationRunner needs. Given our specification, the JVM will wait for a debugger’s attachment before execut- ing the task. You can attach to the JVM any Java debugger that supports the Java Debug Wire Protocol (JDWP). All the major Java IDEs do so. For example, if you’re using jdb , you can attach it to the JVM via jdb -attach 8000 6 Options to configure the Sun JVM for debugging are further explained in Sun’s documentation: http:// java.sun.com/javase/6/docs/technotes/guides/jpda/conninv.html#Invocation. 152 CHAPTER 6 Programming Practices (Of course, this is only an example. I hope you’re using something better than jdb!) Consult your IDE’s documentation for how to connect its debugger to a JVM. 6.3 Tuning for performance After you have developed your MapReduce program and fully debugged it, you may want to start tuning it for performance. Before doing any optimization, note that one of the main attractions of Hadoop is its linear scalability . You can speed up many jobs by adding more machines. This makes economic sense when you have a small cluster. Consider the value of time it takes to optimize your program to gain a 10 percent im- provement. For a 10-node cluster, you can get the same 10 percent performance gain by adding one machine (and this gain applies to all jobs on that cluster). The cost of your development time may well be higher than the cost of the additional computer. On the other hand, for a 1,000-node cluster, squeezing a 10 percent improvement through hardware will take 100 new machines. At that scale the brute force approach of adding hardware to boost performance may be less cost effective. Hadoop has a number of specific levers and knobs for tuning performance, some of which boost the effectiveness of the cluster as a whole. We cover those in the next chapter when we discuss system administration issues. In this section we examine techniques that can be applied on a per-job basis. 6.3.1 Reducing network traffic with combiner Combiner can reduce the amount of data shuffled between the map and reduce phases, and lower network traffic improves execution time. The details and the benefits of using combiner are thoroughly described in section 4.6. We mention it here again for the sake of completeness. 6.3.2 Reducing the amount of input data When processing large data sets , sometimes a nontrivial portion of the processing time is spent scanning data from disk. Reducing the number of bytes to read can enhance overall throughput. There are several ways to do this. The simplest way to reduce the amount of bytes processed is to reduce the amount of data processed. We can choose to process only a sampled subset of the data. This is a viable option for certain analytics applications. For those applications, sampling reduces precision but not accuracy. Their results remain useful for many decision support systems. Often your MapReduce jobs don’t use all the information in the input data set. Recall our patent description data set from chapter 4. It has almost a couple dozen fields, yet most of our jobs access only a few common ones. It’s inefficient for every job on that data set to read the unused fields every time. One can “refactor” the input data into several smaller data sets. Each has only the fields necessary for a particular type of data processing. The exact refactoring will be application dependent. This technique is similar in spirit to vertical partitioning and column-oriented databases in the relational database management system (RDBMS) world . Tuning for performance 153 Finally, you can reduce the amount of disk and network I/O by compressing your data. You can apply this technique to the intermediate as well as output data sets. Hadoop has many options for data compression, and we devote the next subsection to this topic. 6.3.3 Using compression Even with the use of a combiner, the output of the map phase can be large. This inter- mediate data has to be stored on disk and shuffled across the network. Compressing this intermediate data will improve performance for most MapReduce jobs, and it’s easy too. Hadoop has built-in support for compression and decompression. Enabling compression on the mapper’s output involves setting two configuration properties, as you can see in table 6.3. Table 6.3 Configuration properties to control the compression of mapper ’s output Property Description mapred.compress.map.output Boolean property denoting whether the output of mapper should be compressed mapred.map.output. compression.codec Class property denoting which CompressionCodec to use for compressing mapper’s output To enable compression on the mapper’s output, you set mapred.compress.map. output to true. In addition, you should set mapred.map.output.compression.codec to the appropriate codec class. All codec classes in Hadoop implement the Compression- Codec interface. Hadoop supports a number of compression codecs (see table 6.4). For example, to use GZIP compression, you can set the configuration object: conf.setBoolean("mapred.compress.map.output", true); conf.setClass("mapred.map.output.compression.codec", GzipCodec.class, CompressionCodec.class); You can also use the convenience methods setCompressMapOutput() and set- MapOutputCompressorClass() in JobConf instead of setting the properties directly. Table 6.4 List of codecs available under the org.apache.hadoop.io.compress package Codec Hadoop version Description DefaultCodec 0.18, 0.19, 0.20 Works with files in the zlib format. By Hadoop convention filenames for these files end in .deflate. GzipCodec 0.18, 0.19, 0.20 Works with files in the gzip format. These files have a filename extension of .gz. BZip2Codec 0.19, 0.20 Works with files in the bzip2 format. These files have a filename extension of .bz2. This compression format is unique in that it’s splittable for Hadoop, even when used outside the sequence file format. 154 CHAPTER 6 Programming Practices Data output from the map phase of a job is used only internally to the job, so en- abling compression for this intermediate data is transparent to the developer and is a no-brainer. As many MapReduce applications involve multiple jobs, it makes sense for jobs to be able to output and input in compressed form. It’s highly recommended that data that are passed between Hadoop jobs use the Hadoop-specific sequence file format. Sequence file is a compressable binary file format for storing key/value pairs. It is designed to support compression while remaining splittable . Recall that one of the parallelisms of Hadoop is its ability to split an input file for reading and processing by multiple map tasks. If the input file is in a compressed format, Hadoop will have to be able to split the file such that each split can be decompressed by the map tasks independently. Otherwise parallelism is destroyed if Hadoop has to decompress the file as a whole first. Not all compressed file formats are designed for splitting and decompressing in chunks. Sequence files were specially developed to support this feature. The file format provides sync markers to Hadoop to denote splittable boundaries.7 In addition to its compressability and splittability, sequence files support binary keys and values. Therefore, a sequence file is often used for processing binary documents, such as images, and it works great for text documents and other large key/value objects as well. Each document is considered a record within the sequence file. You can make a MapReduce job output a sequence file by setting its output format to SequenceFileOutputFormat . You’ll want to change its compression type from the default RECORD to BLOCK. With record compression, each record is compressed separately. With block compression, a block of records is compressed together and achieves a higher compression ratio. Finally, you have to call the static methods setCompressOutput() and setOutputCompressorClass() in FileOutputFormat (or SequenceFileOutputFormat , which inherits those methods) to enable output compression using a specific codec. The supported codecs are the same as those given in table 6.4. You add these lines to the driver: conf.setOutputFormat(SequenceFileOutputFormat.class); SequenceFileOutputFormat.setOutputCompressionType(conf, ➥ CompressionType.BLOCK); FileOutputFormat.setCompressOutput(conf, true); FileOutputFormat.setOutputCompressorClass(conf, GzipCodec.class); Table 6.5 lists the equivalent properties for configuring for sequence file output. A Streaming program can output sequence files when given the following options: -outputformat org.apache.hadoop.mapred.SequenceFileOutputFormat -D mapred.output.compression.type=BLOCK -D mapred.output.compress=true -D mapred.output.compression.codec=org.apache.hadoop.io.compress.GzipCode 7 All the input files we’ve seen so far are uncompressed text files where each record is a line. The newline character (\n) can trivially be thought of as the sync marker pointing out to both splittable boundaries and record boundaries. Tuning for performance 155 Table 6.5 Configuration properties for outputting compressed sequence file Property Description mapred.output. compression.type String property to denote the sequence file’s compression type. Can be one of NONE, RECORD, or BLOCK. Default is RECORD but BLOCK almost always compresses better. Convenience method: SequenceFileOutputFormat. ➥ setOutputCompressionType() mapred.output.compress Boolean property on whether to compress the job’s output. Convenience method: FileOutputFormat.setCompressOutput() mapred.output. compression.codec Class property that is used to specify which compression codec to use for compressing the job’s output. Convenience method: FileOutputFormat. ➥ setOutputCompressorClass() To read a sequence file as input, set the input format to SequenceFileInputFormat . Use conf.setInputFormat(SequenceFileInputFormat.class); or -inputformat org.apache.hadoop.mapred.SequenceFileInputFormat for Streaming . There’s no need to configure the compression type or codec class, as the SequenceFile.Reader class (used by SequenceFileRecordReader) will auto- matically determine those settings from the file header. 6.3.4 Reusing the JVM By default, the TaskTracker runs each Mapper and Reducer task in a separate JVM as a child process. This necessarily incurs the JVM start-up cost for each task. If the mapper does its own initialization, such as reading into memory a large data structure (see the example of joining using distributed cache in section 5.2.2), that initialization is part of the start-up cost as well. If each task runs only briefly, or if the mapper initialization takes a long time, then the start-up cost can be a significant portion of a task’s total run time. Starting with version 0.19.0, Hadoop allows the reuse of a JVM across multiple tasks of the same job. The start-up cost can, therefore, be amortized across many tasks. A new property, mapred.job.reuse.jvm.num.tasks , specifies the maximum number of tasks (of the same job) a JVM can run. The default value is 1; JVM is not reused. You can enable JVM reuse by setting the property to a higher number. You can also set it to 156 CHAPTER 6 Programming Practices -1, which means there’s no limit to the number of tasks a JVM can be reused for. The JobConf object has a convenience method, setNumTasksToExecutePerJvm(int) , to set the property for a job. This is summarized in table 6.6. Table 6.6 Configuration property for enabling JVM reuse Property Description mapred.job.reuse.jvm.num.tasks Integer property for setting the maximum number of tasks a JVM can run. A value of -1 means no limit. 6.3.5 Running with speculative execution One of the original design assumptions of MapReduce (as stated in the Google MapReduce paper) is that nodes are unreliable and the framework must handle the situation where some nodes fail in the middle of a job. Under this assumption, the original MapReduce framework specifies the map tasks and the reduce tasks to be idempotent . This means that when a task fails, Hadoop can restart that task and the overall job will end up with the same result. Hadoop can monitor the health of running nodes and restart tasks on failed nodes automatically. This makes fault toler- ance transparent to the developer. Often nodes don’t suddenly fail but experience slowdown as I/O devices go bad. In such situations everything works but the tasks run slower. Sometimes tasks also run slow because of temporary congestion. This doesn’t affect the correctness of the running job but certainly affects its performance. Even one slow-running task will delay the completion of a MapReduce job. Until all mappers have finished, none of the reducers will start running. Similarly, a job is not considered finished until all the reducers have finished. Hadoop uses the idempotency property again to mitigate the slow-task problem. Instead of restarting a task only after it has failed, Hadoop will notice a slow-running task and schedule the same task to be run in another node in parallel. Idempotency guarantees the parallel task will generate the same output. Hadoop will monitor the parallel tasks. As soon as one finishes successfully, Hadoop will use its output and kill the other parallel tasks. This entire process is called speculative execution. Note that speculative execution of map tasks will take place only after all map tasks have been scheduled to run, and only for map tasks that are making much less progress than is average on the other map tasks. It’s the same case for speculative execution of reduce tasks. Speculative execution does not “race” multiple copies of a task to get the best completion time. It only prevents the slow tasks from dragging down the job’s completion time. By default, speculative execution is enabled. One can turn it off for map tasks and reduce tasks separately. To do this, set one or both of the properties in table 6.7 to false. They’re applied on a per-job basis, but you can also change the cluster-wide default by setting them in the cluster configuration file. Tuning for performance 157 Table 6.7 Configuration properties for enabling and disabling speculative execution Property Description mapred.map.tasks.speculative. execution Boolean property denoting whether speculative execution is enabled for map tasks mapred.reduce.tasks.speculative. execution Boolean property denoting whether speculative execution is enabled for reduce tasks You should leave speculative execution on in general. The primary reason to turn it off is if your map tasks or reduce tasks have side effects and are therefore not idempotent. For example, if a task writes to external files, speculative execution can cause multiple copies of a task to collide in attempting to create the same external files. You can turn off speculative execution to ensure that only one copy of a task is being run at a time. NOTE If your tasks have side effects, you should also think through how Hadoop’s recovery mechanism would interact with those side effects. For example, if a task writes to an external file, it’s possible that the task dies right after writing to the external file. In that case, Hadoop will restart the task, which will try to write to that external file again. You need to make sure your tasks’ operation remains correct in such situations. 6.3.6 Refactoring code and rewriting algorithms If you’re willing to rewrite your MapReduce programs to optimize performance, some straightforward techniques and some nontrivial, application-dependent rewritings can speed things up. One straightforward technique for a Streaming program is to rewrite it for Hadoop Java. Streaming is great for quickly creating a MapReduce job for ad hoc data analysis, but it doesn’t run as fast as Java under Hadoop. Streaming jobs that start out as one-off queries but end up being run frequently can gain from a Java re-implementation. If you have several jobs that run on the same input data, there are probably opportunities to rewrite them into fewer jobs. For example, if you’re computing the maximum as well as the minimum of a data set , you can write a single MapReduce job that computes both rather than compute them separately using two different jobs. This may sound obvious, but in practice many jobs are originally written to do one function well. This is a good design practice. A job’s conciseness makes it widely applicable to different data sets for different purposes. Only after some usage should you start looking for job groupings that you can rewrite to be faster. One of the most important things you can do to speed up a MapReduce program is to think hard about the underlying algorithm and see if a more efficient algorithm can compute the same results faster. This is true for any programming, but it is more significant for MapReduce programs. Standard text books on algorithm and data 158 CHAPTER 6 Programming Practices structure (sorting, lists, maps, etc.) comprehensively cover design choices for most traditional programming. Hadoop programs, on the other hand, tend to touch on “exotic” areas, such as distributed computing, functional programming, statistics, and data-intensive processing, where best practices are less known to most programmers and there is still exciting research today to explore new approaches. One example we’ve already seen that leverages a new data structure to speed up MapReduce programs is the use of Bloom filters in semijoins (section 5.3). The Bloom filter is well-known in the distributed computing community but relatively unknown outside of it. Another classic example of using a new algorithm to speed up a MapReduce program comes from statistics in the calculation of variance . Non-statisticians may compute variance using its canonical definition: (1/N) * Sumi[(Xi – Xavg)2] where Sumi denotes summation over the data set. The variable Xavg is the average of the data set. If we don’t know that average ahead of time, then a non-statistician may decide to run one MapReduce job to find the average first, and a second MapReduce job to compute the variance. Someone more familiar with computing statistics will use an equivalent definition: (1/N) * Sumi[(Xi)2] – ((1/N) * Sumi[Xi])2 From this definition one needs the sum of X as well as the sum of X2, but you can compute both sums together in one scan of the data, using only a single MapReduce job. (This is analogous to the example of calculating maximum and minimum in a single job.) A little statistical background has halved the processing time in comput- ing variance.8 You should also pay attention to the computational complexity of your algorithms . Hadoop provides “only” linear scalability , and you can still bring it to its knees with large data sets running under computationally intensive algorithms that are quadratic or worse. You certainly should look for more efficient algorithms in those cases, and sometimes you may have to settle for faster algorithms that only give approximate results. 6.4 Summary Development methodologies for Hadoop build on top of best practices for Java pro- gramming, such as unit testing and test-driven development. Hadoop’s central role of processing data calls for more data-centric testing processes. Math and logic errors are more prevalent in data-intensive programs and they’re often inconspicuous. The 8 There’s a lot of nuisance in numerical computation over large data. In this variance calculation example we note our refactored MapReduce job has lower numerical precision and is more likely to run into overflow problems. Summary 159 distributed nature of Hadoop makes debugging much harder. To lessen the burden, you should test in stages, from a nondistributed (i.e., local) mode to a single-node pseudo-distributed mode, and finally to a fully distributed mode. The famous computer scientist Donald Knuth once said that “premature optimization is the root of all evil.” You should tune your Hadoop program for performance only after it’s been fully debugged. Beyond thinking through general algorithmic and computational issues, performance enhancement is platform-specific, and Hadoop has a number of specific techniques to make jobs run more efficiently. 7 160 Cookbook This chapter covers ■ Passing custom parameters to tasks ■ Retrieving task-specific information ■ Creating multiple outputs ■ Making output globally sorted Interfacing with relational databases ■ This book so far has covered the core techniques for making a MapReduce pro- gram. Hadoop is a big framework that supports many more functionalities than those core techniques. In this age of Bing and Google, you can look up specialized MapReduce techniques rather easily, and we don’t try to be an encyclopedic refer- ence. In our own usage and from our discussion with other Hadoop users, we’ve found a number of techniques generally useful, techniques such as being able to take a standard relational database as input or output to a MapReduce job. We’ve collected some of our favorite “recipes” in this cookbook chapter. 7.1 Passing job-specific parameters to your tasks In writing your Mapper and Reducer , you often want to make certain aspects con- figurable. For example, our joining program in chapter 5 is hardcoded to take the Passing job-specific parameters to your tasks 161 first data column as the join key. The program can be more generally applicable if the column for the join key can be specified by the user at run time. Hadoop itself uses a configuration object to store all the configuration properties for a job. You can use the same object to pass parameters to your Mapper and Reducer. We’ve seen how the MapReduce driver configures the JobConf object with properties, such as input format, output format, mapper class, and so forth. To introduce your own property, you give your property a unique name and set it with a value in the same configuration object. This configuration object is passed to all TaskTrackers , so the properties in the configuration object are available to all tasks in that job. Your Mapper and Reducer can read the configuration object and retrieve the property value. The Configuration class (parent of JobConf) has a number of generic setter methods. Properties are key/value pairs, where key has to be a String, but value can be one of a number of common types. Signature for the common setter methods are public void set(String name, String value) public void setBoolean(String name, boolean value) public void setInt(String name, int value) public void setLong(String name, long value) public void setStrings(String name, String... values) Note that Hadoop stores all properties internally as strings. All the other setter methods are convenience methods for set(String, String). For example, the setStrings(String, String...) method takes a String array , turns it into a single comma-separated String, and sets that String as the property value. The get- Strings() retrieval method similarly splits the concatenated string back into an array. With that in mind, don’t keep any commas in the strings in the original array. If you want commas, you should use your own string-encoding function. Your driver will first set the properties in the configuration object to make them available to all tasks. Your Mapper and Reducer have access to the configuration object in the configure() method. When a task initializes, it calls configure(), which you override to retrieve and store your properties . Your map() and reduce() methods will access your copy of those properties later. In the following example we call our new property myjob.myproperty, and it takes an integer value specified by the user. public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, MyJob.class); ... job.setInt("myjob.myproperty", Integer.parseInt(args[2])); JobClient.runJob(job); return 0; q } In MapClass , the configure() method retrieves the property value and stores it in the object’s scope. The getter methods of the Configuration class require specifying Set custom property 162 CHAPTER 7 Cookbook default values, which will be returned if the requested property is not set in the con- figuration object. For this example we use a default of 0: public static class MapClass extends MapReduceBase implements Mapper { int myproperty; public void configure(JobConf job) { myproperty = job.getInt("myjob.myproperty", 0); } ... q } If you want to use the property in the Reducer, it will also have to retrieve the property. public static class Reduce extends MapReduceBase implements Reducer { int myproperty; public void configure(JobConf job) { myproperty = job.getInt("myjob.myproperty", 0); } ... } The Configuration class has a larger list of getter methods than setter methods, although they are largely self-explanatory. Almost all the getter methods require a default value as argument. The exception is get(String) , which returns null if the property with the specified name is not set. public String get(String name) public String get(String name, String defaultValue) public boolean getBoolean(String name, boolean defaultValue) public float getFloat(String name, float defaultValue) public int getInt(String name, int defaultValue) public long getLong(String name, long defaultValue) public String[] getStrings(String name, String... defaultValue) Given that our job class implements the Tool interface and uses ToolRunner , we can also let the user set custom properties directly using the generic options syntax, in the same way the user would set Hadoop configuration properties. bin/hadoop jar MyJob.jar MyJob -D myjob.myproperty=1 input output We can remove the line in the driver that requires the user to always specify the value of this property as an argument. This is more convenient for the user when the default value would work most of the time. public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, MyJob.class); Get custom property Probing for task -specific information 163 ... int myproperty = job.getInt("myjob.myproperty", 0); if (myproperty < 0) { System.err.println("Invalid myjob.myproperty: " + myproperty); System.exit(0); } JobClient.runJob(job); return 0; } When you allow the user to specify custom properties, it’s good practice for the driver to validate any user input. The example above ensures that the user will not be allowed to specify a negative value for myjob.myproperty. 7.2 Probing for task -specific information In addition to retrieving custom properties and global configuration, we can also use the getter methods on the configuration object to obtain certain state information about the current task and job. For example, in the Mapper you can grab the map. input.file property to get the file path to the current map task. This is exactly what the configure() method in the datajoin package’s DataJoinMapperBase does to infer a tag for the data source. this.inputFile = job.get("map.input.file"); this.inputTag = generateInputTag(this.inputFile); Table 7.1 lists some of the other task-specific state information. Table 7.1 Task-specific state information one can get in the configuration object Property Type Description mapred.job.id String The job ID mapred.jar String The jar location in job directory job.local.dir String The job’s local scratch space mapred.tip.id String The task ID mapred.task.id String The task attempt ID mapred.task.is.map boolean Flag denoting whether this is a map task mapred.task.partition int The ID of the task within the job map.input.file String The file path that the mapper is reading from map.input.start long The offset into the file of the start of the current mapper’s input split map.input.length long The number of bytes in the current mapper’s input split mapred.work.output.dir String The task’s working (i.e., temporary) output directory 164 CHAPTER 7 Cookbook Configuration properties are also available to Streaming programs through environment variables. Before executing a script, the Streaming API will have added all configura- tion properties to the running environment. The property names are reformatted such that non-alphanumeric characters are replaced with an underscore (_). For example, a Streaming script should look at the environment variable map_input_file for the full file path that the current mapper is reading from. import os filename = os.environ["map_input_file"] localdir = os.environ["job_local_dir"] The preceding code shows how one would access configuration properties in Python. 7.3 Partitioning into multiple output files Up ‘till now all the MapReduce jobs we’ve seen output a single set of files. However, there are often cases where it’s more convenient to output multiple sets of files, or split a data set into multiple data sets . A popular example is the partitioning of a large log file into distinct sets of log files for each day. MultipleOutputFormat provides a simple way of grouping similar records into separate data sets. Before writing each output record, this OutputFormat class calls an internal method to determine the filename to write to. More specifically, you will extend a particular subclass of MultipleOutputFormat and implement the generateFileNameForKeyValue() method. The subclass you extend will determine the output format. For example, MultipleTextOutputFormat will output text files whereas MultipleSequenceFileOutputFormat will output sequence files. In either case, you’ll override the following method to return the filename for each output key/ value pair: protected String generateFileNameForKeyValue(K key, V value, String name) The default implementation returns the argument name, which is the leaf filename. You can make the method return a filename that’s dependent on the content of the record. For our example here, we take the patent metadata and partition it by country. All patents from U.S. inventors will go into one set of files, all patents from Japan into another pile, and so forth. The skeleton of this example program is a map-only job that takes its input and immediately outputs it. The main change we’ve made is to create our own subclass of MultipleTextOutputFormat called PartitionbyCountryMTOF. (Note that MTOF is an acronym for MultipleTextOutputFormat .) Our subclass will store each record to a location based on the inventing country listed in that record. As we treat the value returned by generateFileNameForKeyValue() as a file path, we’re able to create a subdirectory for each country by returning country + "/" + filename. See listing 7.1. Partitioning into multiple output files 165 Listing 7.1 Partition patent metadata into multiple directories based on country public class MultiFile extends Configured implements Tool { public static class MapClass extends MapReduceBase implements Mapper { public void map(LongWritable key, Text value, OutputCollector output, Reporter reporter) throws IOException { output.collect(NullWritable.get(), value); } } public static class PartitionByCountryMTOF extends MultipleTextOutputFormat { protected String generateFileNameForKeyValue(NullWritable key, Text value, String filename) { String[] arr = value.toString().split(",", -1); String country = arr[4].substring(1,3); return country + "/" + filename; } } public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, MultiFile.class); Path in = new Path(args[0]); Path out = new Path(args[1]); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); job.setJobName("MultiFile"); job.setMapperClass(MapClass.class); job.setInputFormat(TextInputFormat.class); job.setOutputFormat(PartitionByCountryMTOF.class); job.setOutputKeyClass(NullWritable.class); job.setOutputValueClass(Text.class); job.setNumReduceTasks(0); JobClient.runJob(job); return 0; } public static void main(String[] args) throws Exception { int res = ToolRunner.run(new Configuration(), new MultiFile(), args); System.exit(res); } } 166 CHAPTER 7 Cookbook After executing the preceding program, we can see that the output directory now has a separate directory for each country. ls output/ AD BN CS GE IN LC MT PH SV VE AE BO CU GF IQ LI MU PK SY VG AG BR CY GH IR LK MW PL SZ VN AI BS CZ GL IS LR MX PT TC VU AL BY DE GN IT LT MY PY TD YE AM BZ DK GP JM LU NC RO TH YU AN CA DO GR JO LV NF RU TN ZA AR CC DZ GT JP LY NG SA TR ZM AT CD EC GY KE MA NI SD TT ZW AU CH EE HK KG MC NL SE TW AW CI EG HN KN MG NO SG TZ AZ CK ES HR KP MH NZ SI UA BB CL ET HT KR ML OM SK UG BE CM FI HU KW MM PA SM US BG CN FO ID KY MO PE SN UY BH CO FR IE KZ MQ PF SR UZ BM CR GB IL LB MR PG SU VC And within the directory for each country are files with only records (patents) created by those countries. ls output/AD part-00003 part-00005 part-00006 head output/AD/part-00006 5765303,1998,14046,1996,"AD","",,1,12,42,5,59,11,1,0.4545,0,0,1,67.3636,,,, 5785566,1998,14088,1996,"AD","",,1,9,441,6,69,3,0,1,,0.6667,,4.3333,,,, 5894770,1999,14354,1997,"AD","",,1,,82,5,51,4,0,1,,0.625,,7.5,,,, We’ve written this simple partitioning exercise as a map-only program. You can apply the same technique to the output of reducers as well. Be careful not to confuse this with the partitioner in the MapReduce framework. That partitioner looks at the keys of intermediate records and decides which reducer will process them. The partitioning we’re doing here looks at the key/value pair of the output and decides which file to store to. MultipleOutputFormat is simple, but it’s also limited. For example, we were able to split the input data by row, but what if we want to split by column? Let’s say we want to create two data sets from the patent metadata: one containing time-related information (e.g., publication date) for each patent and another one containing geographical information (e.g., country of invention). These two data sets may be of different output formats and different data types for the keys and values. We can look to MultipleOutputs, introduced in version 0.19 of Hadoop, for more powerful capabilities. The approach taken by MultipleOutputs is different from MultipleOutputFormat. Rather than asking for the filename to output each record, MultipleOutputs creates multiple OutputCollectors . Each OutputCollector can have its own OutputFormat and types for the key/value pair. Your MapReduce program will decide what to output to each OutputCollector. Listing 7.2 shows a program that takes our patent metadata Partitioning into multiple output files 167 and outputs two data sets. One has chronological information, such as issued date. The other data set has geographical information associated with each patent. This, too, is a map-only program, but you can apply the multiple output collectors to reducers in a straightforward way. Listing 7.2 Program to project different columns of input data to different files public class MultiFile extends Configured implements Tool { public static class MapClass extends MapReduceBase implements Mapper { private MultipleOutputs mos; private OutputCollector collector; public void configure(JobConf conf) { mos = new MultipleOutputs(conf); } public void map(LongWritable key, Text value, OutputCollector output, Reporter reporter) throws IOException { String[] arr = value.toString().split(",", -1); String chrono = arr[0] + "," + arr[1] + "," + arr[2]; String geo = arr[0] + "," + arr[4] + "," + arr[5]; collector = mos.getCollector("chrono", reporter); collector.collect(NullWritable.get(), new Text(chrono)); collector = mos.getCollector("geo", reporter); collector.collect(NullWritable.get(), new Text(geo)); } public void close() throws IOException { mos.close(); } } public int run(String[] args) throws Exception { Configuration conf = getConf(); JobConf job = new JobConf(conf, MultiFile.class); Path in = new Path(args[0]); Path out = new Path(args[1]); FileInputFormat.setInputPaths(job, in); FileOutputFormat.setOutputPath(job, out); job.setJobName("MultiFile"); job.setMapperClass(MapClass.class); job.setInputFormat(TextInputFormat.class); job.setOutputKeyClass(NullWritable.class); job.setOutputValueClass(Text.class); job.setNumReduceTasks(0); MultipleOutputs.addNamedOutput(job, "chrono", TextOutputFormat.class, NullWritable.class, Text.class); 168 CHAPTER 7 Cookbook MultipleOutputs.addNamedOutput(job, "geo", TextOutputFormat.class, NullWritable.class, Text.class); JobClient.runJob(job); return 0; } public static void main(String[] args) throws Exception { int res = ToolRunner.run(new Configuration(), new MultiFile(), args); System.exit(res); } } To use MultipleOutputs, the driver of the MapReduce program must set up the output collectors it expects to use. Creating the collectors involves a call to MultipleOutputs’ static method addNamedOutput(). We’ve created one output collector called chrono and another one called geo. We’ve created them both to use TextOutputFormat and have the same key/value types, but we can choose to use different output formats or data types. After setting up the output collectors in the driver, we need to get the MultipleOutputs object that tracks them when the mapper is initialized in the configure() method. This object must be available throughout the duration of the map task. In the map() function itself, we call the getCollector() method on the MultipleOutputs object to get back the chrono and the geo OutputCollectors. We will write out different data that’s appropriate for each output collector. We have given a name to each output collector in MultipleOutputs , and MultipleOutputs will automatically generate the output filenames. We can look at the files outputted by our script to see how MultipleOutputs generates the output names: ls -l output/ total 101896 -rwxrwxrwx 1 Administrator None 9672703 Jul 31 06:28 chrono-m-00000 -rwxrwxrwx 1 Administrator None 7752888 Jul 31 06:29 chrono-m-00001 -rwxrwxrwx 1 Administrator None 6884496 Jul 31 06:29 chrono-m-00002 -rwxrwxrwx 1 Administrator None 6933561 Jul 31 06:29 chrono-m-00003 -rwxrwxrwx 1 Administrator None 7164558 Jul 31 06:29 chrono-m-00004 -rwxrwxrwx 1 Administrator None 7273561 Jul 31 06:29 chrono-m-00005 -rwxrwxrwx 1 Administrator None 8281663 Jul 31 06:29 chrono-m-00006 -rwxrwxrwx 1 Administrator None 9428951 Jul 31 06:28 geo-m-00000 -rwxrwxrwx 1 Administrator None 7464690 Jul 31 06:29 geo-m-00001 -rwxrwxrwx 1 Administrator None 6580482 Jul 31 06:29 geo-m-00002 -rwxrwxrwx 1 Administrator None 6448648 Jul 31 06:29 geo-m-00003 -rwxrwxrwx 1 Administrator None 6432392 Jul 31 06:29 geo-m-00004 -rwxrwxrwx 1 Administrator None 6546828 Jul 31 06:29 geo-m-00005 -rwxrwxrwx 1 Administrator None 7450768 Jul 31 06:29 geo-m-00006 Inputting from and outputting to a database 169 -rwxrwxrwx 1 Administrator None 0 Jul 31 06:28 part-00000 -rwxrwxrwx 1 Administrator None 0 Jul 31 06:28 part-00001 -rwxrwxrwx 1 Administrator None 0 Jul 31 06:29 part-00002 -rwxrwxrwx 1 Administrator None 0 Jul 31 06:29 part-00003 -rwxrwxrwx 1 Administrator None 0 Jul 31 06:29 part-00004 -rwxrwxrwx 1 Administrator None 0 Jul 31 06:29 part-00005 -rwxrwxrwx 1 Administrator None 0 Jul 31 06:29 part-00006 We have a set of files prefixed with chrono and another set of files prefixed with geo. Note that the program created the default output files part-* even though it wrote nothing explicitly. It’s entirely possible to write to these files using the original OutputCollector passed in through the map() method. In fact, if this was not a map-only program, records written to the original OutputCollector, and only those records, would be passed to the reducers for processing. One of the trade-offs with MultipleOutputs is that it has a rigid naming structure compared to MultipleOutputFormat. Your output collector’s name cannot be part, because that’s already in use for the default. The output filename is also strictly defined as the output collector’s name followed by m or r depending on whether the output was collected at the mapper or the reducer. It’s finally followed by a partition number. head output/chrono-m-00000 "PATENT","GYEAR","GDATE" 3070801,1963,1096 3070802,1963,1096 3070803,1963,1096 3070804,1963,1096 3070805,1963,1096 3070806,1963,1096 3070807,1963,1096 3070808,1963,1096 3070809,1963,1096 head output/geo-m-00000 "PATENT","COUNTRY","POSTATE" 3070801,"BE","" 3070802,"US","TX" 3070803,"US","IL" 3070804,"US","OH" 3070805,"US","CA" 3070806,"US","PA" 3070807,"US","OH" 3070808,"US","IA" 3070809,"US","AZ" Looking at the output files, we see that we’ve successfully projected out the columns on the patent data set into distinct files. 7.4 Inputting from and outputting to a database Although Hadoop is useful for processing large data, relational databases remain the workhorse of many data processing applications. Oftentimes Hadoop will need to interface with databases. 170 CHAPTER 7 Cookbook Although it’s possible to set up a MapReduce program to take its input by directly querying a database rather than reading a file in HDFS , the performance is less than ideal. More often you would copy a data set from a database to HDFS. You can easily do it with a standard database dump utility to get a flat file. You then upload to HDFS using its file put shell command. But sometimes it is sensible having a MapReduce program write directly to a database. Many MapReduce programs take large data sets and process them into a manageable size for databases to handle. For example, we often use MapReduce in the ETL-like process of taking humongous log files and computing a much smaller and more manageable set of statistics for analysts to look at. The DBOutputFormat is the crucial class for accessing databases. In your driver you set the output format to this class. You’ll need to specify the configuration for connecting to your database. You can do this through the static configureDB() method in DBConfiguration : public static void configureDB(JobConf job, String driverClass, ➥ String dbUrl, String userName, String passwd) After that, you’ll specify what table you’re writing to and what fields are there. This is done with the static setOutput() method in DBOutputFormat. public static void setOutput(JobConf job, String tableName, ➥ String... fieldNames) Your driver should have a few lines that look like this: conf.setOutputFormat(DBOutputFormat.class); DBConfiguration.configureDB(job, "com.mysql.jdbc.Driver", "jdbc:mysql://db.host.com/mydb", "username", "password") DBOutputFormat.setOutput(job, "Events", "event_id", "time"); Using DBOutputFormat forces your output key to implement the DBWritable in- terface. Only the key is written to the database. As usual, the keys have to imple- ment Writable. The signatures for Writable and DBWritable are similar; only the argument types are different. The write() method in Writable takes a DataOutput, whereas write() in DBWritable takes a PreparedStatement. Similarly the read- Fields() method for Writable takes a DataInput, whereas readFields() for DBWritable takes a ResultSet. Unless you plan on fetching input data straight from the database using DBInputFormat, readFields() in DBWritable will never be called. public class EventsDBWritable implements Writable, DBWritable { private int id; private long timestamp; public void write(DataOutput out) throws IOException { out.writeInt(id); out.writeLong(timestamp); Keeping all output in sorted order 171 } public void readFields(DataInput in) throws IOException { id = in.readInt(); timestamp = in.readLong(); } public void write(PreparedStatement statement) throws SQLException { statement.setInt(1, id); statement.setLong(2, timestamp); } public void readFields(ResultSet resultSet) throws SQLException { id = resultSet.getInt(1); timestamp = resultSet.getLong(2); } } We want to emphasize again that reading and writing to databases from within Hadoop is only appropriate for data sets that are relatively small by Hadoop standards. Unless your database setup is as parallel as Hadoop (which can be the case if your Hadoop cluster is relatively small while you have many shards in your database system), your DB will be the performance bottleneck, and you may not gain any scalability advan- tage from your Hadoop cluster. Oftentimes, it’s better to bulk load data into a data- base rather than make direct writes from Hadoop. You’ll need custom solutions for extremely large-scale databases.1 7.5 Keeping all output in sorted order The MapReduce framework guarantees the input to each reducer to be in sorted order based on key. In many cases, the reducer only does a simple computation on the value part of a key/value pair. The output also stays in sorted order. Keep in mind that the MapReduce framework does not guarantee the sorted order of the reducer output. Rather, it’s a byproduct of the sorted input and the typical type of operations reducers perform. For some applications, the sorted order is unnecessary, and sometimes questions are raised about turning off the sorting operation to eliminate an unnecessary step in the reducer. The truth is that the sorting operation is not so much about enforcing the sorted order of the reducer’s input. Rather, sorting is an efficient way to group all records of the same key together. If the grouping function is unnecessary, then we can directly generate an output record from a single input record. In that case, you’ll be able to improve performance by eliminating the entire reduce phase. You can do this by setting the number of reducers to 0, making the application a map- only job. 1 LinkedIn has an interesting blog post on challenges faced in moving massive amounts of data resulting from offline processes (i.e., Hadoop) into live systems: http://project-voldemort.com/blog/2009/06/building-a- 1-tb-data-cycle-at-linkedin-with-hadoop-and-project-voldemort/. 172 CHAPTER 7 Cookbook On the other hand, for some applications it’s desirable that all output is sorted in total . Each output file (generated by one reducer) is already in sorted order; it would be nice to also have all the records in part-00000 be smaller than records in part- 00001, and part-00001 be smaller than part-00002, and so forth. The key to doing this is the partitioner operation in the framework. The job of the partitioner is to deterministically assign a reducer to each key. All records of the same key are grouped and processed together in the reduce stage. An important design requirement of the partitioner is to balance load across reducers; no one reducer is given many more keys than other reducers. Without any prior information about the distribution of keys, the default partitioner uses a hashing function to uniformly assign keys to reducers. This often works well in distributing work evenly across reducers, but the assignment is intentionally arbitrary and not in any order. If we have prior knowledge that the keys are approximately uniformly distributed, we can use a partitioner that assigns key ranges to each reducer and still be certain that the reducers’ loads are fairly balanced. TIP The hash partitioner can also fail to evenly distribute work if certain keys take much more time to process than others. For example, in highly skewed data sets, a significant number of records may have the same key. If possible, you should use a combiner to lessen the load at the reduce phase by doing as much preprocessing as possible at the map phase. In addition, you can also choose to write a special partitioner to distribute keys unevenly in such a way that it balances out the inherent skew of the data and its processing. The TotalOrderPartitioner is a partitioner that ensures sortedness between output partitions, not only within. Sorting of large-scale data (i.e., the TeraSort benchmark) originally used a similar version of this class. This class takes a sequence file with a sorted partition keyset and proceeds to partition keys in different ranges to the reducers. 7.6 Summary This chapter discussed many tools and techniques to make your Hadoop job more user-friendly or make it interface better with other components of your data process- ing infrastructure. The full extent of the capabilities available in a Hadoop job is doc- umented in the Hadoop API: http://hadoop.apache.org/common/docs/current/ api/index.html. You may also want to check out additional abstractions such as Pig and Hive to simplify your programming. We’ll cover these tools in chapters 10 and 11. If your role involves administrating a Hadoop cluster, you will find the tips on managing a Hadoop cluster in the next chapter useful. 8 173 Managing Hadoop This chapter covers ■ Configuring for a production system ■ Setting up a job scheduler Maintaining the HDFS filesystem ■ The installation instructions in chapter 2 produced a running Hadoop cluster fairly quickly. The configuration was relatively simple, but unfortunately it’s not good for a production cluster, which will be under heavy sustained use. There are various configuration parameters that you would want to tune for a production cluster, and section 8.1 will cover those parameters. In addition, like any system, a Hadoop cluster will change over time and you (or some administrator) will have to know how to maintain it to keep it running in good shape. This is particularly true for the HDFS filesystem. In sections 8.2 through 8.5, we cover various standard filesystem maintenance tasks, such as checking its health, setting permissions, quotas, and recovering deleted files (trash). Sections 8.6 through 8.10 will cover the bigger but rarer administrative tasks more specific 174 CHAPTER 8 Managing Hadoop to HDFS. These include adding/removing nodes (capacity) and recovery from NameNode failure. We end the chapter with a section on setting up a scheduler to manage multiple running jobs. 8.1 Setting up parameter values for practical use Hadoop has many different parameters. Their default values tend to target running in standalone mode. They also tend to veer toward being idiotproof. The default values are more likely to work on more systems without causing any errors. However, often- times they’re far from optimal in a production cluster. Table 8.1 shows some of the system properties that you’ll want to change for a production cluster. Table 8.1 Hadoop properties that you can tune for a production cluster Property Description Suggested value dfs.name.dir Directory in NameNode’s local filesystem to store HDFS’s metadata /home/hadoop/ dfs/name dfs.data.dir Directory in a DataNode’s local filesystem to store HDFS’s file blocks /home/hadoop/ dfs/data mapred.system.dir Directory in HDFS for storing shared MapReduce system files /hadoop/ mapred/system mapred.local.dir Directory in a TaskNode’s local filesystem to store temporary data mapred. tasktracker. {map|reduce} .tasks.maximum Maximum number of map and reduce tasks that can run simultaneously in a TaskTracker hadoop.tmp.dir Temporary Hadoop directories /home/hadoop/ tmp dfs.datanode.du .reserved Minimum amount of free space a DataNode should have 1073741824 mapred.child. java.opts Heap size allocated to each child task -Xmx512m mapred.reduce. tasks Number of reduce tasks for a job The default values for dfs.name.dir and dfs.data.dir point to directories under /tmp, which is intended only for temporary storage in almost all Unix systems. You will definitely want to change those properties for a production cluster. 1 In addition, these properties can take comma-separated lists of directories. In the case of dfs.name. dir, multiple directories are good for backup purposes. If a DataNode has multiple drives, you should have a data directory in each one and list them all in dfs.data.dir. 1 The rationale for using /tmp illustrates how default values are idiotproof. Every Unix system has the /tmp directory so you won’t get a “directory not found” error. Setting up parameter values for practical use 175 The DataNode will use them all in parallel to speed up I/O.2 You should also specify directories in multiple drives for mapred.local.dir to speed up processing of tem- porary data. The default configuration for Hadoop’s temporary directories, hadoop.tmp.dir, is dependent on the user name. You should avoid having any Hadoop property that depends on a user name, as there can be mismatches between the user name used to submit a job and the user name used to start a Hadoop node. You should set it to something like /home/hadoop/tmp to be independent of any user name. Another problem with the default value of hadoop.tmp.dir is that it points to the /tmp directory. Although that’s an appropriate place for temporary storage, most default Linux configurations have a quota on /tmp that is too small for Hadoop. Rather than increase the quota for /tmp, it’s better to point hadoop.tmp.dir to a directory that’s known to have a lot of space. By default, HDFS doesn’t require DataNodes to have any reserved free space. In practice, most systems have questionable stability when the amount of free space gets too low. You should set dfs.datanode.du.reserved to reserve 1 GB of free space in a DataNode. A DataNode will stop accepting block writes when its amount of free space falls below the reserved amount. Each TaskTracker is allowed to run a configurable maximum number of map and reduce tasks. Hadoop’s default is four tasks (two map tasks and two reduce tasks). The right number depends on many factors, although most setups call for one to two tasks per core. You can set a quad core machine to have a maximum of six map and reduce tasks (three each), because there will already be one task each allocated for TaskTracker and DataNode, to make a total of eight. Similarly, you can set up a dual quad core machine to have a maximum of fourteen map and reduce tasks. This predicates on most MapReduce jobs being I/O heavy. You should reduce the maximum number of tasks allowed if you expect more CPU-intensive loads . In considering the number of tasks allowed, you should also consider the amount of heap memory allocated to each task. Hadoop’s default of 200 MB per task is quite underwhelming. Many setups bump up the default to 512 MB, some even at 1 GB. This is not a final property. Each job can request more (or less) heap space per task. Be sure that you have sufficient usable memory in your machines for your configuration parameters. Keep in mind that DataNode and TaskTracker each already uses 1 GB of RAM. Although you can set the number of reduce tasks per each individual MapReduce job, it’s desirable to have a default that works well most of the time. Hadoop’s 2 There’s been some discussion in the Hadoop forums about whether one should configure multiple hard drives in a DataNode as RAID or JBOD. Hadoop doesn’t need RAID’s data redundancy because HDFS already replicates data across machines. Furthermore, Yahoo has stated that they were able to get noticeable performance improvement using JBOD. The stated reason is that hard drives, even of the same model, have high variance in their speed. A RAID configuration would slow down the I/O to the slowest drive. On the other hand, letting each drive function independently will allow each one to operate at its top speed, making the overall throughput of the system higher. 176 CHAPTER 8 Managing Hadoop default of one reduce task per job is certainly suboptimal in most cases. The general recommendation is to set the default to either 0.95 or 1.75 times the maximum number of reduce TaskTrackers in the cluster. This means that the number of reduce tasks in a job should be 0.95 or 1.75 multiplied by number of worker nodes multiplied by mapred.tasktracker.reduce.tasks.maximum. A factor of 0.95 will have all the reduce tasks launched immediately and start copying map tasks’ output as they finish. At a factor of 1.75, some reduce tasks will be able to launch immediately whereas others will wait. The faster nodes will finish the first round of reduce tasks earlier and start on the second round. The slowest nodes won’t need to process any reduce tasks from the second round. This can result in better load balancing. 8.2 Checking system’s health Hadoop provides a filesystem checking utility called fsck . You call it with a file path and it’ll recursively check the health of all the files under that path. Call it with the argument / and it’ll check the entire filesystem. An example output looks like this: bin/hadoop fsck / Status: HEALTHY Total size: 143106109768 B Total dirs: 9726 Total files: 41532 Total blocks (validated): 42419 (avg. block size 3373632 B) Minimally replicated blocks: 42419 (100.0 %) Over-replicated blocks: 0 (0.0 %) Under-replicated blocks: 0 (0.0 %) Mis-replicated blocks: 0 (0.0 %) Default replication factor: 3 Average block replication: 3.0 Corrupt blocks: 0 Missing replicas: 0 (0.0 %) Number of data-nodes: 8 Number of racks: 1 Most of the information should be self-explanatory. By default fsck will ignore files still open for writing by a client. You can get a list of such files by running fsck with the -openforwrite argument. As fsck checks the filesystem, it will print out a dot for each file it found healthy (not shown in the above output). It’ll print out a message for each file that is less than healthy, including ones that have over-replicated blocks , under-replicated blocks , mis- replicated blocks , corrupt blocks , and missing replicas . Over-replicated blocks, under- replicated blocks, and mis-replicated blocks are not too alarming as HDFS is self-healing. But, corrupt blocks and missing replicas mean that data has been permanently lost. By default fsck doesn’t act on those corrupt files, but you can run fsck with the -delete option to remove them. Better yet is to run fsck with the -move option, which moves corrupted files into the /lost+found directory for salvaging. You can tell fsck to print out more information by adding -files, -blocks, -locations, and -racks options to fsck. Each successive option requires the Checking system’s health 177 preceding option be used as well. The -blocks option requires the -files option be used. The -locations option requires both -files and -blocks options be used, and so forth. The -files option tells fsck to print out, for each file it checks, a line of information containing the file’s path, the file’s size in bytes and blocks, and the file’s status. The -blocks option tells fsck to go further and print out a line of information for each block in the file. This line will include the block’s name, its length, and its number of replicas. The -locations option will include in each line the location of the block’s replicas. The -racks option will add the rack name to the location information. For example, a short one-block file will have its report as bin/hadoop fsck /user/hadoop/test -files -blocks -locations -racks /user/hadoop/test/part-00000 35792 bytes, 1 block(s): OK 0. blk_-4630072455652803568_97605 len=35792 repl=3 ➥ [/default-rack/, /default-rack/, ➥ /default-rack/] Status: HEALTHY Total size: 35792 B Total dirs: 0 Total files: 1 Total blocks (validated): 1 (avg. block size 35792 B) Minimally replicated blocks: 1 (100.0 %) Over-replicated blocks: 0 (0.0 %) Under-replicated blocks: 0 (0.0 %) Mis-replicated blocks: 0 (0.0 %) Default replication factor: 3 Average block replication: 3.0 Corrupt blocks: 0 Missing replicas: 0 (0.0 %) Number of data-nodes: 8 Number of racks: 1 While fsck reports on each file in HDFS, there is a dfsadmin command for report- ing on each DataNode. You can get it through the -report option on the dfsadmin command: bin/hadoop dfsadmin -report Total raw bytes: 535472824320 (498.7 GB) Remaining raw bytes: 33927731366 (31.6 GB) Used raw bytes: 379948188541 (353.85 GB) % used: 70.96% Total effective bytes: 0 (0 KB) Effective replication multiplier: Infinity ------------------------------------------------- Datanodes available: 8 Name: State : In Service Total raw bytes: 76669841408 (71.4 GB) Remaining raw bytes: 2184594843(2.03 GB) Used raw bytes: 56598956650 (52.71 GB) % used: 73.82% Last contact: Sun Jun 21 16:13:32 PDT 2009 178 CHAPTER 8 Managing Hadoop Name: State : In Service Total raw bytes: 76669841408 (71.4 GB) Remaining raw bytes: 6356175381(5.92 GB) Used raw bytes: 54220537856 (50.5 GB) % used: 70.72% Last contact: Sun Jun 21 16:13:33 PDT 2009 Name: State : In Service Total raw bytes: 76669841408 (71.4 GB) Remaining raw bytes: 6106387206(5.69 GB) Used raw bytes: 52412190091 (48.81 GB) % used: 68.36% Last contact: Sun Jun 21 16:13:33 PDT 2009 ... To look at the NameNode ’s current activity, you can use the -metasave option in dfsadmin: bin/hadoop dfsadmin -metasave filename This will save some of NameNode’s metadata into its log directory under filename. In this metadata, you’ll find lists of blocks waiting for replication, blocks being repli- cated, and blocks awaiting deletion. For replication each block will also have a list of DataNodes being replicated to. Finally, the metasave file will also have summary statistics on each DataNode. 8.3 Setting permissions HDFS has a basic file permission system similar to the POSIX model. Each file has nine permission settings: the read (r), write (w), and execute (x) permissions for each of the file’s associated owner, group, and other users. Not all permission set- tings are meaningful. Under HDFS, we can’t execute files; so we can’t set the x permission. Permission settings for directories also closely follow the POSIX model. The r permission allows listing of the directory. The w permission allows creation or deletion of files or directories. The x permission allows one to access children of the directory. Current HDFS releases don’t provide much in terms of security. You should use the HDFS permission system only to prevent accidental misuse and overwriting of data among trusted users sharing a Hadoop cluster. HDFS doesn’t authenticate users and believes the user identity to be whatever the host operating system says it is. Your Hadoop username is your login name, which is equivalent to what’s shown by whoami . Your group list is equivalent to bash -c groups. An exception is the username that started the name node. That username has a special Hadoop username superuser . This superuser can perform any file operation regardless of permission settings. In addition, the administrator can specify members in a supergroup through the configuration parameter dfs.permissions. supergroup . All members of the supergroup are also superusers. Enabling trash 179 You can change permission settings and ownership using bin/hadoop fs -chmod, -chown, and -chgrp. They behave similarly to Unix commands of the same name. 8.4 Managing quotas By default HDFS doesn’t have any quota to limit how much you can put in a direc- tory. You can enable and specify name quotas on specific directories, which place a hard limit on the number of file and directory names under that directory. The main use case for name quotas is to prevent users from generating too many small files and overwork the NameNode . The following commands are for setting and clearing name quotas: bin/hadoop dfsadmin -setQuota directory [...directory] bin/hadoop dfsadmin -clrQuota directory [...directory] Starting with version 0.19, HDFS also supports space quotas on a per directory basis. This helps manage the amount of storage a user or application can take up. bin/hadoop dfsadmin -setSpaceQuota directory [...directory] bin/hadoop dfsadmin -clrSpaceQuota directory [...directory] The setSpaceQuota command takes an argument for the number of bytes as each directory’s quota. The argument can have a suffix to represent unit. For example, 20g will mean 20 gigabytes, and 5t would mean 5 terabytes. All replicas count towards the quota. To get the quotas associated with a directory as well as a count of the number of names and bytes it uses, use the HDFS shell command count with the -q option. bin/hadoop fs -count -q directory [...directory] 8.5 Enabling trash In addition to file permissions, an additional safeguard against accidental deletion of files in HDFS is the trash feature. By default this feature is disabled. When this feature is enabled, the command line utilities for deleting files don’t delete files immediately. Instead, they move the files temporarily to a .Trash/ folder under the user’s working directory. The files are not permanently removed until after a user-configurable time delay. As long as a file is still in the .Trash/ folder, you can restore it by moving it back to its original location. To enable the trash feature and set the time delay for the trash removal, set the fs.trash.interval property in core-site.xml to the delay (in minutes). For example, if you want users to have 24 hours (1,440 minutes) to restore a deleted file, you should have in core-site.xml fs.trash.interval 1440 Setting the value to 0 will disable the trash feature. 180 CHAPTER 8 Managing Hadoop 8.6 Removing DataNodes You may want to remove DataNodes from your HDFS cluster at some point. For example, you want to take a machine offline for upgrade or maintenance. Remov- ing nodes in Hadoop can be straightforward. Although it’s not recommended, you can kill the nodes or disconnect them from the cluster. HDFS is designed to be re- silient. Taking one or two DataNodes offline will not affect ongoing operation. The NameNode will detect their death and will initiate replication of blocks that have fallen below the desired replication factor. For a smoother and safer operation, par- ticularly when retiring large number of DataNodes, you should use Hadoop’s decom- missioning feature. Decommissioning ensures that all blocks will have the desired replication factor among the remaining active nodes. In order to use this feature, you must create an (initially empty) exclude file in the NameNode’s local file- system, and the configuration parameter dfs.hosts.exclude must point to this file during NameNode’s startup. When you want to retire DataNodes, list them in the exclude file, one node per line. You have to specify the nodes using the full hostname, IP, or IP:port format. Execute bin/hadoop dfsadmin -refreshNodes to force the NameNode to reread the exclude file and start the decommissioning pro- cess. Messages like “Decommission complete for node” will appear in the NameNode log files when it finishes decommissioning, at which point you can remove the nodes from the cluster. If you have started HDFS without setting dfs.hosts.exclude to point to an exclude file, the proper way to decommission DataNodes is this: Shut down the NameNode. Set dfs.hosts.exclude to point to an empty exclude file. Restart NameNode. After NameNode has successfully restarted, follow the procedure above. Note that if you list the retiring DataNodes in the exclude file before restarting NameNode, the NameNode will be confused and throw messages like “ProcessReport from unregistered node: node055:50010” in its logs. The NameNode thinks that it is being contacted by a DataNode outside the system rather than a node to be decommissioned. If the decommissioned machines may rejoin the cluster at some later point, you should remove them from the exclude file and rerun bin/hadoop dfsadmin -refreshNodes now to update the NameNode. When the machines are ready to rejoin the cluster, you can add them using the procedure described in the next section. DataNodes8.7 Adding Besides bringing back a machine from offline maintenance, you may want to add DataNodes to your Hadoop cluster as you use it for more processing jobs with more data. On the new node, install Hadoop and set up the configuration files as you would any DataNode in the cluster. Start the DataNode daemon manually (bin/hadoop datanode). It will automatically contact the NameNode and join the cluster. You should also add the new node to the conf/slaves file in the master server. The script- based commands will recognize the new node. Managing NameNode and Secondary NameNode 181 When you add a new DataNode, it will initially be empty, whereas existing DataNodes will already be filled to some capacity. The filesystem is considered unbalanced. New files will likely go to the new node, but their replicated blocks will still go to the old nodes. One should proactively start the HDFS balancer to balance the cluster for optimal performance. Run the balancer script: bin/start-balancer.sh The script will run in the background until the cluster is balanced. An administrator can also terminate it earlier by running bin/stop-balancer.sh A cluster is considered balanced when the utilization rates of all the DataNodes are within the range of the average utilization rate plus or minus a threshold. This thresh- old is 10 percent by default. You can specify a different threshold when you start the balancer script. For example, to set the threshold to 5 percent for a more evenly distributed cluster, start the balancer with bin/start-balancer.sh -threshold 5 As balancing can be network intensive, we recommend doing it overnight or over a weekend when your cluster may be less busy. Alternatively, you can set the dfs.balance. bandwidthPerSec configuration parameter to limit the bandwidth devoted to balancing. 8.8 Managing NameNode and Secondary NameNode NameNode is one of the most important components in the HDFS architecture. It holds the filesystem’s metadata and caches the cluster’s blockmap in RAM for rea- sonable performance. When you have anything other than a tiny cluster, you should dedicate a machine to run as NameNode and don’t put any DataNode, JobTracker, or TaskTracker service on it. This NameNode machine should be the most powerful machine in the cluster. Give it as much RAM as possible. Although DataNodes may have higher performance with JBOD disk drives, you should definitely use RAID drives in your NameNode for higher reliability against any single drive failure. One approach to reducing the burden on the NameNode is to reduce the amount of filesystem metadata by increasing the block size. Doubling the block size will almost half the amount of metadata. Unfortunately, this also decreases parallelism for files that are not large. The ideal block size will depend on your specific deployment. The block size is set in the configuration parameter dfs.block.size . For example, to double the block size from the default 64 MB to 128 MB, set dfs.block.size to 134217728. By default, the Secondary NameNode3 and the NameNode run on the same machine. For moderate size clusters (10 or more nodes), you should separate the 3 As of this writing, the Secondary NameNode is slated to be deprecated by version 0.21 of Hadoop, which should be released as this book goes to press. The Secondary NameNode will be replaced by a more robust design for warm standby. You should check the online documentation of the version of Hadoop you’re using to confirm whether it’s still using Secondary NameNode or not. The particular patch for this change is at https://issues.apache.org/jira/browse/HADOOP-4539. 182 CHAPTER 8 Managing Hadoop Secondary NameNode into its own machine, the spec of which should be comparable to the NameNode. But, before going into how to set up a separate server as a Secondary NameNode, I should explain what the Secondary NameNode does and doesn’t do, and in turn some of NameNode’s underlying mechanics. Due to its unfortunate naming, the Secondary NameNode (SNN) is sometimes confused with a failover backup for NameNode. It most certainly is not. The SNN only serves to periodically clean up and tighten the filesystem’s state information in NameNode, helping NameNode become more efficient. NameNode manages the filesystem’s state information using two files, FsImage and EditLog . The file FsImage is a snapshot of the filesystem at some checkpoint, and EditLog records each incremental change (delta) to the filesystem after that checkpoint. These two files can completely determine the current state of the filesystem. When you initialize NameNode, it merges these two files to create a new snapshot. At the end of NameNode’s initialization, FsImage will contain the new snapshot and EditLog will be empty. Afterward any operation that changes the state of HDFS is appended to EditLog, whereas FsImage will remain unchanged. When you shut down NameNode and restart it, the consolidation will take place again and make a new snapshot. Note that the two files are only for retaining the filesystem’s state information while NameNode is not running (either intentionally shut down or due to system malfunction ). NameNode keeps in memory a constantly maintained copy of the filesystem’s state information to quickly answer queries about the filesystem. For a busy cluster , the EditLog file will grow quite large, and the next restart of NameNode will take a long time to merge EditLog into FsImage. For busy clusters, it can also be a long time in between NameNode restarts, and you may want more frequent snapshots for archival purposes. This is where SNN comes in. It consolidates FsImage and EditLog into a new snapshot and leaves the NameNode alone to serve live traffic. Therefore, it’s more appropriate to think of the SNN as a checkpointing server. Merging FsImage and EditLog is memory intensive, requiring an amount of memory on the same order as normal NameNode operation. It’s best for the SNN to be on a separate server that is as powerful as the primary NameNode. To configure HDFS to use a separate server as the SNN, first list that server’s host name or IP address in the conf/masters file. Unfortunately, this file name is also confusing. The masters in Hadoop (NameNode and JobTracker) are whichever machine you run bin/start-dfs.sh and bin/start-mapred.sh on. What’s listed in conf/masters is the SNN, not any of the masters. You should also modify the conf/hdfs-site.xml file on the SNN such that the dfs. http.address property points to port 50070 of the NameNode’s host address, like dfs.http.address namenode.hadoop-host.com:50070 You should set this property because the SNN retrieves FsImage and EditLog from the NameNode by sending HTTP Get requests to the URLs: Recovering from a failed NameNode 183 ■ FsImage—http://namenode.hadoop-host.com:50070/getimage?getimage=1 ■ EditLog—http://namenode.hadoop-host.com:50070/getimage?getedit=1 The SNN also updates the NameNode with the merged metadata using the same ad- dress and port. 8.9 Recovering from a failed NameNode Failures happen, and Hadoop has been designed to be quite resilient. The NameNode, unfortunately, remains a weak point. HDFS is out of commission if the NameNode is down. A common design for setting up a backup NameNode server is by reusing the SNN.4 After all, the SNN has similar hardware specs as the NameNode, and Hadoop should’ve already been installed with the same directory configurations. If we do some additional work of maintaining the SNN to be a functional mirror image of the NameNode, we can quickly start this backup machine as a NameNode instance in the case of a NameNode failure. Some manual intervention and time are necessary to start the backup node as the new NameNode, but at least we wouldn’t lose any data. NameNode keeps all the filesystem’s metadata, including the FsImage and EditLog files, under the dfs.name.dir directory. Note that the SNN server doesn’t use that directory at all. It downloads the system’s metadata into the fs.checkpoint.dir directory and proceeds to merge FsImage and EditLog there. As the dfs.name.dir directory on the SNN is unused, we can expose it to the NameNode via the Network File System (NFS). We instruct the NameNode to always write to this mounted directory in addition to writing to the NameNode’s local metadata directory. HDFS supports this ability to write the metadata to multiple directories. You have to specify dfs.name.dir on the NameNode with a comma separated list, like dfs.name.dir /home/hadoop/dfs/name,/mnt/hadoop-backup true This works, assuming the local dfs.name.dir directory at both the NameNode and the Secondary NameNode are at /home/hadoop/dfs/name, and that the directory on the SNN is mounted to the NameNode at /mnt/hadoop-backup. When HDFS sees a comma- separated list in dfs.name.dir, it writes its metadata to every directory on the list. Given this setup, when the NameNode dies, the local dfs.name.dir directory at both the NameNode and the backup node (SNN) should have the same content. To have the backup node serve as the replacement NameNode, you’ll have to switch its IP address to the original NameNode’s IP address. (Unfortunately, changing only the hostname is not sufficient as the DataNodes cache the DNS entry.) You’ll also have to run the backup node as a NameNode by executing bin/start-dfs.sh on it. 4 Unfortunately, this common design also contributes to the misperception of the Secondary NameNode as a backup node. You can set up the backup node in a totally different machine from the NameNode and SNN, but that machine would be idle almost all the time. 184 CHAPTER 8 Managing Hadoop To be safer, this new NameNode should also have a backup node set up before you start it. Otherwise you’ll be in trouble if this new NameNode fails too. If you don’t have a machine readily available as a backup, you should at least set up an NFS-mounted directory. This way the filesystem’s state information is in more than one location. As HDFS writes its metadata to all directories listed in dfs.name.dir, if your NameNode has multiple hard drives, you can specify directories from different drives to hold replicas of the metadata. This way if one drive fails, it’s easier to restart the NameNode without the bad drive than to switch over to the backup node, which involves moving the IP address, setting up a new backup node, and so on. Recall that the SNN creates a snapshot of the filesystem’s metadata in the fs.checkpoint.dir directory. As it checkpoints only periodically (once an hour under the default setup), the metadata is too stale to rely on for failover. But it’s still a good idea to archive this directory periodically over to remote storage . In catastrophic situations, recovering from stale data is better than no data at all. This can be true if both the NameNode and the backup fail simultaneously (say, a power surge affecting both machines). Another unfortunate scenario is if the filesystem’s metadata has been corrupted (say, by human error or a software bug) and has poisoned all the replicas. Recovering from a checkpoint image is explained in http://issues.apache.org/jira/ browse/HADOOP-2585. HDFS’s backup and recovery mechanism is undergoing active improvements as of this writing. You should check with HDFS’s online documentation for the latest news. There have also been applications of specialized Linux software such as DRBD 5 to Hadoop clusters for high availability. You can find one example in http://www. cloudera.com/blog/2009/07/22/hadoop-ha-configuration/. 8.10 Designing network layout and rack awareness When your Hadoop cluster gets big, the nodes will be spread out in more than one rack and the cluster’s network topology starts to affect reliability and performance. You may want the cluster to survive the failure of an entire rack. You should place your backup server for NameNode, as described in the previous section, in a separate rack from the NameNode itself. This way the failure of any one rack will not destroy all cop- ies of the filesystem’s metadata. With more than one rack, the placement of both block replicas and tasks becomes more complex. Replicas of a block should be placed in separate racks to reduce the potential of data loss. For the standard replication value of 3, the default placement policy for writing a block is this: If the client performing the write operation is part of the Hadoop cluster, place the first replica on the DataNode where the client resides. Otherwise randomly place the replica in the cluster. Place the second replica on a random rack different from the rack where the first replica resides. Write the third replica to a different node on the same rack as the second replica. For replication values higher than 3, place the subsequent replicas on random nodes. As of this 5 http://www.drbd.org. Designing network layout and rack awareness 185 writing, this block placement policy is baked into the NameNode. A pluggable policy is targeted for version 0.21.6 Besides block placement, task placement is also rack aware . A task is usually placed on a node that has a copy of the block the task is assigned to process. When no such node is available to take on the new task, the task is randomly assigned to a node on a rack where a copy of the block is available somewhere on that rack. That is, when data locality can’t be enforced at a node level, Hadoop tries to enforce it at the rack level. Failing that, a task would be randomly assigned to one of the remaining nodes. At this point you may wonder how Hadoop knows which rack a node is at. It requires you to tell it. It assumes a hierarchical network topology for your Hadoop cluster, structurally similar to figure 8.1. Each node has a rack name similar to a file path. For example, the nodes H1, H2, and H3 in figure 8.1 all have a rack name of /D1/R1. Figure 8.1 shows a case where you have multiple datacenters (D1 and D2) each with multiple racks (R1 to R4). In most cases you’ll be dealing with multiple racks co-located together. Your rack names will be in a flat namespace, such as /R1 and /R2. To help Hadoop know the location of each node, you have to provide an executable script that can map IP addresses into rack names. This network topology script must reside on the master node and its location is specified in the topology.script.file.name property in core-site.xml. Hadoop will call this script with a set of IP addresses as separate arguments. The script should print out (through STDOUT) the rack name corresponding to each IP address in the same order, separated by whitespace. The topology.script.number.args property controls the maximum number of IP addresses Hadoop will ask for at any one time . It’s convenient to simplify your script by setting that value to 1. Here is an example a network topology script. / D1 R1 H1 D2 R2 R3 R4 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12 Figure 8.1 A cluster with a hierarchical network topology. This cluster spans multiples datacenters (D1 and D2). Each datacenter has multiple racks (R), and each rack has multiple machines. 6 See http://issues.apache.org/jira/browse/HDFS-385 for the description of this change. 186 CHAPTER 8 Managing Hadoop #!/bin/bash ipaddr=$1 segments='echo $ipaddr | cut --delimiter=. --fields=4' if [ "$segments" -lt 128 ]; then echo /rack-1 else echo /rack-2 fi This bash script takes an IPv4 address and looks at the last of the four octets (assuming dot-decimal notation). The node is considered in rack 1 if the last octet is less than 128, and the node is considered in rack 2 otherwise. A table lookup may make more sense for more complex cluster topologies. On the other hand, if there is no net- work topology script, Hadoop assumes a flat topology where all nodes are assigned to /default-rack. 8.11 Scheduling jobs from multiple users As you have more and more jobs coming from multiple users for your Hadoop cluster, you’ll need some control to prevent contention. Under Hadoop’s default FIFO sched- uler , as soon as a job is sent to Hadoop for execution, the JobTracker will assign as many TaskTrackers as necessary to process that job. This works fairly well when things are not busy and you have a good amount of processing capacity to spare. But some big Hadoop jobs can easily tie up the cluster for a long time and force the smaller jobs to wait. Wouldn’t it be great if something akin to an express checkout existed for smaller jobs in a Hadoop cluster? 8.11.1 Multiple JobTracker s Back in the days before Hadoop version 0.19, you had to physically set up multiple MapReduce clusters to provide rudimentary CPU allocation among jobs. To keep storage utilization reasonably efficient though, there would still be one single HDFS cluster. Let’s say, you have Z slave nodes available for your Hadoop cluster. You’ll have a single NameNode that takes all Z nodes as DataNodes. All Z nodes will also be TaskTrackers. Until now, all these TaskTrackers would point to the same/only JobTracker. The trick in making a multicluster setup is having multiple JobTrackers, and each JobTracker controls a (mutually exclusive) subset of TaskTrackers. For example, to create two MapReduce clusters, you have X TaskTrackers point to one JobTracker (via the mapred.job.tracker property) and Y TaskTrackers configured to use the second JobTracker. The slave nodes between the two MapReduce clusters are distinct to give X+Y=Z. To use this setup, you submit certain jobs to one JobTracker, whereas other jobs go to the other JobTracker. This limits the number of TaskTrackers available to each type of job. The job type need not necessarily determine the assignment of jobs to the MapReduce pool. More typical is to allocate each MapReduce pool to one user group. This enforces a limit to how much resource one group can take up. Scheduling jobs from multiple users 187 Physically setting up multiple MapReduce clusters this way has many drawbacks. It’s not very user-friendly as one has to remember which pool to use. It’s less likely that a task is data-local. (It may be the case that all replicas are in DataNodes outside of one’s pool.) The setup is not flexible to changing resource requirements. Fortunately, starting with version 0.19, Hadoop has a pluggable architecture for the scheduler, and two new schedulers have become available for resolving job contention. One is the Fair Scheduler developed at Facebook, and another one is the Capacity Scheduler developed at Yahoo. 8.11.2 Fair Scheduler The Fair Scheduler introduced the notion of pools . Jobs are tagged to belong to specific pools, and each pool is configured to have a guaranteed capacity of a certain number of map slots and a certain number of reduce slots. When task slots are freed up, the Fair Scheduler will allocate them to meet these minimum guarantees first. After it meets the guarantees, slots are allocated between jobs using “fair sharing,” such that each job gets roughly an equal amount of compute resource. You can set priority on jobs to give more capacity to higher priority jobs. (Some jobs are more equal than others.) The Fair Scheduler is available as the jar file hadoop-*-fairscheduler.jar under the contrib/fairscheduler directory of the Hadoop installation. To install it, you can move the jar file directly into Hadoop’s lib/ directory. Alternatively, you can modify HADOOP_CLASSPATH in the script conf/hadoop-env.sh to include this jar. You’ll need to set a few properties in hadoop-site.xml to fully enable and configure the Fair Scheduler. You first instruct Hadoop to use the Fair Scheduler instead of the default one by setting mapred.jobtracker.taskScheduler to org.apache.hadoop. mapred.FairScheduler. You then configure a few Fair Scheduler properties. The most important is mapred.fairscheduler.allocation.file, which points to the file that defines the different pools. This file is typically named pools.xml and specifies each pool’s name and capacity. The mapred.fairscheduler.poolnameproperty defines the jobconf property the scheduler will use to determine which pool to use for a job. A useful configuration pattern is to set this to a new property, say pool.name, and assign pool.name to have a default value of ${user.name}. The Fair Scheduler automatically gives each user her own individual pool. This particular pool.name will by default assign each job to its owner’s pool. You can change the pool.name property in a job’s jobconf to assign the job to a different pool.7 Finally, the mapred. fairscheduler.assignmultiple property, when set to true, allows the scheduler to assign both a map task and a reduce task on each heartbeat, which improves ramp-up speed and throughput. To summarize, your mapred-site.xml will have the following properties set: 7 Yes, you can run your job in another user’s pool, but that’s not very polite. The main usage is to assign special jobs to specific pools. For example, you may want all cron jobs to go to a single pool rather than have them run under each individual user’s pool. 188 CHAPTER 8 Managing Hadoop mapred.jobtracker.taskScheduler org.apache.hadoop.mapred.FairScheduler mapred.fairscheduler.allocation.file HADOOP_CONF_DIR/pools.xml mapred.fairscheduler.assignmultiple true mapred.fairscheduler.poolnameproperty pool.name pool.name ${user.name} The allocation file pools.xml defines the pools for the scheduler. It gives each pool a name and capacity constraints. The constraints can include the minimum number of map slots or reduce slots. They can also include the maximum number of running jobs. In addition, you can set the maximum number of running jobs per user, and also override this maximum for specific users. An example pools.xml looks like this: 2 2 2 2 2 6 3 This pools.xml defines two special pools, “ads” and “hive”. Each is guaranteed to have at least two map slots and two reduce slots. The “hive” pool is limited to running at most two jobs at once. To use these pools, you set the pool.name property in a job’s configuration to either “ads” or “hive”. This pools.xml also caps the number of simul- taneous running jobs a user can have to three, but the user “chuck” is given a higher cap of six. Note that the pools.xml file is reread every 15 seconds. You can modify this file and dynamically reallocate capacity at run time. Any pool not defined in this file has no guaranteed capacity and no limit on number of jobs running at once. Summary 189 Figure 8.2 The Web UI to monitor Hadoop’s fair scheduler. The top table shows all the available pools and each pool’s usage. The table showing “Running Jobs” has a Pool column where you can monitor or change the pool of each job. When you have your Hadoop cluster running with the Fair Scheduler , there’s a Web UI available to administer the scheduler . The page is at http:/// scheduler. Besides letting you know how the jobs are scheduled, it also allows you to change the pool a job belongs to and the job’s priority. Figure 8.2 shows an example screenshot of this page. The Capacity Scheduler shares similar goals with the Fair Scheduler. The Capacity Scheduler works on queues rather than pools. The interested reader can learn more about the Capacity Scheduler from the online documentation at http://hadoop. apache.org/common/docs/r0.20.0/capacity_scheduler.html. 8.12 Summary Managing distributed clusters is complicated and Hadoop is no different. We’ve covered many common administrative tasks in this chapter. If you have a complex setup and have more sophisticated questions, a useful resource is the Hadoop mailing lists.8 Many Hadoop administrators with deep expertise are active on those mailing lists, and chances are that one of them will have encountered your situation. On the other hand, if you mostly want a basic Hadoop cluster without all the hassle of administration, you may want to consider using the Cloudera distribution9 or checking out one of the Hadoop cloud services, which we cover in the next chapter. 8 http://hadoop.apache.org/common/mailing_lists.html. 9 http://www.cloudera.com/distribution. Part 3 Hadoop Gone Wild Part 3 explores the larger ecosystem around Hadoop. Cloud services provide an alternative to buying and hosting your own hardware to create a Hadoop clus- ter. Many add-on packages provide higher-level programming abstractions over MapReduce. We show several case studies where Hadoop solves real business problems in practice. 193 9Running Hadoop in the cloud This chapter covers ■ Setting up a compute cloud with Amazon Web Services (AWS) ■ Transferring data into and out of an AWS Hadoop cloud Running Hadoop in the AWS cloud ■ Depending on your data processing needs, your Hadoop workload can vary widely over time. You may have a few large data processing jobs that occasionally take ad- vantage of hundreds of nodes, but those same nodes will sit idle the rest of the time. You may be new to Hadoop and want to get familiar with it first before investing in a dedicated cluster. You may own a startup that needs to conserve cash and wants to avoid the capital expense of a Hadoop cluster. In these and other situations, it makes more sense to rent a cluster of machines rather than buy it. The general paradigm of provisioning compute resources as a remote service in a flexible, cost-effective manner is called cloud computing. The best-known infrastructure service platform for cloud computing is Amazon Web Services (AWS). You can rent computing and storage services from AWS on demand as your requirement scales. As of this writing, renting a compute unit with the equivalent power of a 1.0 GHz 194 CHAPTER 9 Running Hadoop in the cloud 32-bit Opteron with 1.7 GB RAM and 160 GB disk storage costs $0.10 per hour. Using a cluster of 100 such machines for an hour will cost a measly $10! It wasn’t too long ago that only a select few had access to clusters on this scale. Thanks to AWS and other such services, large-scale compute power is available to many people today. Because of its flexibility and cost effectiveness, running Hadoop on the AWS cloud is a popular setup, and we learn how to install and configure this configuration in this chapter. 9.1 Introducing Amazon Web Services Learning all the capabilities of Amazon Web Services is worthy of a book itself. Amazon is constantly adding new services and features. We recommend you to explore the AWS website (http://aws.amazon.com) to get more details and the latest offerings. We only cover enough basics here to get a Hadoop cluster running. Of all the services AWS offers, its Elastic Compute Cloud (EC2) and Simple Storage Service (S3 ) are the two core services you need to understand to run Hadoop in the cloud . The EC2 service provides compute capacity for running Hadoop nodes. You can think of EC2 as a large farm of virtual machines. An EC2 instance is the AWS terminology for a virtual compute unit. Each Hadoop node will take up an EC2 instance. You rent an EC2 instance for only as long as you need and pay on an hourly basis. A car rental company throws out whatever you leave in the trunk when you return it. Similarly, all your data on an EC2 instance is deleted when you terminate the instance. If you want the data to be around for future use, you have to ensure that it’s in some persistent storage. The Amazon S3 service is a cloud storage service that you may use for such purposes. Each EC2 instance functions like a commodity Intel machine that you can access and control over the internet. You boot up an instance using an Amazon Machine Image , also known as an AMI or an image. More demanding users can create their own images, but most users are well served by one of the many preconfigured ones. Some images are only basic bare-bones operating systems. Supported operating systems on EC2 include more than six variants of Linux, plus Windows Server and OpenSolaris. Other images include one of the operating systems plus pre-installed software, such as database server, Apache HTTP server, Java application server, and others. AWS offers preconfigured images of Hadoop running on Linux, and Hadoop has built-in support for working with both EC2 and S3. 9.2 Setting up AWS This section is a quick introduction to setting up AWS. We only cover the essentials to get a Hadoop cluster running. If you are already familiar with launching and using EC2 instances, you should skip directly to the next section on setting up Hadoop on AWS. To start using AWS, you first have to sign up for an account. Go to http://aws. amazon.com/ and click on the button urging you to “Sign Up Now.” The process is self-explanatory. It’s no more complicated than buying a book from Amazon. The sign-up process sets up your Amazon account (which you may already have if you have bought stuff from Amazon before) and activates it to pay for your usage of AWS. Setting up AWS 195 NOTE Amazon has introduced the Elastic MapReduce (EMR) service that vastly simplifies the use of Hadoop on AWS. The most important simplification is that you no longer have to set up and launch your own cluster of EC2 instances. The trade-off is that you lose some control over how the cluster works and you have to pay extra for this EMR service. We discuss EMR in section 9.6. But we highly encourage you to keep on reading and understand how to set up your own EC2 cluster running Hadoop, even if you don’t go through the process of setting it up. At the very least, knowing more details about how Hadoop runs on an EC2 cluster will clarify what EMR is doing underneath the hood. After you have activated your Amazon account for AWS, there are three more steps before you can start creating machine instances and using them: 1 Obtain your account IDs and your authentication keys and certificates. You’ll set these up on your local machine to secure your communication with AWS. These security mechanisms ensure that only you can rent compute resources with your account. 2 Download and set up command line tools to manage your EC2 instances. These include special programs to start and stop EC2 instances in your virtual cluster. 3 Generate an SSH key pair. After you have started an EC2 instance, you’ll log into it using SSH (either directly or indirectly through the use of special tools). The default SSH mechanism uses the SSH key pair to authenticate you to your EC2 instance in lieu of using a password. We go through each step in a subsection below. 9.2.1 Getting your AWS authentication credentials AWS supports two types of authentication mechanisms: the AWS Access Key Identifier and X.509 certificate . To run Hadoop on AWS you’ll need both mechanisms, and they can be set up from the Access Identifiers page where you manage your AWS account (http://aws.amazon.com/account/). The AWS Access Key Identifier consists of an Access Key ID and a Secret Access Key. Figure 9.1 shows a section of the Access Identifiers page. An Access Key ID is a 20-character alphanumeric sequence whereas a Secret Ac- cess Key is a 40-character sequence. Don’t share your Secret Access Key with anyone. The web page requires an extra click on Show to display it (in case anyone is looking over your shoulder). You should generate a new Secret Access Key if the one you have has been compromised. You’ll need to specify your Access Key ID and Secret Access Key later when you set up the Hadoop cluster. TIP In some situations when you want Hadoop to access S3, you’ll tell Hadoop your AWS Access Key ID and Secret Access Key in a specially formatted URI. Unfortunately, AWS allows slashes (/) in its Secret Access Keys, which will cause confusion inside a URI. Although there are ways you can tell Hadoop your AWS Access Key ID without using a URI, it may be more convenient to regenerate your Secret Access Key until you get one without a slash (/) in it. 196 CHAPTER 9 Running Hadoop in the cloud Figure 9.1 Getting AWS Access Key ID and Secret Access Key Setting up the X.509 Certificate involves a bit more work. At the same Access Identi- fiers page is a section titled X.509 Certificate, as seen in figure 9.2. You click on Create New to generate a new X.509 certificate. A certificate has two keys: a public key and a private key. Unlike the Access Key ID and the Secret Access Key, the public and private keys in your X.509 certificate are hundreds of characters long, and they have to be stored and managed as files. After creating a new X.509 certificate, you’ll arrive at a page to download both keys/files. See figure 9.3. Figure 9.2 Managing your X.509 certificate. You can upload your own, or ask AWS to create one. Figure 9.3 Download the private key and certificate file of your X.509 Certificate. The public key is also called a certificate file . Your private key, as the name implies, should not be shared with anyone. Even Amazon itself doesn’t store a copy of it. AWS specifies filenames for the certificate and the private key , and you should keep those names when you save them. The filenames for the certificate and the private key are prefixed with cert- and pk-, respectively, and they have the .pem file extension. You should create a directory called .ec2 under your home directory on your local machine and save those two files in the new directory. On Linux you’ll have saved the following files to your local machine: ~/.ec2/cert-HKZYKTAIG2ECMXYIBH3HXV4ZBZQ55CLO.pem ~/.ec2/pk-HKZYKTAIG2ECMXYIBH3HXV4ZBZQ55CLO.pem Finally, you should also note your AWS Account Number . It’s near the top right side of the Access Identifiers page, and it’s a 12-digit hyphenated number that looks like “4952-1993-3132.” Your Account ID is the unhyphenated version of that—something like “495219933132.” Your Account ID is what you’ll use in setting up Hadoop on EC2. Setting up AWS 197 198 CHAPTER 9 Running Hadoop in the cloud You may think these are a lot of values to generate and write down. To summarize, here are the five values you should have at this point: ■ Your 20-character, alphanumeric Access Key ID ■ Your 40-character Secret Access Key ■ Your X.509 Certificate file under the .ec2 directory ■ Your X.509 private key file under the .ec2 directory ■ Your 12-digit (unhyphenated) AWS Account ID You’ll use these values later to authenticate yourself to AWS and control your Hadoop cluster. 9.2.2 Getting command line tools After getting all the security credentials, you should download and configure the AWS command line tools to instantiate and manage your EC2 instances. These tools are written in Java, which presumably is already installed on your local machine. The EC2 command line tools are self-contained in one downloadable ZIP file from the AWS EC2 Resource Center.1 Unzip the file into a directory you use for your AWS work. In the unzipped files you’ll see the Java tools plus shell scripts for Windows, Linux, and Mac OS X. You don’t have to configure the command line tools, but you do have to set a few environment variables before using them. The environment variable EC2_HOME should be a path pointing to the directory where the command line tools were unzipped. Unless you have renamed that directory, its name is ec2-api-tools-A.B-nnnn, where A, B, and n are version/release numbers. You should also set EC2_CERT and EC2_PRIVATE_KEY to point to your X.509 certificate and private key files, respectively. I’ve found it useful to have a script for setting up all the necessary environment variables to use the AWS command line tools. A version for Linux, Unix, and Mac OS X, called ec2-init.sh, is in listing 9.1. You run this script before using any AWS related tools by executing source ec2-init.sh or . ec2-init.sh Listing 9.1 ec2-init.sh: a Unix script setting up variables for EC2 tools export JAVA_HOME = /Library/Java/Home export EC2_HOME = ~/Projects/Hadoop/aws/ec2-api-tools-1.3-30349 export PATH = $PATH:$EC2_HOME/bin:$HADOOP_HOME/src/contrib/ec2/bin export EC2_PRIVATE_KEY = ~/.ec2/pk-HKZYKTAIG2ECMXYIBH3HXV4ZBZQ55CLO.pem export EC2_CERT = ~/.ec2/cert-HKZYKTAIG2ECMXYIBH3HXV4ZBZQ55CLO.pem 1 http://developer.amazonwebservices.com/connect/entry.jspa?externalID=351&categoryID=88. An analogous version for Windows is in listing 9.2. You can run it in a command prompt by executing ec2-init.bat. Listing 9.2 ec2-init.bat: a Windows script setting up variables for EC2 tools set JAVA_HOME = "C:\Program Files\Java\jdk1.6.0_08" set EC2_HOME = "C:\Program Files\Hadoop\aws\ec2-api-tools-1.3-30349" set PATH = %PATH%;%EC2_HOME%\bin;%HADOOP_HOME%\src\contrib\ec2\bin set EC2_PRIVATE_KEY = c:\ec2\pk-HKZYKTAIG2ECMXYIBH3HXV4ZBZQ55CLO.pem set EC2_CERT = c:\ec2\cert-HKZYKTAIG2ECMXYIBH3HXV4ZBZQ55CLO.pem If you’ll be working with AWS often, you may choose to not use a separate script and instead integrate it directly into your operating system’s start-up script (for example, .profile and autoexec.bat). The pathnames in the script will be different for your particular installation. The environment variable JAVA_HOME needs to be set for the AWS command line tools to work. We set it here although most likely it has already been set elsewhere. The script adds the command line tools’ bin directory to your system PATH. This makes executing the tools much easier, as you don’t need to specify the full path every time. The Hadoop EC2 tools’ directory is also added to PATH, although we won’t cover them until the next section. AWS has machines located in different regions of the world. As of this writing, AWS supports two regions, the U.S. and the E.U. As an optional step, you can choose which region to run your EC2 instances to reduce latency. After you have run the preceding script to set up the environment variables, let’s run our first AWS command line tool to ask Amazon what regions are currently available: ec2-describe-regions You’ll get back something like this: REGION us-east-1 us-east-1.ec2.amazonaws.com REGION eu-west-1 eu-west-1.ec2.amazonaws.com The second column is the region names (us-east-1 and eu-west-1) and the third column is the corresponding service endpoints. The default region is us-east-1. Set the environment variable EC2_URL to the service endpoint of a different region if you choose to. You can do this within the preceding AWS initialization shell script. TIP Besides the official command line tools, there are also GUI tools for managing your EC2 and S3 usage. These GUI tools tend to be more user- friendly. Two of the more popular ones are both Firefox extensions: Elasticfox and S3Fox. Elasticfox (http://developer.amazonwebservices.com/connect/ entry.jspa?entryID=609) provides basic EC2 management features, such as launching new EC2 instances and listing currently running ones. S3Fox (http://www.suchisoft.com/ext/s3fox.php) is a third-party tool for organizing your S3 storage. It functions much like a GUI-based FTP client in managing remote storage. Setting up AWS 199 200 CHAPTER 9 Running Hadoop in the cloud 9.2.3 Preparing an SSH key pair After starting an EC2 instance, you’ll want to log into it to run programs and services. The default login mechanism (of public images) uses SSH with a key pair. Half of this key pair (public key ) is embedded in the EC2 instance, and the other half (private key ) is in your local machine. Together, the key pair secures the communication between your local machine and the EC2 instance. NOTE Some of you may be more familiar with logging into a remote machine using SSH with a password. Using SSH with a key pair is an alternative mechanism. Instead of a password, you authenticate yourself with a private key that’s stored as a file on your local machine. Like your password, your private key file should not be accessible by unauthorized people. Each SSH key pair has a key name to identify it. When requesting Amazon EC2 to cre- ate an instance, you have to specify the public key to be embedded in that instance by its corresponding key name. The SSH public key has to exist and be registered with Amazon before creating any EC2 instances. The following command generates an SSH public/private key pai r and registers the public key to Amazon EC2 under the key name gsg-keypair. ec2-add-keypair gsg-keypair Interestingly, the command doesn’t save the private key to a local file. Instead, it generates a standard output (stdout) similar to figure 9.4, part of which is the pri- vate key. You’ll have to manually save it to a file using a text editor. Specifically, copy and paste the output between the following two lines, inclusive, to a new file named id_rsa-gsg-keypair. -----BEGIN RSA PRIVATE KEY----- -----END RSA PRIVATE KEY----- For ease of management, you should save the file to the same .ec2 directory for your X.509 private key and certificate. You’ll also need to lock down the file permission such that it’s only readable by you. On Linux and other Unix-based systems, use the following command: chmod 600 ~/.ec2/id_rsa-gsg-keypair All EC2 instances in a single Hadoop cluster will have the same public key . A single private key can log into any node in the cluster, and only one SSH key pair is needed. You can also choose to use more than one SSH key pair when working with multiple Hadoop clusters, or when you use extra EC2 instances outside of your Hadoop cluster. At this point you have finished the one-time setup of credentials and certificates to start a compute cluster in the Amazon cloud. You can manually use the AWS tools to launch EC2 instances and log into them to launch your Hadoop cluster . This approach is time consuming and error prone, though. Fortunately, Hadoop has included tools to work with AWS, which we discuss in the next section. Before we finish this section, we Setting up Hadoop on EC2 201 Figure 9.4 Example output of ec2-add-keypair. The first line is a key signature and the rest is the private key. do recommend that you spend some time reading the EC2 documentation, includ- ing the Getting Started Guide.2 EC2 has many configuration and customization options. Understanding them will come in handy when you start to tune your AWS Hadoop cluster. 9.3 Setting up Hadoop on EC2 To run Hadoop on an EC2 cluster, you first need to install Hadoop on your local ma- chine and be able to run it in standalone mode. Your local Hadoop installation con- tains scripts to help you launch and log into an EC2 Hadoop cluster. These scripts are in the directory src/contrib/ec2/bin under your Hadoop installation. 9.3.1 Configuring security parameters You must configure the single initialization script at src/contrib/ec2/bin/hadoop- ec2-env.sh. Inside that script, set the following three variables to values you obtained in section 9.2.1: 2 http://docs.amazonwebservices.com/AWSEC2/latest/GettingStartedGuide/ 202 CHAPTER 9 Running Hadoop in the cloud ■ AWS_ACCOUNT_ID—Your 12-digit AWS account ID ■ AWS_ACCESS_KEY_ID—Your 20-character, alphanumeric Access Key ID ■ AWS_SECRET_ACCESS_KEY—Your 40-character Secret Access Key The tools for Hadoop on EC2 get the other security parameters from environment variables (which should be set when you source aws-init.sh) or are based on defaults that should work fine if you have followed the AWS setup in section 9.2. 9.3.2 Configuring cluster type You’ll need to specify the configuration of your Hadoop cluster in hadoop-ec2-env. sh. You need to set three main parameters: HADOOP_VERSION, INSTANCE_TYPE, and S3_BUCKET. Before telling you how to set these parameters, let’s go over a little background. Before the creation of an instance, Amazon EC2 must know the instance type and the image used to boot up the instance. Instance type is the physical configuration of the virtual machine (CPU, RAM, disk space, etc.). As of this writing, five instance types are available, grouped into two families: standard and high-CPU. High-CPU types are for compute-intensive work. Rarely are they used for Hadoop applications, which tend to be data-intensive. The standard family has three instance types, and table 9.1 lists their attributes. The more powerful instance types cost more, and you should look up the AWS website to find the latest pricing. Only Amazon’s S3 storage service can store images for booting up EC2 instance. Many existing images are available for all kinds of setups. You can use one of the public images, or pay for special custom images, or even create your own. Similar images are stored in the same S3 bucket .3 The standard public Hadoop images are either in the hadoop-ec2-images bucket or the hadoop-images bucket. In fact, we only use the hadoop- images bucket because the newer versions of Hadoop (after 0.17.1) aren’t available in the hadoop-ec2-images bucket. The Hadoop team puts new EC2 images in the hadoop- images bucket when significant versions of Hadoop are released. At any point in time, execute the following EC2 command to see the available Hadoop images: ec2-describe-images -x all | grep hadoop-images Table 9.1 Specification for various EC2 instance types Type CPU Memory Storage Platform I/O Name Small 1 EC2 Compute Unit 1.7 GB 160 GB 32-bit Moderate m1.small Large 4 EC2 Compute Unit 7.5 GB 850 GB 64-bit High m1.large Extra Large 8 EC2 Compute Unit 15 GB 1690 GB 64-bit High m1.xlarge 3 An S3 bucket is the top-level partition in S3’s namespace. A bucket is owned by exactly one AWS account and must have a globally unique name. Figure 9.5 Some of the available Hadoop images in AWS. Figure 9.5 shows an example output from the previous command. Each row describes one available EC2 image. Each image lists eleven properties, most of which are useful only for advanced AWS users. For our purpose, all the information we need can be read from the third column, also known as the manifest location of the image. These are ex- pressed in a two-level hierarchy, in which the top level is the S3 bucket where the image resides. As mentioned earlier, the hadoop-images bucket is the one we focus on. The manifest location includes the Hadoop version number. The manifest location also includes a term that’s either i386 or x86_64. This tells you whether the image is for a 32-bit instance or a 64-bit instance. An example image that’s available as of this writing has a manifest location of hadoop-images/hadoop-0.19.0- i386.manifest.xml. That image uses Hadoop version 0.19.0 and can run on 32-bit EC2 instances. After seeing the available Hadoop images, we’re ready to set HADOOP_VERSION, INSTANCE_TYPE, and S3_BUCKET in hadoop-ec2-env.sh. Unless you’re using a custom image, you should set S3_BUCKET to hadoop-images. INSTANCE_TYPE defaults to m1.small, which should work fine. The main point to remember is that the instance type directly specifies whether the CPU is 32-bit or 64-bit and must be booted from a compatible image (i386 or x86_64). Finally, HADOOP_VERSION should be set to the Hadoop version you want to use. The particular combination of HADOOP_VERSION, INSTANCE_TYPE, and S3_BUCKET has to be available as seen in the output of the ec2-describe-images command. 9.4 Running MapReduce programs on EC2 The Hadoop EC2 tools are in the directory src/contrib/ec2/bin under your Hadoop installation. Recall that our ec2-init.sh script has already added that directory to your system PATH. Within that directory is hadoop-ec2, which is a meta-command for execut- ing other commands. To launch a Hadoop cluster on EC2, use hadoop-ec2 launch-cluster This first launches a master EC2 instance. After this it boots the requested number of slave nodes to point to the master node. When this command returns, it will print out the public DNS name to the master node, which we denote . At this point, not all slave nodes necessarily have been fully booted. You can view the master node’s JobTracker web interface at http://:50030/ to monitor the cluster and the operational status of the slave nodes. Running MapReduce programs on EC2 203 204 CHAPTER 9 Running Hadoop in the cloud NOTE New EC2 users can’t run more than 20 concurrent instances. You can request a higher limit through the Amazon EC2 Instance Request Form at http://www.amazon.com/gp/html-forms-controller/ec2-request. After launching a Hadoop cluster, you’re ready to log into the master node and use the cluster as if you’ve set it up on your own machines. To log in, use this command: hadoop-ec2 login The $HADOOP_HOME for a Hadoop EC2 instance is /usr/local/hadoop-x.y.z, where x.y.z is the Hadoop version number. We run a quick test to show that Hadoop is running in this cluster: # cd /usr/local/hadoop-* # bin/hadoop jar hadoop-*-examples.jar pi 10 10000000 For the rest of this chapter, a hash mark (#) character in the beginning of a command line denotes that line should be executed in the master node of your Hadoop EC2 cluster, rather than at your local machine. The commands above run an example Hadoop program to estimate the value of pi. You can monitor the job at http://:50030/. 9.4.1 Moving your code to the Hadoop cluster All Hadoop applications consist of two components: code and data. We first move our code to the Hadoop cluster. In the next subsection we discuss making our data acces- sible (which may or may not involve moving data to the cluster). You’ll copy your application code to the master node in your Hadoop EC2 cluster using scp. Execute the following commands from your local machine: source hadoop-ec2-env.sh scp $SSH_OPTS root@$MASTER_HOST: where points to the application code on your local machine and points to the destination file path on the master node. 9.4.2 Accessing your data from the Hadoop cluster As the Hadoop EC2 cluster is being rented, data stored in the cluster (including in HDFS) is not persistent. Your input data has to persist somewhere else and be brought into the EC2 cluster for processing. Many options exist for where to put your data and bring it into the Hadoop cluster, and each option has its trade-offs. MOVING DATA DIRECTLY INTO HDFS When the input data is small (<100 GB) and is processed only once, the simplest approach is to copy the data into the master node and then copy it from there to the cluster’s HDFS. Copying data into the master node is no different than copying application code into the master node (see section 9.4.1). Once the data is in the master node, you log into the master node and copy the data into HDFS using the standard Hadoop command: # bin/hadoop fs –put Hadoop MapReduce HDFS EC2 Cloud Data Master Node Figure 9.6 Transferring data to the Hadoop EC2 cloud directly The data flowpath is depicted graphically in figure 9.6. Some issues are worth noting in this data flowpath. One is that AWS charges for bandwidth between AWS and the outside world (in addition to hourly charges of each EC2 instance), but bandwidth within AWS is free. In this case there’s a monetary cost to copying data into the mas- ter node but not the copying of the data from the master node into HDFS. (There’s also no cost to any of the communication within MapReduce processing and between MapReduce and HDFS.) Whichever way you get data into the Hadoop cluster, you’ll incur this bandwidth cost at least once. The time it takes to move data into the master node will also be relatively long, as the connection between your machine and AWS is much slower than the connections within AWS. Again, this sunk time will be incurred at least once no matter how you architect the dataflow. The problem with the current dataflow architecture is that you’ll incur the time and monetary costs each time you bring up the Hadoop cluster. If the input data will be processed in different ways over multiple sessions, this dataflow is not recommended. Another shortcoming to the existing flowpath is the size limitation on the input data. All the data must be able to reside at the master node first, and a small EC instance only has 150 GB of disk partition. You can overcome this limitation if you can divide your input data over several chunks and move one chunk at a time. You may also choose to use bigger instances, which have multiple 420 GB disk partitions. But before trying these more complicated schemes, you should consider using S3 in your datapath. MOVING DATA INTO HDFS THROUGH S3 S3 is a cloud storage service offered by AWS. You’ve already seen it in action as storage for EC2 images. Storing data in S3 is charged by bandwidth for data I/O with non- AWS machines, plus a monthly storage charge based on the size of the data. The cost Running MapReduce programs on EC2 205 206 CHAPTER 9 Running Hadoop in the cloud model makes it an attractive storage service for many applications. More particularly, it’s well suited for use with Hadoop EC2 clusters. You can see the dataflow model in figure 9.7. The main change from the dataflow of figure 9.6 is that your input data is first transferred to the S3 cloud instead of the master node. Note that, unlike the master node, the S3 cloud storage persists independently of your Hadoop EC2 cluster. You can create and terminate multiple Hadoop EC2 clusters over time, and they can all read the same input data from S3. The benefit of this setup is that you incur the monetary and time costs of copying your input data into AWS only once, when it’s copied into S3, whereas in the dataflow of figure 9.6 they’re incurred on every session of the Hadoop EC2 cluster. After the input data is copied into S3, copying it from the S3 cloud to the cluster’s HDFS is fast and free, because both S3 and EC2 are managed within the AWS system. There’s now an additional monthly storage cost for hosting your input data in S3, but it’s usually minimal. If you need to have a scalable archival storage for your data, S3 can function in that role under this dataflow architecture, further justifying its cost model. The default Hadoop installation has built-in support for using S3. There’s a special Hadoop filesystem for S3, called S3 Block FileSystem , built on top of S3 to enable large file sizes. (S3 imposes a file size limit of 5 GB.) You’ll need to consider the S3 Block FileSystem a separate filesystem from S3, just as HDFS is treated distinctly from the underlying Unix filesystem. The S3 Block FileSystem requires a dedicated S3 bucket . Once you’ve created that S3 bucket, you can move your data from the local machine to S3: bin/hadoop fs –put ➥ s3://:@/ Hadoop MapReduce HDFS EC2 Cloud S3 Cloud Data Figure 9.7 Using Hadoop on AWS with both S3 and HDFS Running MapReduce programs on EC2 207 Recall that and are authentication param- eters from section 9.2.1, and is the name of the S3 bucket you’ve created for the S3 Block FileSystem. After your data is in S3, you can copy it to any Hadoop EC2 cluster. From the master node of the cluster, run # bin/hadoop distcp s3://:@/ ➥ After the data is in HDFS, you can run your Hadoop program in the cluster in the usual fashion. ACCESSING DATA DIRECTLY FROM S3 Up ’till now we’ve always copied data into the cluster’s HDFS before running our Hadoop applications. This preserves data locality between storage and the MapReduce program. For very small jobs, you may choose to bypass HDFS and forego data locality, in return for skipping the copying of data from S3 to HDFS. You can see this datapath in figure 9.8. To work in this architecture, specify S3 as the input filepath when executing your Hadoop application: # bin/hadoop jar s3://: ➥ @/ The preceding command will store the output file in HDFS, but you can change that to be S3 as well. MORE OPTIONS IN USING S3 FOR HADOOP There are a couple of variations on how you can use S3 that may be useful in some situations. Hadoop MapReduce HDFS HDFS EC2 Cloud S3 Cloud Data Figure 9.8 Hadoop running on EC2 can directly access data in S3. 208 CHAPTER 9 Running Hadoop in the cloud Up ’till now we’ve used a special Hadoop S3 filesystem (the S3 Block FileSystem) to store data in S3. An alternative is to use S3’s native filesystem. The main disadvantage with the native system is a limitation on file size of 5 GB. If the files for your input data are smaller than that limit, the native system can be an excellent option. It’s compat- ible with all the standard S3 tools, whereas Hadoop’s S3 filesystem is in a special and unique format. Standard S3 tools make the S3 native filesystem more transparent and easier to understand. To use the S3 native filesystem instead of the S3 Block FileSystem, substitute s3 with s3n when specifying file locations. For example, use s3n://:@/ in place of s3://:@/ If you’re using S3 often, you’ll find it cumbersome to type out the long URI for each file you want to access. One way to shorten it is to add the following to your conf/core-site.xml file: fs.s3.awsAccessKeyId AWS_ACCESS_KEY_ID fs.s3.awsSecretAccessKey AWS_SECRET_ACCESS_KEY Note that you have to replace AWS_ACCESS_KEY_ID with your 20-character Access Key ID and AWS_SECRET_ACCESS_KEY with your 40-character Secret Access Key . After add- ing the two properties to core-site.xml, the URI for S3 files can be shortened to s3:/// or s3n:/// for the native S3 filesystem. NOTE If you’re unfortunate enough to be stuck using a Secret Access Key that has a slash (/) in it, you can’t embed the secret key inside a URI. The only way to use that AWS/S3 account is by embedding the secret key in core- site.xml as described before. (Some documentation has suggested escaping the secret key by replacing the slash (/) with the string %2F inside a URI, although that doesn’t seem to work in practice.) For an additional shortcut, it may be appropriate to make S3 your default filesystem , in place of HDFS. To do this, change the fs.default.name property in conf/core- site.xml after adding the two properties above: fs.default.name s3://S3_BUCKET where S3_BUCKET is the S3 bucket you had chosen to use as your Hadoop S3 Block FileSystem (which we had denoted earlier). 9.5 Cleaning up and shutting down your EC2 instances Hadoop stores the output data of your Hadoop job in the cluster’s HDFS by default, and you should save it to somewhere more persistent. The options for retrieving out- put data are the same as the options for copying input data into the Hadoop EC2 cluster, only running in reverse. The main difference is that the output data is usually orders of magnitude smaller than the input data. Given generally small output data, copying through the master node may turn out to be your best option. As you’re renting your EC2 instances from AWS on an hourly basis, it’s important that you shut down the instances when you’re done and tell AWS to stop charging you. It’s easy to log out of a cluster and forget that the instances are still running and you’re being charged! To properly terminate a cluster, use the following command: bin/hadoop-ec2 terminate-cluster All the EC2 instances in the cluster will shut down and data in them lost. No further cleanup is necessary. 9.6 Amazon Elastic MapReduce and other AWS services Amazon Web Services is constantly adding new capabilities, many of which will make life easier for Hadoop developers. Two of the newest services that they’ve announced during the writing of this book include Amazon Elastic MapReduce (EMR) and AWS Import/Export. 9.6.1 Amazon Elastic MapReduce For a small extra fee, the EMR service will launch a preconfigured Hadoop cluster for you to run your MapReduce programs. The major simplification this service provides is that you don’t need to worry about setting up EC2 instances, and therefore you don’t need to deal with all the certificates and command line tools and so forth. You interact with EMR purely through a web-based console at https://console.aws.amazon. com/elasticmapreduce/home. You can see its initial screen in figure 9.9. The design targets the processing of single jobs. You submit a MapReduce job, either as a (Streaming, Pig, or Hive) script or a JAR file, and EMR will set up a cluster to run the job. By default the cluster will shut down at the end of the job . The input (output) of the job is read (written) directly to S3. A heavy user of Hadoop usually has many jobs running against the same data, making this setup relatively inefficient, as explained in section 9.4.2. But a light user will find EMR dramatically simplifies running MapReduce in the cloud. In addition, it’s not difficult to imagine that the sophistication of EMR will only grow in the future and eventually become the primary way to run Hadoop on AWS. Amazon Elastic MapReduce and other AWS services 209 210 CHAPTER 9 Running Hadoop in the cloud Figure 9.9 The introductory screen of the web console to Amazon Elastic MapReduce. You can follow the steps onscreen to create a job flow. You can find more information about Amazon Elastic MapReduce at these sites: http://aws.amazon.com/elasticmapreduce/ http://docs.amazonwebservices.com/ElasticMapReduce/latest/GettingStartedGuide/ 9.6.2 AWS Import/Export One of the main obstacles to large-scale data processing in the cloud is the difficulty of moving large data sets into the cloud. If you already have an existing process that stores your data in S3, then it’s relatively straightforward to run Hadoop on EC2 to access that data. On the other hand, if you have to move data to the Amazon cloud for the sole purpose of analyzing it, then the data transfer itself can be a signifi- cant hurdle. Amazon introduced the AWS Import/Export service by which you can physically send a hard drive to them and they upload the data to S3 using their high-speed internal network. The point at which this service makes sense depends on your available internet connection speed. Table 9.2 is a rough guideline given by AWS. Table 9.2 Size of data set at which AWS Import/Export is more practical than internet upload Available internet connection Theoretical min. number of days to transfer 1 TB at 80% network utilization When to consider AWS Import/Export T1 (1.544 Mbps) 82 days 100 GB or more 10 Mbps 13 days 600 GB or more T3 (44.736 Mbps) 3 days 2 TB or more 100 Mbps 1 to 2 days 5 TB or more 1000 Mbps Less than 1 day 60 TB or more You can find more details about AWS Import/Export at http://aws.amazon.com/ importexport/. 9.7 Summary Cloud infrastructure is a great place for running Hadoop, as it allows you to easily scale to hundreds of nodes and gives you the financial flexibility to avoid upfront investments. Hadoop has native support for Amazon Web Services (AWS). This chapter started with the basics of setting up an account and renting compute services from AWS. Once you’re ready to rent computing nodes from AWS, you’ll find Hadoop tools for automating the setting up and running of a Hadoop cluster. AWS also has a cloud storage service (S3) which can be used in conjunction with or in place of HDFS. You’ll find pros and cons to the different setups. Finally, it’s important to remember to shut down your Hadoop cluster when you’re finished. You’re renting the cloud infrastruc- ture by the hour, and fees will continue to accrue unless you explicitly shut down the machines. Summary 211 10 212 Programming with Pig This chapter covers ■ Installing Pig and using the Grunt shell ■ Understanding the Pig Latin language ■ Computing similar documents efficiently, using a simple Pig Latin script Extending the Pig Latin language with user-defined functions ■ One frequent complaint about MapReduce is that it’s difficult to program. When you first think through a data processing task, you may think about it in terms of data flow operations, such as loops and filters. However, as you implement the program in MapReduce, you’ll have to think at the level of mapper and reducer functions and job chaining. Certain functions that are treated as first-class operations in higher- level languages become nontrivial to implement in MapReduce, as we’ve seen for joins in chapter 5. Pig is a Hadoop extension that simplifies Hadoop programming by giving you a high-level data processing language while keeping Hadoop’s simple scalability and reliability. Yahoo , one of the heaviest user of Hadoop (and a backer of both the Hadoop Core and Pig), runs 40 percent of all its Hadoop jobs with Pig. Twitter is also another well-known user of Pig.1 1 http://www.slideshare.net/kevinweil/hadoop-pig-and-twitter-nosql-east-2009. Thinking like a Pig 213 Pig has two major components: 1 A high-level data processing language called Pig Latin . 2 A compiler that compiles and runs your Pig Latin script in a choice of evaluation mechanisms . The main evaluation mechanism is Hadoop. Pig also supports a local mode for development purposes. Pig simplifies programming because of the ease of expressing your code in Pig Latin. The compiler helps to automatically exploit optimization opportunities in your script. This frees you from having to tune your program manually. As the Pig compiler im- proves, your Pig Latin program will also get an automatic speed-up. 10.1 Thinking like a Pig Pig has a certain philosophy about its design. We expect ease of use, high perfor- mance, and massive scalability from any Hadoop subproject. More unique and crucial to understanding Pig are the design choices of its programming language (a data flow language called Pig Latin), the data types it supports, and its treatment of user-defined functions (UDFs ) as first-class citizens. 10.1.1 Data flow language You write Pig Latin programs in a sequence of steps where each step is a single high- level data transformation. The transformations support relational-style operations, such as filter, union, group, and join. An example Pig Latin program that processes a search query log may look like log = LOAD 'excite-small.log' AS (user, time, query); grpd = GROUP log BY user; cntd = FOREACH grpd GENERATE group, COUNT(log); DUMP cntd; Even though the operations are relational in style, Pig Latin remains a data flow language. A data flow language is friendlier to programmers who think in terms of algorithms, which are more naturally expressed by the data and control flows. On the other hand, a declarative language such as SQL is sometimes easier for analysts who prefer to just state the results one expects from a program. Hive is a different Hadoop subproject that tar- gets users who prefer the SQL model. We’ll learn about Hive in detail in chapter 11. 10.1.2 Data types We can summarize Pig’s philosophy toward data types in its slogan of “Pigs eat any- thing.” Input data can come in any format. Popular formats, such as tab-delimited text files, are natively supported. Users can add functions to support other data file formats as well. Pig doesn’t require metadata or schema on data, but it can take advantage of them if they’re provided. Pig can operate on data that is relational, nested, semistructured, or unstructured. To support this diversity of data, Pig supports complex data types, such as bags and tuples that can be nested to form fairly sophisticated data structures. 214 CHAPTER 10 Programming with Pig 10.1.3 User-defined functions Pig was designed with many applications in mind—processing log data, natural lan- guage processing, analyzing network graphs, and so forth. It’s expected that many of the computations will require custom processing. Pig is architected from the ground up with support for user-defined functions. Knowing how to write UDFs is a big part of learning to use Pig. 10.2 Installing Pig You can download the latest release of Pig from http://hadoop.apache.org/pig/ releases.html. As of this writing, the latest versions of Pig are 0.4 and 0.5. Both of them re- quire Java 1.6. The main difference between them is that Pig version 0.4 targets Hadoop version 0.18 whereas Pig version 0.5 targets Hadoop version 0.20. As usual, make sure to set JAVA_HOME to the root of your Java installation, and Windows users should install Cygwin . Your Hadoop cluster should already be set up. Ideally it’s a real cluster in fully distributed mode, although a pseudo-distributed setup is fine for practice. You install Pig on your local machine by unpacking the downloaded distribution. There’s nothing you have to modify on your Hadoop cluster. Think of the Pig distribution as a compiler and some development and deployment tools. It enhances your MapReduce programming but is otherwise only loosely coupled with the production Hadoop cluster. Under the directory where you unpacked Pig, you should create the subdirectories logs and conf (unless they’re already there). Pig will take custom configuration from files in conf. If you are creating the conf directory just now, there’s obviously no configuration file, and you’ll need to put in conf a new file named pig-env.sh. This script is executed when you run Pig, and it can be used to set up environment variables for configuring Pig. Besides JAVA_HOME, the environment variables of particular interest are PIG_HADOOP_VERSION and PIG_CLASSPATH . You set these variables to instruct Pig about your Hadoop cluster. For example, the following statements in pig-env.sh will tell Pig the version of Hadoop used by the cluster is 0.18, and to add the configuration directory of your local installation of Hadoop to Pig’s classpath: export PIG_HADOOP_VERSION=18 export PIG_CLASSPATH=$HADOOP_HOME/conf/ We assume HADOOP_HOME is set to Hadoop’s installation directory on your local ma- chine. By adding Hadoop’s conf directory to Pig’s classpath, Pig can automatically pick up the location of your Hadoop cluster’s NameNode and JobTracker . Instead of using Pig’s classpath, you can also specify the location of your Hadoop cluster by creating a pig.properties file . This properties file will be under the conf directory you created earlier. It should define fs.default.name and mapred.job.tracker, the filesystem (i.e., HDFS’s NameNode ) and the location of the JobTracker. An example pig. properties file pointing to a Hadoop set up in pseudo-distributed mode is fs.default.name=hdfs://localhost:9000 mapred.job.tracker=localhost:9001 Running Pig 215 For the sake of convenience, let’s add the Pig installation’s bin directory to your path. Assume $PIG_HOME is pointing to your Pig’s installation: export PATH=$PATH:$PIG_HOME/bin With Pig’s bin directory set as part of your command line path, you can start Pig with the command pig. You may want to first see its usage options: pig -help Let’s start Pig’s interactive shell to see that it’s reading the configurations properly. pig 2009-07-11 22:33:04,797 [main] INFO ➥ org.apache.pig.backend.hadoop.executionengine.HExecutionEngine - ➥ Connecting to hadoop file system at: hdfs://localhost:9000 2009-07-11 22:33:09,533 [main] INFO ➥ org.apache.pig.backend.hadoop.executionengine.HExecutionEngine - ➥ Connecting to map-reduce job tracker at: localhost:9001 grunt> The filesystem and the JobTracker Pig reports should be consistent with your configu- ration setup. You’re now inside Pig’s interactive shell, also known as Grunt. Pig10.3 Running We can run Pig Latin commands in three ways—via the Grunt interactive shell, through a script file, and as embedded queries inside Java programs. Each way can work in one of two modes—local mode and Hadoop mode . (Hadoop mode is sometimes called Mapreduce mode in the Pig documentation.) At the end of the previous section we’ve entered the Grunt shell running in Hadoop mode. The Grunt shell allows you to enter Pig commands manually. This is typically used for ad hoc data analysis or during the interactive cycles of program development. Large Pig programs or ones that will be run repeatedly are run in script files. To enter Grunt, use the command pig. To run a Pig script, execute the same pig command with the file name as the argument, such as pig myscript.pig. The convention is to use the .pig extension for Pig scripts. You can think of Pig programs as similar to SQL queries , and Pig provides a PigServer class that allows any Java program to execute Pig queries. Conceptually this is analogous to using JDBC to execute SQL queries. Embedded Pig programs is a fairly advanced topic and you can find more details at http://wiki.apache.org/pig/ EmbeddedPig. When you run Pig in local mode, you don’t use Hadoop at all.2 Pig commands are compiled to run locally in their own JVM , accessing local files. This is typically used for development purposes, where you can get fast feedback by running locally against 2 There are plans to change Pig such that it uses Hadoop even in local mode, which helps to make some programming more consistent. The discussion for this topic is taking place at https://issues.apache.org/ jira/browse/PIG-1053. 216 CHAPTER 10 Programming with Pig a small development data set. Running Pig in Hadoop mode means the compile Pig program will physically execute in a Hadoop installation. Typically the Hadoop installation is a fully distributed cluster . (The pseudo-distributed Hadoop setup we used in section 10.2 was purely for demonstration. It’s rarely used except to debug configurations.) The execution mode is specified to the pig command via the -x or -exectype option. You can enter the Grunt shell in local mode through: pig -x local Entering the Grunt shell in Hadoop mode is pig -x mapreduce or use the pig command without arguments, as it chooses the Hadoop mode by default. 10.3.1 Managing the Grunt shell In addition to running Pig Latin statements (which we’ll look at in a later section), the Grunt shell supports some basic utility commands.3 Typing help will print out a help screen of such utility commands. You exit the Grunt shell with quit. You can stop a Hadoop job with the kill command followed by the Hadoop job ID. Some Pig param- eters are set with the set command . For example, grunt> set debug on grunt> set job.name 'my job' The debug parameter states whether debug-level logging is turned on or off. The job. name parameter takes a single-quoted string and will use that as the Pig program’s Hadoop job name. It’s useful to set a meaningful name to easily identify your Pig job in Hadoop’s Web UI . The Grunt shell also supports file utility commands, such as ls and cp. You can see the full list of utility commands and file commands in table 10.1. The file commands are mostly a subset of the HDFS filesystem shell commands, and their usage should be self-explanatory. Table 10.1 Utility and file commands in the Grunt shell Utility commands help quit kill jobid set debug [on|off] set job.name 'jobname' File commands cat, cd, copyFromLocal, copyToLocal, cp, ls, mkdir, mv, pwd, rm, rmf, exec, run 3 Technically these are still considered Pig Latin commands, but you’ll not likely use them outside of the Grunt shell. Learning Pig Latin through Grunt 217 Two new commands are exec and run. They run Pig scripts while inside the Grunt shell and can be useful in debugging Pig scripts. The exec command executes a Pig script in a separate space from the Grunt shell. Aliases defined in the script aren’t visible to the shell and vice versa. The command run executes a Pig script in the same space as Grunt (also known as interactive mode ). It has the same effect as manually typ- ing in each line of the script into the Grunt shell. 10.4 Learning Pig Latin through Grunt Before formally describing Pig’s data types and data processing operators, let’s run a few commands in the Grunt shell to get a feel for how to process data in Pig. For the purpose of learning, it’s more convenient to run Grunt in local mode: pig -x local You may want to first try some of the file commands, such as pwd and ls, to orient yourself around the filesystem. Let’s look at some data. We’ll later reuse the patent data we introduced in chapter 4, but for now let’s dig into an interesting data set of query logs from the Excite search engine. This data set already comes with the Pig installation, and it’s in the file tutorial/data/excite-small.log under the Pig installation directory. The data comes in a three-column, tab-separated format. The first column is an anonymized user ID. The second column is a Unix timestamp, and the third is the search query. A decidedly non-random sample from the 4,500 records of this file looks like 3F8AAC2372F6941C 970916093724 minors in possession C5460576B58BB1CC 970916194352 hacking telenet 9E1707EE57C96C1E 970916073214 buffalo mob crime family 06878125BE78B42C 970916183900 how to make ecstacy From within Grunt, enter the following statement to load this data into an “alias” (i.e., variable) called log. grunt> log = LOAD 'tutorial/data/excite-small.log' AS (user, time, query); Note that nothing seems to have happened after you entered the statement. In the Grunt shell, Pig parses your statements but doesn’t physically execute them until you use a DUMP or STORE command to ask for the results. The DUMP command prints out the content of an alias whereas the STORE command stores the content to a file. The fact that Pig doesn’t physically execute any command until you explicitly request some end result will make sense once you remember that we’re processing large data sets. There’s no memory space to “load” the data, and in any case we want to verify the logic of the execution plan before spending the time and resources to physically execute it. We use the DUMP command usually only for development. Most often you’ll STORE significant results into a directory. (Like Hadoop, Pig will automatically partition the data into files named part-nnnnn.) When you DUMP an alias, you should be sure that 218 CHAPTER 10 Programming with Pig its content is small enough to be reasonably printed to screen. The common way to do that is to create another alias through the LIMIT command and DUMP the new, smaller alias. The LIMIT command allows you to specify how many tuples (rows) to return back. For example, to see four tuples of log grunt> lmt = LIMIT log 4; grunt> DUMP lmt; (2A9EABFB35F5B954,970916105432L,+md foods +proteins) (BED75271605EBD0C,970916001949L,yahoo chat) (BED75271605EBD0C,970916001954L,yahoo chat) (BED75271605EBD0C,970916003523L,yahoo chat) Table 10.2 summarizes the read and write operators in Pig Latin. LIMIT is techni- cally not a read or write operator, but as it’s often used alongside, we’ve included it in the table. Table 10.2 Data read/write operators in Pig Latin LOAD alias = LOAD 'file' [USING function] [AS schema]; Load data from a file into a relation. Uses the PigStorage load function as default unless specified otherwise with the USING option. The data can be given a schema using the AS option. LIMIT alias = LIMIT alias n; Limit the number of tuples to n. When used right after alias was processed by an ORDER operator, LIMIT returns the first n tuples. Otherwise there’s no guarantee which tuples are returned. The LIMIT operator defies categorization because it’s certainly not a read/write operator but it’s not a true relational operator either. We include it here for the practical reason that a reader looking up the DUMP operator, explained later, will remember to use the LIMIT operator right before it. DUMP DUMP alias; Display the content of a relation. Use mainly for debugging. The relation should be small enough for printing on screen. You can apply the LIMIT operation on an alias to make sure it’s small enough for display. STORE STORE alias INTO 'directory' [USING function]; Store data from a relation into a directory. The directory must not exist when this command is executed. Pig will create the directory and store the relation in files named part-nnnnn in it. Uses the PigStorage store function as default unless specified otherwise with the USING option. You may find loading and storing data not terribly exciting. Let’s execute a few data processing statements and see how we can explore Pig Latin through Grunt. grunt> log = LOAD 'tutorial/data/excite-small.log' ➥ AS (user:chararray, time:long, query:chararray); grunt> grpd = GROUP log BY user; grunt> cntd = FOREACH grpd GENERATE group, COUNT(log); grunt> STORE cntd INTO 'output'; The preceding statements count the number of queries each user has searched for. The content of the output files (you’ll have to look at the file from outside Grunt) look like this: Learning Pig Latin through Grunt 219 002BB5A52580A8ED 18 005BD9CD3AC6BB38 18 00A08A54CD03EB95 3 011ACA65C2BF70B2 5 01500FAFE317B7C0 15 0158F8ACC570947D 3 018FBF6BFB213E68 1 Conceptually we’ve performed an aggregating operation similar to the SQL query : SELECT user, COUNT(*) FROM excite-small.log GROUP BY user; Two main differences between the Pig Latin and SQL versions are worth pointing out. As we’ve mentioned earlier, Pig Latin is a data processing language. You’re specifying a series of data processing steps instead of a complex SQL query with clauses. The other difference is more subtle—relations in SQL always have fixed schemas . In SQL, we define a relation’s schema before it’s populated with data. Pig takes a much looser approach to schema. In fact, you don’t need to use schemas if you don’t want to, which may be the case when handling semistructured or unstructured data. Here we do specify a schema for the relation log, but it’s only in the load statement and it’s not enforced until we’re loading in the data. Any field that doesn’t obey the schema in the load operation is casted to a null. In this way the relation log is guaranteed to obey our stated schema for subsequent operations. As much as possible, Pig tries to figure out the schema for a relation based on the operation used to create it. You can expose Pig’s schema for any relation with the DESCRIBE command. This can be useful in understanding what a Pig statement is doing. For example, we’ll look at the schemas for grpd and cntd. Before doing this, let’s first see how the DESCRIBE command describes log. grunt> DESCRIBE log; log: {user: chararray,time: long,query: chararray} As expected, the load command gives log the exact schema we’ve specified. The rela- tion log consists of three fields named user, time, and query. The fields user and query are both strings (chararray in Pig) whereas time is a long integer. A GROUP BY operation on the relation log generates the relation grpd. Based on the operation and the schema for log, Pig infers a schema for grpd: grunt> DESCRIBE grpd; grpd: {group: chararray,log: {user: chararray,time: long,query: chararray}} group and log are two fields in grpd. The field log is a bag with subfields user, time, and query. As we haven’t covered Pig’s type system and the GROUP BY operation, we don’t expect you to understand this schema yet. The point is that relations in Pig can have fairly complex schemas, and DESCRIBE is your friend in understanding the rela- tions you’re working with: grunt> DESCRIBE cntd; cntd: {group: chararray,long} 220 CHAPTER 10 Programming with Pig Finally, the FOREACH command operates on the relation grpd to give us cntd. Having looked at the output of cntd, we know it has two fields—the user ID and a count of the number of queries. Pig’s schema for cntd, as given by DESCRIBE, also has two fields. The first one’s name—group—is taken from grpd’s schema. The second field has no name, but it has a type of long. This field is generated by the COUNT function , and the function doesn’t automatically provide a name, although it does tell Pig that the type has to be a long. Whereas DESCRIBE can tell you the schema of a relation, ILLUSTRATE does a sample run to show a step-by-step process on how Pig would compute the relation. Pig tries to simulate the execution of the statements to compute a relation, but it uses only a small sample of data to make the execution fast. The best way to understand ILLUSTRATE is by applying it to a relation. In this case we use cntd. (The output is reformatted to fit the width of a printed page.) grunt> ILLUSTRATE cntd; ---------------------------------------------------------------------- | log | user: bytearray | time: bytearray | query: bytearray | ---------------------------------------------------------------------- | | 0567639EB8F3751C | 970916161410 | "conan o'brien" | | | 0567639EB8F3751C | 970916161413 | "conan o'brien" | | | 972F13CE9A8E2FA3 | 970916063540 | finger AND download | ---------------------------------------------------------------------- ------------------------------------------------------------------- | log | user: chararray | time: long | query: chararray | ------------------------------------------------------------------- | | 0567639EB8F3751C | 970916161410 | "conan o'brien" | | | 0567639EB8F3751C | 970916161413 | "conan o'brien" | | | 972F13CE9A8E2FA3 | 970916063540 | finger AND download | ------------------------------------------------------------------- --------------------------------------------------------------------------- | grpd | group: chararray | log: bag({user: chararray,time: long, | | | | query: chararray}) | --------------------------------------------------------------------------- | | 0567639EB8F3751C | {(0567639EB8F3751C, 970916161410, | | | | "conan o'brien"), | | | | (0567639EB8F3751C,970916161413, | | | | "conan o'brien")} | | | 972F13CE9A8E2FA3 | {(972F13CE9A8E2FA3, 970916063540, | | | | finger AND download)} | --------------------------------------------------------------------------- --------------------------------------- | cntd | group: chararray | long | --------------------------------------- | | 0567639EB8F3751C | 2 | | | 972F13CE9A8E2FA3 | 1 | --------------------------------------- The ILLUSTRATE command shows there to be four transformations to arrive at cntd. The header row of each table describes the schema of the output relation after trans- formation, and the rest of the table shows example data. The log relation is shown as two transformations. The data hasn’t changed from one to the next, but the schema has changed from a generic bytearray (Pig’s type for binary objects) to the specified Speaking Pig Latin 221 schema. The GROUP operation on log is executed on the three sample log tuples to arrive at the data for grpd. Based on this we can infer the GROUP operation to have taken the user field and made it the group field. In addition, it groups all tuples in log with the same user value into a bag in grpd. Seeing sample data in a simulated run by ILLUSTRATE can greatly aid the understanding of different operations. Finally, we see the FOREACH operation applied to grpd to arrive at cntd. Having seen the data in grpd in the previous table, one can easily infer that the COUNT() function provided the size of each bag. Although DESCRIBE and ILLUSTRATE are your workhorses in understanding Pig Latin statements, Pig also has an EXPLAIN command to show the logical and physical execution plan in detail. We summarize the diagnostic operators in table 10.3. Table 10.3 Diagnostic operators in Pig Latin DESCRIBE DESCRIBE alias; Display the schema of a relation. EXPLAIN EXPLAIN [-out path] [-brief] [-dot] [-param ...] [-param_file ...] alias; Display the execution plan used to compute a relation. When used with a script name, for example, EXPLAIN myscript.pig, it will show the execution plan of the script. ILLUSTRATE ILLUSTRATE alias; Display step-by-step how data is transformed, starting with a load command, to arrive at the resulting relation. To keep the display and processing manageable, only a (not completely random) sample of the input data is used to simulate the execution. In the unfortunate case where none of Pig’s initial sample will survive the script to generate meaningful data, Pig will “fake” some similar initial data that will survive to generate data for alias. For example, consider these operations: A = LOAD 'student.data' as (name, age); B = FILTER A by age > 18; ILLUSTRATE B; If every tuple Pig samples for A happens to have age less than or equal to 18, B is empty and not much is “illustrated.” Instead Pig will construct for A some tuples with age greater than 18. This way B won’t be an empty relation and users can see how the script works. In order for ILLUSTRATE to work, the load command in the first step must include a schema. The subsequent transformations must not include the LIMIT or SPLIT operators, or the nested FOREACH operator, or the use of the map data type (to be explained in section 10.5.1). 10.5 Speaking Pig Latin You now know how to use Grunt to run Pig Latin statements and investigate their execution and results. We can come back and give a more formal treatment of the language. You should feel free to use Grunt to explore these language concepts as we present them. 222 CHAPTER 10 Programming with Pig 10.5.1 Data types and schemas Let’s first look at Pig data types from a bottom-up view. Pig has six simple atomic types and three complex types, shown in tables 10.4 and 10.5 respectively. The atomic types include numeric scalars as well as string and binary objects. Type casting is sup- ported and done in the usual manner. Fields default to bytearray unless specified otherwise. Table 10.4 Atomic data types in Pig Latin int Signed 32-bit integer long Signed 64-bit integer float 32-bit floating point double 64-bit floating point chararray Character array (string) in Unicode UTF-8 bytearray Byte array (binary object) The three complex types are tuple, bag, and map. A field in a tuple or a value in a map can be null or any atomic or complex type. This enables nesting and complex data structures. Whereas data structures can be arbitrarily complex, some are definitely more useful and occur more often than others, and nesting usually doesn’t go deeper than two levels. In the Excite log example earlier, the GROUP BY operator generated a relation grpd where each tuple has a field that is a bag. The schema for the relation seems more natural once you think of grpd as the query history of each user. Each tuple represents one user and has a field that is a bag of the user’s queries. We can also look at Pig’s data model from the top down. At the top, Pig Latin statements work with relations, which is a bag of tuples. If you force all the tuples in a bag Table 10.5 Complex data types in Pig Latin Tuple (12.5,hello world,-2) A tuple is an ordered set of fields. It’s most often used as a row in a relation. It’s represented by fields separated by commas, all enclosed by parentheses. Bag {(12.5,hello world,-2),(2.87,bye world,10)} A bag is an unordered collection of tuples. A relation is a special kind of bag, sometimes called an outer bag. An inner bag is a bag that is a field within some complex type. A bag is represented by tuples separated by commas, all enclosed by curly brackets. Tuples in a bag aren’t required to have the same schema or even have the same number of fields. It’s a good idea to do this though, unless you’re handling semistructured or unstructured data. Map [key#value] A map is a set of key/value pairs. Keys must be unique and be a string (chararray). The value can be any type. Speaking Pig Latin 223 to have a fixed number of fields and each field has a fixed atomic type, then it behaves like a relational data model —the relation is a table, tuples are rows (records), and fields are columns. But, Pig’s data model has more power and flexibility by allowing nested data types . Fields can themselves be tuples, bags, or maps. Maps are helpful in processing semistructured data such as JSON, XML, and sparse relational data. In ad- dition, it isn’t necessary that tuples in a bag have the same number of fields. This allows tuples to represent unstructured data. Besides declaring types for fields, schemas can also assign names to fields to make them easier to reference. Users can define schemas for relations using the AS keyword with the LOAD, STREAM, and FOREACH operators. For example, in the LOAD statement for getting the Excite query log, we defined the data types for the fields in log, as well as named the fields user, time, and query. grunt> log = LOAD 'tutorial/data/excite-small.log' ➥ AS (user:chararray, time:long, query:chararray); In defining a schema, if you leave out the type, Pig will default to bytearray as the most generic type. You can also leave out the name, in which case a field would be unnamed and you can only reference it by position. 10.5.2 Expressions and functions You can apply expressions and functions to data fields to compute various values. The simplest expression is a constant value . Next is to reference the value of a field . You can reference the named fields’ value directly by the name. You can reference an unnamed field by $n, where n is its position inside the tuple. (Position is num- bered starting at 0.) For example, this LOAD command provides named fields to log through the schema. log = LOAD 'tutorial/data/excite-small.log' ➥ AS (user:chararray, time:long, query:chararray); The three named fields are user, time, and query. For example, we can refer to the time field as either time or $1, because the time field is the second field in log (posi- tion number 1). Let’s say we want to extract the time field into its own relation; we can use this statement: projection = FOREACH log GENERATE time; We can also achieve the same with projection = FOREACH log GENERATE $1; Most of the time you should give names to fields. One use of referring to fields by posi- tion is when you’re working with unstructured data. When using complex types, you use the dot notation to reference fields nested inside tuples or bags. For example, recall earlier that we’d grouped the Excite log by user ID and arrived at relation grpd with a nested schema . 224 CHAPTER 10 Programming with Pig --------------------------------------------------------------------------- | grpd | group: chararray | log: bag({user: chararray,time: long, | | | | query: chararray}) | --------------------------------------------------------------------------- | | 0567639EB8F3751C | {(0567639EB8F3751C, 970916161410, | | | | "conan o'brien"), | | | | (0567639EB8F3751C,970916161413, | | | | "conan o'brien")} | | | 972F13CE9A8E2FA3 | {(972F13CE9A8E2FA3, 970916063540, | | | | finger AND download)} | --------------------------------------------------------------------------- The second field in grpd is named log, of type bag. Each bag has tuples with three named fields: user, time, and query. In this relation, log.query would refer to the two copies of “conan” “o’brien” when applied to the first tuple. You can get the same field with log.$2. You reference fields inside maps through the pound operator instead of the dot operator . For a map named m, the value associated with key k is referenced through m#k. Being able to refer to values is only a first step. Pig supports the standard arithmetic, comparison, conditional, type casting , and Boolean expressions that are common in most popular programming languages. See table 10.6. Table 10.6 Expressions in Pig Latin Constant 12, 19.2, 'hello world' Constant values such as 19 or “hello world.” Numeric values without decimal point are treated as int unless l or L is appended to the number to make it a long. Numeric values with a decimal point are treated as double unless f or F is appended to the number to make it a float. Basic arithmetic +,-,*,/ Plus, minus, multiply, and divide. Sign +x, -x Negation (-) changes the sign of a number. Cast (t)x Convert the value of x into type t. Modulo x % y The remainder of x divided by y. Conditional (x ? y : z) Returns y if x is true, z otherwise. This expression must be enclosed in parentheses. Comparison ==,!=,<,>, <=,>= Equals to, not equals to, greater than, less than, etc. Pattern matching x matches regex Regular expression matching of string x. Uses Java’s regular expression format (under the java.util.regex. Pattern class) to specify regex. Null x is null, x is not null Check if x is null (or not). Boolean x and y, x or y not x Boolean and, or, not. Speaking Pig Latin 225 Furthermore, Pig also supports functions. Table 10.7 shows Pig’s built-in functions, most of which are self-explanatory. We’ll discuss user-defined functions (UDF) in section 10.6. Table 10.7 Built-in functions in Pig Latin AVG Calculate the average of numeric values in a single-column bag. CONCAT Concatenate two strings (chararray) or two bytearrays. COUNT Calculate the number of tuples in a bag. See SIZE for other data types. DIFF Compare two fields in a tuple. If the two fields are bags, it will return tuples that are in one bag but not the other. If the two fields are values, it will emit tuples where the values don’t match. MAX Calculate the maximum value in a single-column bag. The column must be a numeric type or a chararray. MIN Calculate the minimum value in a single-column bag. The column must be a numeric type or a chararray. SIZE Calculate the number of elements. For a bag it counts the number of tuples. For a tuple it counts the number of elements. For a chararray it counts the number of characters. For a bytearray it counts the number of bytes. For numeric scalars it always returns 1. SUM Calculate the sum of numeric values in a single-column bag. TOKENIZE Split a string (chararray) into a bag of words (each word is a tuple in the bag). Word separators are space, double quote ("), comma, parentheses, and asterisk (*). IsEmpty Check if a bag or map is empty. You can’t use expressions and functions alone. You must use them within relational operators to transform data. 10.5.3 Relational operators The most salient characteristic about Pig Latin as a language is its relational opera- tors. These operators define Pig Latin as a data processing language. We’ll quickly go over the more straightforward operators first, to acclimate ourselves to their style and syntax. Afterward we’ll go into more details on the more complex operators such as COGROUP and FOREACH. UNION combines multiple relations together whereas SPLIT partitions a relation into multiple ones. An example will make it clear: grunt> a = load 'A' using PigStorage(',') as (a1:int, a2:int, a3:int); grunt> b = load 'B' using PigStorage(',') as (b1:int, b2:int, b3:int); grunt> DUMP a; (0,1,2) (1,3,4) grunt> DUMP b; (0,5,2) (1,7,8) grunt> c = UNION a, b; grunt> DUMP c; 226 CHAPTER 10 Programming with Pig (0,1,2) (0,5,2) (1,3,4) (1,7,8) grunt> SPLIT c INTO d IF $0 == 0, e IF $0 == 1; grunt> DUMP d; (0,1,2) (0,5,2) grunt> DUMP e; (1,3,4) (1,7,8) The UNION operator allows duplicates. You can use the DISTINCT operator to remove duplicates from a relation. Our SPLIT operation on c sends a tuple to d if its first field ($0) is 0, and to e if it’s 1. It’s possible to write conditions such that some rows will go to both d and e or to neither. You can simulate SPLIT by multiple FILTER operators. The FILTER operator alone trims a relation down to only tuples that pass a certain test: grunt> f = FILTER c BY $1 > 3; grunt> DUMP f; (0,5,2) (1,7,8) We’ve seen LIMIT being used to take a specified number of tuples from a relation. SAMPLE is an operator that randomly samples tuples in a relation according to a speci- fied percentage. The operations ‘till now are relatively simple in the sense that they operate on each tuple as an atomic unit. More complex data processing, on the other hand, will require working on groups of tuples together. We’ll next look at operators for grouping. Unlike previous operators, these grouping operators will create new schemas in their output that rely heavily on bags and nested data types. The generated schema may take a little time to get used to at first. Keep in mind that these grouping operators are almost always for generating intermediate data. Their complexity is only temporary on your way to computing the final results. The simpler of these operators is GROUP. Continuing with the same set of relations we used earlier, grunt> g = GROUP c BY $2; grunt> DUMP g; (2,{(0,1,2),(0,5,2)}) (4,{(1,3,4)}) (8,{(1,7,8)}) grunt> DESCRIBE c; c: {a1: int,a2: int,a3: int} grunt> DESCRIBE g; g: {group: int,c: {a1: int,a2: int,a3: int}} We’ve created a new relation, g, from grouping tuples in c having the same value on the third column ($2, also named a3). The output of GROUP always has two fields. The first field is group key, which is a3 in this case. The second field is a bag containing Speaking Pig Latin 227 all the tuples with the same group key. Looking at g’s dump, we see that it has three tuples, corresponding to the three unique values in c’s third column. The bag in the first tuple represents all tuples in c with the third column equal to 2. The bag in the second tuple represents all tuples in c with the third column equal to 4. And so forth. After you understand how g’s data came about, you’ll feel more comfortable looking at its schema. The first field of GROUP’s output relation is always named group, for the group key. In this case it may seem more natural to call the first field a3, but currently Pig doesn’t allow you to assign a name to replace group. You’ll have to adapt yourself to refer to it as group. The second field of GROUP’s output relation is always named af- ter the relation it’s operating on, which is c in this case, and as we said earlier it’s always a bag. As we use this bag to hold tuples from c, the schema for this bag is exactly the schema for c —three fields of integers named a1, a2, and a3. Before moving on, we want to note that one can GROUP by any function or expression. For example, if time is a timestamp and there exists a function DayOfWeek , one can conceivably do this grouping that would create a relation with seven tuples. GROUP log BY DayOfWeek(time); Finally, one can put all tuples in a relation into one big bag. This is useful for aggregate analysis on relations, as functions work on bags but not relations. For example: grunt> h = GROUP c ALL; grunt> DUMP h; (all,{(0,1,2),(0,5,2),(1,3,4),(1,7,8)}) grunt> i = FOREACH h GENERATE COUNT($1); grunt> dump i; (4L) This is one way to count the number of tuples in c. The first field in GROUP ALL’s out- put is always the string all. Now that you’re comfortable with GROUP, we can look at COGROUP , which groups together tuples from multiple relations . It functions much like a join. For example, let’s cogroup a and b on the third column. grunt> j = COGROUP a BY $2, b BY $2; grunt> DUMP j; (2,{(0,1,2)},{(0,5,2)}) (4,{(1,3,4)},{}) (8,{},{(1,7,8)}) grunt> DESCRIBE j; j: {group: int,a: {a1: int,a2: int,a3: int},b: {b1: int,b2: int,b3: int}} Whereas GROUP always generates two fields in its output, COGROUP always generates three (more if cogrouping more than two relations). The first field is the group key, whereas the second and third fields are bags. These bags hold tuples from the co- grouping relations that match the grouping key. If a grouping key matches only tuples from one relation but not the other, then the field corresponding to the nonmatching relation will have an empty bag. To ignore group keys that don’t exist for a relation, one can add the INNER keyword to the operation, like 228 CHAPTER 10 Programming with Pig grunt> j = COGROUP a BY $2, b BY $2 INNER; grunt> dump j; (2,{(0,1,2)},{(0,5,2)}) (8,{},{(1,7,8)}) grunt> j = COGROUP a BY $2 INNER, b BY $2 INNER; grunt> dump j; (2,{(0,1,2)},{(0,5,2)}) Conceptually, you can think of the default behavior of COGROUP as an outer join, and the INNER keyword can modify it to be left outer join, right outer join, or inner join. Another way to do inner join in Pig is to use the JOIN operator. The main difference between JOIN and an inner COGROUP is that JOIN creates a flat set of output records, as indicated by looking at the schema: grunt> j = JOIN a BY $2, b BY $2; grunt> dump j; (0,1,2,0,5,2) grunt> DESCRIBE j; j: {a::a1: int,a::a2: int,a::a3: int,b::b1: int,b::b2: int,b::b3: int} The last relational operator we look at is FOREACH. It goes through all tuples in a relation and generates new tuples in the output. Behind this seeming simplicity lies tremendous power though, particularly when it’s applied to complex data types outputted by the grouping operators. There’s even a nested form of FOREACH designed for handling com- plex types. First let’s familiarize ourselves with different FOREACH operations on simple relations. grunt> k = FOREACH c GENERATE a2, a2 * a3; grunt> DUMP k; (1,2) (5,10) (3,12) (7,56) FOREACH is always followed by an alias (name given to a relation) followed by the key- word GENERATE. The expressions after GENERATE control the output. At its simplest, we use FOREACH to project specific columns of a relation into the output. We can also apply arbitrary expressions, such as multiplication in the preceding example. For relations with nested bags (e.g., ones generated by the grouping operations), FOREACH has special projection syntax, and a richer set of functions. For example, applying nested projection to have each bag retain only the first field: grunt> k = FOREACH g GENERATE group, c.a1; grunt> DUMP k; (2,{(0),(0)}) (4,{(1)}) (8,{(1)}) To get two fields in each bag: grunt> k = FOREACH g GENERATE group, c.(a1,a2); grunt> DUMP k; (2,{(0,1),(0,5)}) Speaking Pig Latin 229 (4,{(1,3)}) (8,{(1,7)}) Most built-in Pig functions are geared toward working on bags. grunt> k = FOREACH g GENERATE group, COUNT(c); grunt> DUMP k; (2,2L) (4,1L) (8,1L) Recall that g is based on grouping c on the third column. This FOREACH statement therefore generates a frequency count of the values in c’s third column. As we said earlier, grouping operators are mainly for generating intermediate data that will be simplified by other operators such as FOREACH. The COUNT function is one of the ag- gregate functions. As we’ll see, you can create many other functions via UDFs. The FLATTEN function is designed to flatten nested data types. Syntactically it looks like a function, such as COUNT and AVG, but it’s a special operator as it can change the structure of the output created by FOREACH...GENERATE. Its flattening behavior is also different depending on how it’s applied and what it’s applied to. For example, consider a relation with tuples of the form (a, (b, c)). The statement FOREACH... GENERATE $0, FLATTEN($1) will create one output tuple of the form (a, b, c) for each input tuple. When applied to bags, FLATTEN modifies the FOREACH...GENERATE statement to generate new tuples. It removes one layer of nesting and behaves almost the opposite of grouping operations. If a bag contains N tuples, flattening it will remove the bag and create N tuples in its place. grunt> k = FOREACH g GENERATE group, FLATTEN(c); grunt> DUMP k; (2,0,1,2) (2,0,5,2) (4,1,3,4) (8,1,7,8) grunt> DESCRIBE k; k: {group: int,c::a1: int,c::a2: int,c::a3: int} Another way to understand FLATTEN is to see that it produces a cross-product. This view is helpful when we use FLATTEN multiple times within a single FOREACH statement. For example, let’s say we’ve somehow created a relation l. grunt> dump l; (1,{(1,2)},{(3)}) (4,{(4,2),(4,3)},{(6),(9)}) (8,{(8,3),(8,4)},{(9)}) grunt> describe l; d: {group: int,a: {a1: int,a2: int},b: {b1: int}} The following statement that flattens two bags outputs all combinations of those two bags for each tuple: grunt> m = FOREACH l GENERATE group, FLATTEN(a), FLATTEN(b); grunt> dump m; 230 CHAPTER 10 Programming with Pig (1,1,2,3) (4,4,2,6) (4,4,2,9) (4,4,3,6) (4,4,3,9) (8,8,3,9) (8,8,4,9) We see that the tuple with group key 4 in relation l has a bag of size 2 in field a and also a bag size 2 in field b. The corresponding output in m therefore has four rows rep- resenting the full cross-product. Finally, there’s a nested form of FOREACH to allow for more complex processing of bags. Let’s assume you have a relation (say l) and one of its fields (say a) is a bag, a FOREACH with nested block has this form: alias = FOREACH l { tmp1 = operation on a; [more operations...] GENERATE expr [, expr...] } The GENERATE statement must always be present at the end of the nested block. It will create some output for each tuple in l. The operations in the nested block can create new relations based on the bag a. For example, we can trim down the a bag in each element of l’s tuple. grunt> m = FOREACH l { tmp1 = FILTER a BY a1 >= a2; GENERATE group, tmp1, b; } grunt> DUMP m; (1,{},{(3)}) (4,{(4,2),(4,3)},{(6),(9)}) (8,{(8,3),(8,4)},{(9)}) You can have multiple statements in the nested block. Each one can even be operating on different bags. grunt> m = FOREACH l { tmp1 = FILTER a BY a1 >= a2; tmp2 = FILTER b by $0 < 7; GENERATE group, tmp1, tmp2; }; grunt> DUMP m; (1,{},{(3)}) (4,{(4,2),(4,3)},{(6)}) (8,{(8,3),(8,4)},{}) As of this writing, only five operators are allowed in the nested block: DISTINCT, FILTER, LIMIT, ORDER, and SAMPLE. It’s expected that more will be supported in the future. NOTE Sometimes the output of FOREACH can have a completely different schema from its input. In those cases, users can specify the output schema using Doesn’t require the relations to have the same schema or even the same Speaking Pig Latin 231 the AS option after each field. This syntax differs from the LOAD command where the schema is specified as a list after the AS option, but in both cases we use AS to specify a schema. Table 10.8 summarizes the relational operators in Pig Latin. On many operators you’ll see an option for PARALLEL n . The number n is the degree of parallelism you want for executing that operator. In practice n is the number of reduce tasks in Hadoop that Pig will use. If you don’t set n it’ll default to the default setting of your Hadoop cluster. Pig documentation recommends setting the value of n according to the fol- lowing guideline: n = (#nodes - 1) * 0.45 * RAM where #nodes is the number of nodes and RAM is the amount of memory in GB on each node. Table 10.8 Relational operators in Pig Latin SPLIT SPLIT alias INTO alias IF expression, alias IF expression [, alias IF expression ...]; Splits a relation into two or more relations, based on the given Boolean expressions. Note that a tuple can be assigned to more than one relation, or to none at all. UNION alias = UNION alias, alias, [, alias ...] Creates the union of two or more relations. Note that ■ As with any relation, there’s no guarantee to the order of tuples ■ number of fields Doesn’t remove duplicate tuples■ FILTER alias = FILTER alias BY expression; Selects tuples based on Boolean expression. Used to select tuples that you want or remove tuples that you don’t want. DISTINCT alias = DISTINCT alias [PARALLEL n]; Remove duplicate tuples. SAMPLE alias = SAMPLE alias factor; Randomly sample a relation. The sampling factor is given in factor. For example, a 1% sample of data in relation large_data is small_data = SAMPLE large_data 0.01; The operation is probabilistic in such a way that the size of small_data will not be exactly 1% of large_data, and there’s no guarantee the operation will return the same number of tuples each time. FOREACH alias = FOREACH alias GENERATE expression [,expression ...] [AS schema]; Loop through each tuple and generate new tuple(s). Usually applied to transform columns of data, such as adding or deleting fields. One can optionally specify a schema for the output relation; for example, naming new fields. 232 CHAPTER 10 Programming with Pig Table 10.8 Relational operators in Pig Latin (continued) FOREACH (nested) alias = FOREACH nested_alias { alias = nested_op; [alias = nested_op; ...] GENERATE expression [, expression ...]; }; Loop through each tuple in nested_alias and generate new tuple(s). At least one of the fields of nested_alias should be a bag. DISTINCT, FILTER, LIMIT, ORDER, and SAMPLE are allowed operations in nested_op to operate on the inner bag(s). JOIN alias = JOIN alias BY field_alias, alias BY field_alias [, alias BY field_alias …] [USING "replicated"] [PARALLEL n]; Compute inner join of two or more relations based on common field values. When using the replicated option, Pig stores all relations after the first one in memory for faster processing. You have to ensure that all those smaller relations together are indeed small enough to fit in memory. Under JOIN, when the input relations are flat, the output relation is also flat. In addition, the number of fields in the output relation is the sum of the number of fields in the input relations, and the output relation’s schema is a concatenation of the input relations’ schemas. GROUP alias = GROUP alias { [ALL] | [BY {[field_alias [, field_alias]] | * | [expression]] } [PARALLEL n]; Within a single relation, group together tuples with the same group key. Usually the group key is one or more fields, but it can also be the entire tuple (*) or an expression. One can also use GROUP alias ALL to group all tuples into one group. The output relation has two fields with autogenerated names. The first field is always named “group” and it has the same type as the group key. The second field takes the name of the input relation and is a bag type. The schema for the bag is the same as the schema for the input relation. COGROUP alias = COGROUP alias BY field_alias [INNER | OUTER] , alias BY field_alias [INNER | OUTER] [PARALLEL n]; Group tuples from two or more relations, based on common group values. The output relation will have a tuple for each unique group value. Each tuple will have the group value as its first field. The second field is a bag containing tuples from the first input relation with matching group value. Ditto for the third field of the output tuple. In the default OUTER join semantic, all group values appearing in any input relation are represented in the output relation. If an input relation doesn’t have any tuple with a particular group value, it will have an empty bag in the corresponding output tuple. If the INNER option is set for a relation, then only group values that exist in that input relation are allowed in the output relation. There can’t be an empty bag for that relation in the output. You can group on multiple fields. For this, you have to specify the fields in a comma-separated list enclosed by parentheses for field_alias. COGROUP (with INNER) and JOIN are similar except that COGROUP generates nested output tuples. CROSS alias = CROSS alias, alias [, alias …] [PARALLEL n]; Compute the (flat) cross-product of two or more relations. This is an expensive operation and you should avoid it as far as possible. Working with user-defined functions 233 Table 10.8 Relational operators in Pig Latin (continued) ORDER alias = ORDER alias BY { * [ASC|DESC] | field_alias [ASC|DESC] [, field_alias [ASC|DESC] …] } [PARALLEL n]; Sort a relation based on one or more fields. If you retrieve the relation right after the ORDER operation (by DUMP or STORE), it’s guaranteed to be in the desired sorted order. Further processing (FILTER, DISTINCT, etc.) may destroy the ordering. STREAM alias = STREAM alias [, alias …] THROUGH {'command' | cmd_alias } [AS schema] ; Process a relation with an external script. At this point you’ve learned various aspects of the Pig Latin language—data types, expressions, functions, and relational operators. You can extend the language further with user-defined functions. But before discussing that we’ll end this section with a note on Pig Latin compilation and optimization. 10.5.4 Execution optimization As with many modern compilers, the Pig compiler can reorder the execution sequence to optimize performance, as long as the execution plan remains logically equivalent to the original program. For example, imagine a program that applies an expensive func- tion (say, encryption) to a certain field (say, social security number) of every record, followed by a filtering function to select records based on a different field (say, limit only to people within a certain geography). The compiler can reverse the execution order of those two operations without affecting the final result, yet performance is much improved. Having the filtering step first can dramatically reduce the amount of data and work the encryption step will have to do. As Pig matures, more optimization will be added to the compiler. Therefore it’s important to try to always use the latest version. But there’s always a limit to a compiler’s ability to optimize arbitrary code. You can read Pig’s web documentation for techniques to improve performance. A list of tips for enhancing performance under Pig version 0.3 is at http://hadoop.apache.org/pig/docs/ r0.3.0/cookbook.html. 10.6 Working with user-defined functions Fundamental to Pig Latin’s design philosophy is its extensibility through user-defined functions (UDFs ), and there’s a well-defined set of APIs for writing UDFs. This doesn’t mean that you’ll have to write all the functions you need yourself. Part of Pig’s ecosys- tem4 is PiggyBank, 5 an online repository for users to share their functions. You should 4 I thought about calling it a Pig pen, but PigPen is actually the name of an Eclipse plug-in for editing Pig Latin scripts. See http://wiki.apache.org/pig/PigPen. 5 http://wiki.apache.org/pig/PiggyBank. 234 CHAPTER 10 Programming with Pig check PiggyBank first for any function you need. Only if you don’t find an appropriate function should you consider writing your own. You should also consider contributing your UDF back to PiggyBank to benefit others in the Pig community. 10.6.1 Using UDFs As of this writing UDFs are always written in Java and packaged in jar files. To use a particular UDF you’ll need the jar file containing the UDF’s class file(s). For example, when using functions from PiggyBank you’ll most likely obtain a piggy- bank.jar file. To use a UDF, you must first register the jar file with Pig using the REGISTER statement . Afterward, you invoke the UDF by its fully qualified Java class name. For example, there’s an UPPER function in PiggyBank that transforms a string to uppercase: REGISTER piggybank/java/piggybank.jar; b = FOREACH a GENERATE ➥ org.apache.pig.piggybank.evaluation.string.UPPER($0); If you need to use a function multiple times, it’ll get annoying to write out the fully qualified class name every time. Pig offers the DEFINE statement to assign a name to a UDF. You can rewrite the above statements to REGISTER piggybank/java/piggybank.jar; DEFINE Upper org.apache.pig.piggybank.evaluation.string.UPPER(); b = FOREACH a GENERATE Upper($0); Table 10.9 summarizes the UDF-related statements. Table 10.9 UDF statements in Pig Latin DEFINE DEFINE alias { function | 'command' [...] }; Assign an alias to a function or command. REGISTER REGISTER alias; Register UDFs with Pig. Currently UDFs are only written in Java, and alias is the path of the JAR file. All UDFs must be registered before they can be used. If you’re only using UDFs written by other people, this is all you need to know. But if you can’t find the UDF you need, you’ll have to write your own. 10.6.2 Writing UDFs Pig supports two main categories of UDFs: eval 6 and load/store . We use the load/ store functions only in LOAD and STORE statements to help Pig read and write special formats. Most UDFs are eval functions that take one field value and return another field value. 6 Some eval functions are quite common and have special considerations. They’re sometimes described in their own categories. These include filter functions (eval functions that return a Boolean) and aggregate functions (eval functions that take a bag and return a scalar value). Working with user-defined functions 235 As of this writing, you can only write a UDF using Pig’s Java API. 7 To create an eval UDF you make a Java class that extends the abstract EvalFunc class. It has only one abstract method which you need to implement: abstract public T exec(Tuple input) throws IOException; This method is called on each tuple in a relation, where each tuple is represented by a Tuple object. The exec() method processes the tuple and returns a type T corre- sponding to a valid Pig Latin type. T can be any one of the Java classes in table 10.10, some of which are native Java classes and some of which are Pig extensions. Table 10.10 Pig Latin types and their equivalent classes in Java. Pig Latin type Java class Bytearray DataByteArray Chararray String Int Integer Long Long Float Float Double Double Tuple Tuple Bag DataBag Map Map The best way to learn about writing UDFs is to dissect one of the existing UDFs in PiggyBank. Even when writing your own, it’s often useful to start with a working UDF that’s functionally similar to what you want and only modify the processing logic. For our purpose, let’s explore the UPPER UDF we used earlier from PiggyBank. The exec() method looks like this: public class UPPER extends EvalFunc { public String exec(Tuple input) throws IOException { if (input == null || input.size() == 0) return null; try { String str = (String)input.get(0); return str.toUpperCase(); } catch(Exception e){ System.err.println("Failed to process input; error - " + ➥ e.getMessage()); return null; } } } 7 The Javadoc for the API is at http://hadoop.apache.org/pig/javadoc/docs/api/. 236 CHAPTER 10 Programming with Pig The object input belongs to the Tuple class, which has two methods for retrieving its content. List getAll(); Object get(int fieldNum) throws ExecException; The getAll() method return all fields in the tuple as an ordered list. UPPER instead uses the get() method to request for a specific field (at position 0). This method would throw an ExecException if the requested field number is greater than the num- ber of fields in the tuple. In UPPER the retrieved field is casted to a Java String, which usually works but may cause a cast exception if we were casting between incompatible data types. We’ll see later how to use Pig to ensure that our casting works. In any case, the try/catch block would’ve caught and handled any exception. If everything works, UPPER’s exec() method will return a String with characters uppercased. In addition, most UDFs should implement the default behavior that the output is null when the input tuple is null. In addition to implementing exec(), UPPER also overrides a couple methods from EvalFunc, one of which is getArgToFuncMapping: @Override public List getArgToFuncMapping() throws FrontendException { List funcList = new ArrayList(); funcList.add(new FuncSpec(this.getClass().getName(), ➥ new Schema(new Schema.FieldSchema(null, DataType.CHARARRAY)))); return funcList; } The getArgToFuncMapping() method returns a List of FuncSpec objects repre- senting the schema of each field in the input tuple. Pig will handle typecasting for you by converting the types of all fields in a tuple to conform to this schema before passing it to exec(). It will pass fields that can’t be converted to the desired type as null. UPPER only cares about the type of the first field, so it adds only one FuncSpec to the list, and this FuncSpec states that the field must be of type chararray, represented as DataType.CHARARRAY. The instantiation of FuncSpec is quite convoluted, which is due to Pig’s ability to handle complex nested types. Fortunately, unless you work with unusually complicated types, you’ll probably find a FuncSpec instantiation for the type you want already in one of PiggyBank ’s UDFs. Reuse that in your code. You can even reuse the entire getArgToFuncMapping() function if you have the same tuple schema as another UDF. Besides telling Pig the input schema, you can also tell Pig the schema of your output. You may not need to do this if the output of your UDF is a simple scalar , as Pig will use Java’s Reflection mechanism to infer the schema automatically. But if your UDF returns a tuple or a bag, the Reflection mechanism will fail to figure out the schema completely. In that case you should specify it so that Pig can propagate the schema correctly. Working with scripts 237 In UPPER’s case it only outputs a simple String, so it’s not necessary to specify the output schema. But UPPER does do this by overriding outputSchema() to tell Pig that it’s returning a string (DataType.CHARARRAY). @Override public Schema outputSchema(Schema input) { return new Schema( new Schema.FieldSchema( getSchemaName(this.getClass().getName().toLowerCase(), input), DataType.CHARARRAY ) ); } Again, the Schema object construction looks convoluted because of Pig’s ability to have complex nested types. One special case is if the schema of your UDF’s output is the same as the input. We can return a copy of the input schema: public Schema outputSchema(Schema input) { return new Schema(input); } As with the construction of FuncSpec, you’ll probably find some preexisting UDFs in PiggyBank with your desired output schema. A few types of UDFs call for special considerations. Filter functions are eval functions that return a Boolean, and we use them in Pig Latin’s FILTER and SPLIT statements. They should extend FilterFunc instead of EvalFunc. Aggregate functions are eval functions that take in a bag and return a scalar. They’re usually used for computing aggregate metrics, such as COUNT, and we can sometimes optimize them in Hadoop by using a combiner. We haven’t covered the load/save UDFs for reading and writing data sets. These more advanced topics are covered in Pig’s documentation on UDFs: http://hadoop.apache.org/pig/docs/r0.3.0/udf.html. 10.7 Working with scripts Writing Pig Latin scripts is largely about packaging together the Pig Latin statements that you’ve successfully tested in Grunt. Pig scripting does have a few unique topics though. They’re comments, parameter substitution, and multiquery execution. 10.7.1 Comments As you’ll reuse your Pig Latin script, it’s obviously a good idea to leave comments for other people (or yourself) to understand it in the future. Pig Latin supports two forms of comments, single-line and multiline. You start the single-line comment by a double hyphen and the comment ends at the end of the line. You enclose the multiline com- ment by the /* and */ markers, similar to multiline comments in Java. For example, a Pig Latin script with comments can look like /* * Myscript.pig * Another line of comment 238 CHAPTER 10 Programming with Pig */ log = LOAD 'excite-small.log' AS (user, time, query); lmt = LIMIT log 4; -- Only show 4 tuples DUMP lmt; -- End of program 10.7.2 Parameter substitution When you write a reusable script, it’s generally parameterized such that you can vary its operation for each run. For example, the script may take the file paths of its input and output from the user each time. Pig supports parameter substitution to allow the user to specify such information at runtime. It denotes such parameters by the $ prefix within the script. For example, the following script displays a user-specified number of tuples from a user-specified log file: log = LOAD '$input' AS (user, time, query); lmt = LIMIT log $size; DUMP lmt; The parameters in this script are $input and $size. If you run this script using the pig command, you specify the parameters using the -param name=value argument. pig -param input=excite-small.log -param size=4 Myscript.pig Note that you don’t need the $ prefix in the arguments. You can enclose a param- eter value in single or double quotes, if it has multiple words. A useful technique is to use Unix commands to generate the parameter values, particularly for dates. This is accomplished through Unix’s command substitution, which executes commands enclosed in back ticks (`). pig -param input=web-'date +%y-%m-%d'.log -param size=4 Myscript.pig By doing this, the input file for Myscript.pig will be based on the date the script is run. For example, the input file will be web-09-07-29.log if the script is run on July 29, 2009. If you have to specify many parameters, it may be more convenient to put them in a file and tell Pig to execute the script using parameter substitution based on that file. For example, we can create a file Myparams.txt with the following content: # Comments in a parameter file start with hash input=excite-small.log size=4 The parameter file is passed to the pig command with the -param_file filename argument. pig -param_file Myparams.txt Myscript.pig You can specify multiple parameter files as well as mix parameter files with direct specifi- cation of parameters at the command line using -param. If you define a parameter mul- tiple times, the last definition takes precedence. When in doubt about what parameter values a script ends up using, you can run the pig command with the -debug option. Working with scripts 239 This tells Pig to run the script and also output a file named original_script_name. substituted that has the original script but with all the parameters fully substituted. Exe- cuting pig with the -dryrun option outputs the same file but doesn’t execute the script. The exec and run commands allow you to run Pig Latin scripts from within the Grunt shell, and they support parameter substitution using the same -param and -param_file arguments; for example: grunt> exec -param input=excite-small.log -param size=4 Myscript.pig However, parameter substitution in exec and run doesn’t support Unix commands, and there’s no debug or dryrun option. 10.7.3 Multiquery execution In the Grunt shell, a DUMP or STORE operation processes all previous statements need- ed for the result. On the other hand, Pig optimizes and processes an entire Pig script as a whole. This difference would have no effect at all if your script has only one DUMP or STORE command at the end. If your script has multiple DUMP/STORE, Pig script’s mul- tiquery execution improves efficiency by avoiding redundant evaluations. For example, let’s say you have a script that stores intermediate data: a = LOAD ... b = some transformation of a STORE b ... c = some further transformation of b STORE c ... If you enter the statements in Grunt, where there’s no multiquery execution, it will generate a chain of jobs on the STORE b command to compute b. On encountering STORE c, Grunt will run another chain of jobs to compute c, but this time it will evalu- ate both a and b again! You can manually avoid this reevaluation by inserting a b = LOAD ... statement right after STORE b, to force the computation of c to use the saved value of b. This works on the assumption that the stored value of b has not been modi- fied, because Grunt, by itself, has no way of knowing. On the other hand, if you run all the statements as a script, multiquery execution can optimize the execution by intelligently handling intermediate data. Pig compiles all the statements together and can locate the dependency and redundancy. Multiquery execution is enabled by default and usually has no effect on the computed results. But multiquery execution can fail if there are data dependencies that Pig is not aware of. This is quite rare but can happen with, for example, UDFs. Consider this script: STORE a INTO 'out1'; b = LOAD ... c = FOREACH b GENERATE MYUDF($0,'out1'); STORE c INTO 'out2'; If the custom function MYUDF is such that it accesses a through the file out1, the Pig compiler would have no way of knowing that. Not seeing the dependency, the Pig com- piler may erroneously think it OK to evaluate b and c before evaluating a. To disable multiquery execution, run the pig command with -M or -no_multiquery option. 240 CHAPTER 10 Programming with Pig 10.8 Seeing Pig in action—example of computing similar patents Given the extra power that Pig provides, we can take on more challenging data process- ing applications. One interesting application from the patent data set is finding similar patents based on citation data. Patents that are often cited together must be similar (or at least related) in some way. This application has the essence of the Amazon.com style collaborative filtering (“Customers who have bought this have also bought that.”) and finding similar documents (by looking for documents with a similar set of words). For our purpose here, let’s suppose we want to look into patents that are cited together more than N times, where N is a fixed number we specify.8 For applications that involve pair-wise computations (e.g., computing number of cocitations for each pair of patents), it’s often easy to imagine an implementation involving a pair of nested loops enumerating all pair combinations and performing the computation on each pair. Even though Hadoop makes it easy to scale by adding more hardware, we should continue to remember fundamental concepts in computational complexity. Quadratic complexity will still bring linear scalability to its knees. Even a small data set of 3 million patents can lead to 9 trillion pairs. We need smarter algorithms. The main insight to leverage is that the resulting data is sparse . Most pairs will have zero similarity as most pairs of patents are never cited together. Our similarity computation will become much more manageable if we redesign it to only work on patent pairs that are known to have been cited together. Looking at our data, this approach is quite natural. This implementation involves these steps for each patent: 1 Get the list of patents it cites 2 Generate all pair-wise combinations of the list and record each pair 3 Count how many of each pair we have If each patent cites a fixed number of patents, say 10, this implementation would gen- erate 45 pairs for each patent. (45 is the number of pair combinations possible from 10 items, which mathematically is derived as 10 x 9 / 2.) With 3 million patents this creates 135 million pairs, which is orders of magnitude smaller than the brute force ap- proach. This advantage would be even more apparent if the patent data set is larger. Even though we’ve figured out the algorithm for this application, implementing it in MapReduce can still be tedious. It’ll require chaining multiple jobs together, and each job will require its own class. Pig Latin, on the other hand, takes only a dozen lines to implement the three-step program (listing 10.1), and further optimization can eliminate more lines and increase efficiency still. Listing 10.1 Pig Latin script to find patents that are often cited together cite = LOAD 'input/cite75_99.txt' USING PigStorage(',') ➥ AS (citing:int, cited:int); cite_grpd = GROUP cite BY citing; 8 Variations of this may involve more advanced scoring functions, such as normalizing for frequent items, or computing a similarity ranking rather than a simple cutoff. The simple cutoff criterion we chose here is easier to implement and illustrates the essence of computing similarity. Seeing Pig in action—example of computing similar patents 241 cite_grpd_dbl = FOREACH cite_grpd GENERATE group, cite.cited AS cited1, ➥ cite.cited AS cited2; cocite = FOREACH cite_grpd_dbl ➥ GENERATE FLATTEN(cited1), FLATTEN(cited2); cocite_fltrd = FILTER cocite BY cited1 != cited2; cocite_grpd = GROUP cocite_fltrd BY *; cocite_cnt = FOREACH cocite_grpd ➥ GENERATE group, COUNT(cocite_fltrd) as cnt; cocite_flat = FOREACH cocite_cnt GENERATE FLATTEN(group), cnt; cocite_cnt_grpd = GROUP cocite_flat BY cited1; cocite_bag = FOREACH cocite_cnt_grpd ➥ GENERATE group, cocite_flat.(cited2, cnt); cocite_final = FOREACH cocite_cnt_grpd { similar = FILTER cocite_flat BY cnt > 5; GENERATE group, similar; } STORE cocite_final INTO 'output'; Pig Latin, and probably complex data processing in general, can be hard to read. For- tunately, we can use Grunt’s ILLUSTRATE command on cocite_bag to get a simulated sample run of the statements and see what each operation is generating. (We’ve refor- matted the output to fit the width of the printed page.) --------------------------------------------------- | cite | citing: bytearray | cited: bytearray | --------------------------------------------------- | | 3858554 | 3601095 | | | 3858554 | 3685034 | | | 3859004 | 1730866 | | | 3859004 | 3022581 | | | 3859572 | 3206651 | --------------------------------------------------- --------------------------------------- | cite | citing: int | cited: int | --------------------------------------- | | 3858554 | 3601095 | | | 3858554 | 3685034 | | | 3859004 | 1730866 | | | 3859004 | 3022581 | | | 3859572 | 3206651 | --------------------------------------- ------------------------------------------------------------------------- | cite_grpd | group: int | cite: bag({citing: int,cited: int}) | ------------------------------------------------------------------------- | | 3858554 | {(3858554, 3601095), (3858554, 3685034)} | | | 3859004 | {(3859004, 1730866), (3859004, 3022581)} | | | 3859572 | {(3859572, 3206651)} | ------------------------------------------------------------------------- --------------------------------------------------------------------------- | cite_grpd_dbl | group: | cited1: | cited2: | | | int | bag({cited: int}) | bag({cited: int})| --------------------------------------------------------------------------- | | 3858554 | {(3601095), (3685034)} | {(3601095),(3685034)}| | | 3859004 | {(1730866), (3022581)} | {(1730866),(3022581)}| | | 3859572 | {(3206651)} | {(3206651)} | --------------------------------------------------------------------------- 242 CHAPTER 10 Programming with Pig The relation cite_grpd contains a bag for each patent, and in this bag are the cited patents. From this relation (in this example run), we can see that patents 3601095 and 3685034 are cited together in patent 3858554. Grouping cocited patents was done by the GROUP operation in creating cite_grpd . The relation cite_grpd_dbl only re- moves the redundant “citing” patent and creates a duplicate column. The columns cited1 and cited2 have the same values. This duplication will allow the cross-product operation to generate all pair-wise combinations . -------------------------------------------------------- | cocite | cited1::cited: int | cited2::cited: int | -------------------------------------------------------- | | 3601095 | 3601095 | | | 3601095 | 3685034 | | | 3685034 | 3601095 | | | 3685034 | 3685034 | | | 1730866 | 1730866 | | | 1730866 | 3022581 | | | 3022581 | 1730866 | | | 3022581 | 3022581 | | | 3206651 | 3206651 | -------------------------------------------------------- The cross-product from flattening each row of cite_grpd_dbl creates cocite.9 This is the record of all pairs of patents that have been cited together and is a major check- point for our algorithm. We know that cocite is a big relation, even under our scheme which is more efficient than brute force. There are three ways to trim down cocite further. We’ll discuss them all but implement only one. The first potential reduction is to notice that each cited patent is considered to have been cocited with itself. As we know that it’s quite pointless for our application to figure out that a patent is similar to itself, we can ignore all such pairs. Note that if we keep these “identity” pairs in the calculation, the cocitation count for them will end up being exactly the citation count. These numbers can still be useful if we’re looking for the percentage of times patents are cocited. As we’re not computing percentages, that consideration wouldn’t affect us. As cocitation is symmetric , pairs always appear twice, in reverse order. For example, we see both (3601095,3685034) and (3685034,3601095) when they appear together once. Given our current application need to find patent pairs that are cocited more than N times together, we can put in a simple rule retaining only one of the two redundant pairs and trim cocite’s size by half. This rule can be thus: retain only pairs where the first field is smaller than the second field. But keeping the redundant pairs can be useful for lookup later in some applications. For example, we can find all patents cocited with X by searching for X in the first field. In the more condensed version we’d have to look for X in both fields. Finally, we can use heuristics to remove cocitation pairs that we don’t think are important. We compromise final precision to gain efficiency. The applicability 9 Note that cocite can be computed from cite_grpd directly by using a more complicated FOREACH statement, and you may choose to do it when you feel more comfortable reading Pig Latin. Seeing Pig in action—example of computing similar patents 243 and usefulness of heuristics will depend on the application semantics and the data distribution. In our case, a patent that cites many patents together will generate a quadratic number of rows in cocite. If we believe that such “verbose” patents don’t help us understand similar patent pairs, removing them can significantly reduce the size of data to process with little impact on final results. The benefit of this heuristic is much greater if we’re looking at reverse patent citation or text documents, where frequency of items are extremely skewed and quadratic expansion on a few popular items can dominate the amount of data processed. In fact, in such situations approximate heuristics are almost necessary. An important process check is to note that we’ve focused on a higher level of data processing issues. We’ve obviated any low-level discussion about MapReduce. -------------------------------------------------------------- | cocite_fltrd | cited1::cited: int | cited2::cited: int | -------------------------------------------------------------- | | 3601095 | 3685034 | | | 3685034 | 3601095 | | | 1730866 | 3022581 | | | 3022581 | 1730866 | -------------------------------------------------------------- We’ve decided to only filter out “identity” patent pairs : --------------------------------------------------------------------------- | cocite_grpd | group: | cocite_fltrd: | | | tuple({cited1::cited: int, | bag({cited1::cited: int, | | | cited2::cited: int}) | cited2::cited: int}) | --------------------------------------------------------------------------- | | (1730866, 3022581) | {(1730866, 3022581)} | | | (3022581, 1730866) | {(3022581, 1730866)} | | | (3601095, 3685034) | {(3601095, 3685034)} | | | (3685034, 3601095) | {(3685034, 3601095)} | --------------------------------------------------------------------------- --------------------------------------------------------------------------- | cocite_cnt | group: | cnt: | | | tuple({cited1::cited: int,cited2::cited: int}) | long | --------------------------------------------------------------------------- | | (1730866, 3022581) | 1 | | | (3022581, 1730866) | 1 | | | (3601095, 3685034) | 1 | | | (3685034, 3601095) | 1 | --------------------------------------------------------------------------- --------------------------------------------------------------------------- | cocite_flat | group::cited1::cited: | group::cited2::cited: | cnt: | | | int | int | long | --------------------------------------------------------------------------- | | 1730866 | 3022581 | 1 | | | 3022581 | 1730866 | 1 | | | 3601095 | 3685034 | 1 | | | 3685034 | 3601095 | 1 | --------------------------------------------------------------------------- 244 CHAPTER 10 Programming with Pig We grouped the patent pair citations together, counted them, and flattened out the relation. Unfortunately, ILLUSTRATE generates sample data that only has coci- tation counts of 1. However, we see that the operations are doing basically what we wanted. If we stick to the original application requirement of only looking for patent pairs that have been cocited more than N times, we would apply a filter on cocite_flat and be finished. But we want to show how we can further group the tuples, which would be needed for other types of filtering. For example, you may want to find the K most cocited patents for each patent. Let’s look at the rest of the output: --------------------------------------------------------------------------- | cocite_cnt_grpd | group: int | cocite_flat: bag({group::cited1::cited: | | | | int,group::cited2::cited: int,cnt: long})| --------------------------------------------------------------------------- | | 1730866 | {(1730866, 3022581, 1)} | | | 3022581 | {(3022581, 1730866, 1)} | | | 3601095 | {(3601095, 3685034, 1)} | | | 3685034 | {(3685034, 3601095, 1)} | --------------------------------------------------------------------------- --------------------------------------------------------------------------- | cocite_bag | group: int | cocite_flat: | | | | bag({group::cited2::cited: int,cnt: long}) | --------------------------------------------------------------------------- | | 1730866 | {(3022581, 1)} | | | 3022581 | {(1730866, 1)} | | | 3601095 | {(3685034, 1)} | | | 3685034 | {(3601095, 1)} | --------------------------------------------------------------------------- If we had wanted to find each patent’s K most cocited patents, we would use a FOREACH statement to process each tuple in cocite_bag and write our own UDF to take in a bag (cocite_flat) and return a bag of at most K tuples (the most cocited ones). You can do this final step as an exercise. Let’s see an example of a nested FOREACH statement to filter out tuples inside bags that have counts of 5 or less. cocite_final = FOREACH cocite_cnt_grpd { similar = FILTER cocite_flat BY cnt > 5; GENERATE group, similar; } As you can see, Pig has simplified the implementation of this data processing ap- plication tremendously. This “similar item” feature has been known to be useful in different applications, but it’s also quite challenging to implement. Using Pig and Hadoop, this turns into only an afternoon’s work. Furthermore, its improved ease of development enables rapid prototyping of alternative features. For your own ex- ercise, instead of finding patents that are often cited together, can you find patents that have similar citations? Summary 245 10.9 Summary Pig is a higher-level data processing layer on top of Hadoop. Its Pig Latin language provides programmers a more intuitive way to specify data flows. It supports schemas in processing structured data, yet it’s flexible enough to work with unstructured text or semistructured XML data. It’s extensible with the use of UDFs. It vastly simplifies data joining and job chaining—two aspects of MapReduce programming that many developers found overly complicated. To demonstrate its usefulness, our example of computing patent cocitation shows a complex MapReduce program written in a dozen lines of Pig Latin. 11 246 Hive and the Hadoop herd This chapter covers ■ What Hive is ■ Setting up Hive ■ Other software packages related to Hadoop Using Hive for data warehousing ■ As powerful as Hadoop is, it doesn’t offer everything for everybody. Many projects have sprung up to extend Hadoop for specific purposes. The most prominent and well-supported ones have officially become subprojects under the umbrella of the Apache Hadoop project.1 These subprojects include ■ Pig—A high-level data flow language ■ Hive—A SQL-like data warehouse infrastructure ■ HBase—A distributed, column-oriented database modeled after Google’s Bigtable 1 What we’ve referred to in this book as “Hadoop” so far (HDFS and MapReduce) is technically called the “Hadoop Core” subproject of Apache Hadoop, although colloquially people tend to call it Hadoop. Hive 247 ■ ZooKeeper—A reliable coordination system for managing shared state between distributed applications ■ Chukwa—A data collection system for managing large distributed systems We covered Pig in detail in chapter 10, and we’ll learn about Hive in this chapter. Furthermore, section 11.2 will briefly describe other Hadoop-related projects. Some of these aren’t associated with Apache (e.g., Cascading, CloudBase). Some are in their nascent stages (e.g., Hama, Mahout). You’ll see some of these tools in action in the case studies of chapter 12. 11.1 Hive Hive2 is a data warehousing package built on top of Hadoop. It began its life at Facebook processing large amount of user and log data. It’s now a Hadoop subproject with many contributors. Its target users remain data analysts who are comfortable with SQL and who need to do ad hoc queries , summarization , and data analysis on Hadoop-scale data.3 You interact with Hive by issuing queries in a SQL-like language called HiveQL . For example, a query to get all active users from a user table looks like this: INSERT OVERWRITE TABLE user_active SELECT user.* FROM user WHERE user.active = 1; Hive’s design reflects its targeted use as a system for managing and querying structured data. By focusing on structured data , Hive can add certain optimization and usability features that MapReduce, being more general, doesn’t have. Hive’s SQL-inspired lan- guage separates the user from the complexity of MapReduce programming. It reuses familiar concepts from the relational database world, such as tables, rows, columns, and schema, to ease learning. In addition, whereas Hadoop naturally works on flat files, Hive can use directory structures to “partition” data to improve performance on certain queries. To support these additional features, a new and important component of Hive is a metastore for storing schema information. This metastore typically resides in a relational database. You can interact with Hive using several methods, including a Web GUI and Java Database Connectivity (JDBC) interface . Most interactions, though, tend to take place over a command line interface (CLI ), which is what we focus on. You can see a high-level architecture diagram of Hive in figure 11.1. 2 http://hadoop.apache.org/hive/. 3 Note that because Hive is built on top of Hadoop, it’s still designed for a low-latency, batch-oriented type of processing. It’s therefore not a direct replacement for traditional SQL data warehouses, such as ones offered by Oracle. 248 CHAPTER 11 Hive and the Hadoop herd JDBC/ ODBC Web GUICLI Parser Planner Optimizer Metastore QueriesDDL Hive Hadoop Figure 11.1 Hive architecture. Queries are parsed and executed on Hadoop. The metastore is an important component that helps determine how queries will be run. 11.1.1 Installing and configuring Hive Hive requires Java 1.6 and Hadoop version 0.17 or above. You can find the latest release of Hive at http://hadoop.apache.org/hive/releases.html. Download and extract the tarball into a directory that we call HIVE_HOME. Hadoop needs to be up and running already. In addition, you need to set up a couple directories in HDFS for Hive to use. bin/hadoop fs -mkdir /tmp bin/hadoop fs -mkdir /user/hive/warehouse bin/hadoop fs -chmod g+w /tmp bin/hadoop fs -chmod g+w /user/hive/warehouse If you let Hive manage your data completely for you, Hive will store your data under the /user/hive/warehouse directory. Hive can automatically add compression and special directory structures (such as partitions ) to those data to improve query perfor- mance. It’s good to let Hive manage your data if you plan on using Hive to query it. But if you already have your data in some other directories in HDFS and want to keep them there, Hive can work with them too. In that case, Hive will take your data as is and won’t try to optimize your data storage for query processing . Some casual users don’t understand this distinction, and believe that Hive requires data to be in some special Hive format. This is definitely not true. Hive 249 Hive stores metadata in a standard relational database. Out-of-the-box Hive comes with an open source, lightweight, embedded SQL database called Derby, 4 which is installed and run on the client machine along with Hive. If you are the only Hive user, this default setup should be fine. But beyond the initial testing and evaluation, you’ll most likely deploy Hive in a multi-user environment, where you wouldn’t want each user to have their own version of the metadata. You’ll need a centralized location for storing metadata. Typically, you use a shared SQL database such as MySQL, but any JDBC-compliant database will do. You’ll need a database server and you’ll need to create a database dedicated as a Hive metastore. This database is typically named metastore_db. Once you have created this, configure every Hive installation to point to it as the metastore. You configure the installations by modifying the files hive-site. xml and jpox.properties. Both are under the $HIVE_HOME/conf directory. A raw installation doesn’t have a hive-site.xml file, and you’ll have to create it. Properties in this file override the properties in hive-default.xml, in the same way that hadoop- site.xml overrides hadoop-default.xml. The file hive-site.xml should override three properties and look like the following: hive.metastore.local false javax.jdo.option.ConnectionURL jdbc:mysql:///metastore_db javax.jdo.option.ConnectionDriverName com.mysql.jdbc.Driver Table 11.1 explains these properties. Specify the javax.jdo.option.ConnectionURL and javax.jdo.option.ConnectionDriverName properties again in the file jpox. properties . In addition, the username and password to log into the database are also specified in jpox.properties. The jpox.properties file should contain the following four lines: javax.jdo.option.ConnectionDriverName=com.mysql.jdbc.Driver javax.jdo.option.ConnectionURL=jdbc:mysql:///metastore_db javax.jdo.option.ConnectionUserName= javax.jdo.option.ConnectionPassword= 4 http://db.apache.org/derby/. 250 CHAPTER 11 Hive and the Hadoop herd Table 11.1 Configuration for using a MySQL database as a metadata store in multi-user mode Property Description hive.metastore.local Controls whether to create and use a local metastore server in the client machine. Set this to false to use a remote metastore server. javax.jdo.option.ConnectionURL JDBC connection URL specifying the database for the metastore.5 For example, jdbc:mysql:/// metastore_db. javax.jdo.option. ConnectionDriverName The class name of the JDBC driver. For example, com.mysql.jdbc.Driver. javax.jdo.option.ConnectionUserName Username for logging into the database. javax.jdo.option.ConnectionPassword Password for logging into the database. Once you have the database set up, or if you’re only evaluating Hive and can use its default single-user mode, you’re ready to go into its CLI . Type in bin/hive in the $HIVE_ HOME directory. You’ll receive the Hive prompt, ready to take your Hive commands. bin/hive Hive history file=/tmp/root/hive_job_log_root_200908240830_797162695.txt hive> 11.1.2 Example queries Before we formally explain HiveQL, it’s useful to run a few commands from the CLI. You’ll get a feel of how it works and can explore on your own. Let’s assume you have the patent citation data cite75_99.txt on your local machine. Recall that this is a comma-separated data set of patent citations. In Hive, we first define a table that will store this data: hive> CREATE TABLE cite (citing INT, cited INT) > ROW FORMAT DELIMITED > FIELDS TERMINATED BY ',' > STORED AS TEXTFILE; OK Time taken: 0.246 seconds HiveQL statements are terminated by semicolons. You can have a statement that goes over multiple lines as long as you type in a semicolon only at the end, as we’ve done here. Most of the action in this four-line command is in the first line. Here we define a two- column table called cite. The first column is called citing and is of type INT , whereas the second column is called cited and is also of type INT. The other lines in this command 5 The complete format for the MySQL JDBC driver is described in http://dev.mysql.com/doc/refman/5.0/ en/connector-j-reference-configuration-properties.html. Hive 251 tell Hive how the data is stored (as a text file) and how it should be parsed (fields are separated by commas). We can see what tables are currently in Hive with the SHOW TABLES command: hive> SHOW TABLES; OK cite Time taken: 0.053 seconds Between Hive’s “OK” and “Time taken” messages, we see the cite table. We can check its schema with the DESCRIBE command : hive> DESCRIBE cite; OK citing int cited int Time taken: 0.13 seconds The table has the two columns from our definition, as expected. Managing and defin- ing tables in HiveQL are similar to standard relational databases. Let’s load the patent citation data into this table. hive> LOAD DATA LOCAL INPATH 'cite75_99.txt' > OVERWRITE INTO TABLE cite; Copying data from file:/root/cite75_99.txt Loading data to table cite OK Time taken: 9.51 seconds This tells Hive to load data from a file called cite75_99.txt in the local filesystem into our cite table . Underneath the hood, the local machine uploads this data into HDFS, under some directory managed by Hive. (Unless you’ve changed the configuration, this will be some directory under /user/hive/warehouse.) When loading data, Hive will not let any data into a table that violates its schema. In place of those data Hive will substitute a null. We can use a simple SELECT statement to browse data in the cite table: hive> SELECT * FROM cite LIMIT 10; OK NULL NULL 3858241 956203 3858241 1324234 3858241 3398406 3858241 3557384 3858241 3634889 3858242 1515701 3858242 3319261 3858242 3668705 3858242 3707004 Time taken: 0.17 seconds Our schema defines the two columns to be integers. We see that there’s a row with nulls , in- dicating that a record violated the schema. This is due to the first line of cite75_99.txt, which has the column names rather than patent numbers. Overall this shouldn’t be alarming. 252 CHAPTER 11 Hive and the Hadoop herd Now that we’re pretty confident that Hive has read the data and is managing it, we can run all kinds of queries on it. Let’s start by counting how many rows are in the table. In SQL this is accomplished by the familiar SELECT COUNT(*). HiveQL has a slightly different syntax in this case: hive> SELECT COUNT(1) FROM cite; Total MapReduce jobs = 1 Number of reduce tasks determined at compile time: 1 In order to change the average load for a reducer (in bytes): set hive.exec.reducers.bytes.per.reducer= In order to limit the maximum number of reducers: set hive.exec.reducers.max= In order to set a constant number of reducers: set mapred.reduce.tasks= Starting Job = job_200908250716_0001, Tracking URL = http://ip-10-244-199- 143.ec2.internal:50030/jobdetails.jsp?jobid=job_200908250716_0001 Kill Command = /usr/lib/hadoop/bin/hadoop job -Dmapred.job.tracker=ip-10- 244-199-143.ec2.internal:9001 -kill job_200908250716_0001 map = 0%, reduce =0% map = 12%, reduce =0% map = 25%, reduce =0% map = 30%, reduce =0% map = 34%, reduce =0% map = 43%, reduce =0% map = 53%, reduce =0% map = 62%, reduce =0% map = 71%, reduce =0% map = 75%, reduce =0% map = 79%, reduce =0% map = 88%, reduce =0% map = 97%, reduce =0% map = 99%, reduce =0% map = 100%, reduce =0% map = 100%, reduce =67% map = 100%, reduce =100% Ended Job = job_200908250716_0001 OK 16522439 Time taken: 85.153 seconds Reading the messages, you can see that this query had created a MapReduce job . The beauty of Hive is that the user doesn’t need to know anything about MapReduce at all. As far as she’s concerned, she’s only querying a database using a language similar to SQL. The result of the previous query was printed directly to the screen. In most cases the query result should be saved to disk, which usually would be some other Hive table. Our next query finds the citation frequency of each patent. We first create a table to store this result: hive> CREATE TABLE cite_count (cited INT, count INT); OK Time taken: 0.027 seconds Hive 253 We can execute a query to find the citation frequency. The query uses the COUNT and GROUP BY features, again in a way similar to SQL. There’s an additional INSERT OVERWRITE TABLE clause to tell Hive to write the result to a table: hive> INSERT OVERWRITE TABLE cite_count > SELECT cited, COUNT(citing) > FROM cite > GROUP BY cited; ... map = 100%, reduce =89% map = 100%, reduce =90% map = 100%, reduce =100% Ended Job = job_200908250716_0002 Loading data to table cite_count 3258984 Rows loaded to cite_count OK Time taken: 103.331 seconds The query execution helpfully tells us that 3,258,984 rows were loaded into the cita- tion frequency table. We can execute more HiveQL statements to browse this citation frequency table: hive> SELECT * FROM cite_count WHERE count > 10 LIMIT 10; Total MapReduce jobs = 1 Number of reduce tasks is set to 0 as there's no reduce operator ... map = 80%, reduce =0% map = 100%, reduce =100% Ended Job = job_200908250716_0003 OK 163404 13 164184 16 217584 13 246144 14 288134 11 347644 11 366494 11 443764 11 459844 13 490484 13 An interesting part about this query is that Hive is intelligent enough to know that “Number of reduce tasks is set to 0 as there’s no reduce operator.” When you’re finished with using a table, you can delete it with the DROP TABLE command : hive> DROP TABLE cite_count; OK Time taken: 0.024 seconds Be careful when using this command. It doesn’t ask you for confirmation whether you really want to delete the table or not. It’s difficult to recover a table once you have dropped it. Finally, you can exit your Hive session with the exit command. 254 CHAPTER 11 Hive and the Hadoop herd 11.1.3 HiveQL in details Having seen Hive in action, we’re ready to formally look at different aspects and usage of HiveQL. DATA MODEL We’ve already seen that Hive supports tables as a fundamental data model. Physically, Hive stores tables as directories under /user/hive/warehouse. For example, the cite ta- ble we created earlier would have its data under the /user/hive/warehouse/cite direc- tory. The output table cite_count would be under /user/hive/warehouse/cite_count. In the most basic setup, the directory hierarchy under a table is only one level deep, and the table’s data are spread out over many files under that one directory. Relational databases use indexes on columns to speed up queries on those columns. Hive, instead, uses a concept of partition columns , which are columns whose values would divide the table into separate partitions. For example, a state column would partition a table into 50 partitions, one for each state.6 A date column is a popular partition column for log data; data for each day would belong to its own partition. Hive treats partition columns differently than regular data columns, and executes queries involving partition columns much more efficiently. This is because Hive physically stores different partitions in different directories. For example, let’s say you have a table named users that has two partition columns date and state (plus the regular data columns). Hive will have a directory structure like this for that table: /user/hive/warehouse/users/date=20090901/state=CA /user/hive/warehouse/users/date=20090901/state=NY /user/hive/warehouse/users/date=20090901/state=TX ... /user/hive/warehouse/users/date=20090902/state=CA /user/hive/warehouse/users/date=20090902/state=NY /user/hive/warehouse/users/date=20090902/state=TX ... /user/hive/warehouse/users/date=20090903/state=CA /user/hive/warehouse/users/date=20090903/state=NY /user/hive/warehouse/users/date=20090903/state=TX ... All user data for California (state=CA) on September 1, 2009 (date=20090901) re- sides in one directory, and data for other partitions is in other directories. If a query comes in asking about California users on September 1, 2009, Hive only has to process data in that one directory and ignore data in the users table that have been stored in other partitions. Queries over ranges in the partition columns will involve processing multiple directories, but Hive will still avoid a full scan of all data in users. In some sense partitioning brings similar benefits to Hive as indexing provides to a traditional relational database, although partitioning works at a much less granular level. You’ll want each partition to still be big enough that a MapReduce job on it can be reason- ably efficient. 6 In practice you’ll also have to handle District of Columbia and various territories. Hive 255 In addition to partitions, the Hive data model also has a concept of buckets , which provide efficiency to queries that can work well on a random sample of data. (For example, in computing the average of a column, a random sample of data can provide a good approximation.) Bucketing divides data into a specified number of files based on the hash of the bucket column. If we specify 32 buckets based on user id in our users table, the full file structure for our table in Hive will look like /user/hive/warehouse/users/date=20090901/state=CA/part-00000 ... /user/hive/warehouse/users/date=20090901/state=CA/part-00031 /user/hive/warehouse/users/date=20090901/state=NY/part-00000 ... /user/hive/warehouse/users/date=20090901/state=NY/part-00031 /user/hive/warehouse/users/date=20090901/state=TX/part-00000 ... Each partition will have 32 buckets. By bucketing on user id, Hive will know that each file in part-00000 ... part-00031 has a random sample of users. The computation of many aggregate statistics remains fairly accurate on a sampled data set. Bucketing is particularly useful for speeding up those queries. For example, Hive can run a query on 1/32 of all the users in a partition by using only the data in part-00000, without having to even read the other files. Hive can still do sampling without buckets (or on columns other than the bucket column), but this involves scanning in all the data and randomly ignoring much of it. Much of the efficiency advantage of sampling would therefore be lost. MANAGING TABLES We’ve already seen how to create a simple table for the patent citation data set. Let’s now break down the different parts of a more complicated table creation statement. This one creates a table called page_view . CREATE TABLE page_view(viewTime INT, userid BIGINT, page_url STRING, referrer_url STRING, ip STRING COMMENT 'IP Address of the User') COMMENT 'This is the page view table' PARTITIONED BY (dt STRING, country STRING) CLUSTERED BY (userid) INTO 32 BUCKETS ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t' LINES TERMINATED BY '\n' STORED AS SEQUENCEFILE; The first part looks much like the SQL equivalent: CREATE TABLE page_view(viewTime INT, userid BIGINT, page_url STRING, referrer_url STRING, ip STRING COMMENT 'IP Address of the User') It specifies the name of the table (page_view) and its schema, which includes the name of the columns as well as their type. Hive supports the following data types: ■ TINYINT —1 byte integer ■ SMALLINT —2 byte integer 256 CHAPTER 11 Hive and the Hadoop herd ■ INT —4 byte integer ■ BIGINT —8 byte integer ■ DOUBLE —Double precision floating point ■ STRING —Sequence of characters Noticeably missing is the Boolean type, which is usually handled as TINYINT. Hive also has complex types , such as structs , maps , and arrays that can be nested. But they’re currently not well supported in the language and are considered advanced topics. We can attach a descriptive comment to each column, as was done here for the ip column. In addition, we also add a descriptive comment to the table: COMMENT 'This is the page view table' The next part of the CREATE TABLE statement specifies the partition columns: PARTITIONED BY (dt STRING, country STRING) As we’ve discussed previously, partition columns are optimized for querying. They’re distinct from the data columns of viewTime, userid, page_url, referrer_url, and ip. The value of a partition column for a particular row is not explicitly stored with the row; it’s implied from the directory path. But there’s no syntactical difference in queries over partition columns or data columns. CLUSTERED BY (userid) INTO 32 BUCKETS The CLUSTERED BY (...) INTO ... BUCKETS clause specifies the bucketing informa- tion, including the column that random samples will be taken from and also how many buckets to create. The choice of the number of buckets will depend on the following: 1 The size of your data under each partition 2 The size of sample you intend to use The first criterion is important because after you divide a partition into the specified number of buckets, you wouldn’t want each bucket file to be so small that it becomes inefficient for Hadoop to handle. On the other hand, a bucket should be the same size or smaller than your intended sample size. Bucketing by user into 32 buckets is a good setting if your sample size will be about 3 percent (~1/32) of your user base. NOTE Unlike partitioning, Hive doesn’t automatically enforce bucketing when data is written to a table. Specifying bucketing information merely tells Hive that you’ll manually enforce the bucketing (sampling) criteria when data is written to a table and that Hive can take advantage of it in processing queries. To enforce the bucketing criteria you need to correctly set the number of reducers when populating the table. More detail can be found in http://wiki. apache.org/hadoop/Hive/LanguageManual/DDL/BucketedTables. The ROW FORMAT clause tells Hive how the table data is stored per row. Without this clause, Hive defaults to the newline character as the row delimiter and an ASCII value Hive 257 of 001 (ctrl-A) as the field delimiter. Our clause tells Hive to use the tab character as the field delimiter instead. We also tell Hive to use the newline character as the line delim- iter, but that’s already the default and we include it here only for illustrative purposes: ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t' LINES TERMINATED BY '\n' Finally, the last clause tells Hive the file format to store the table data: STORED AS SEQUENCEFILE; Currently Hive supports two formats, SEQUENCEFILE and TEXTFILE . Sequence file is a compressed format and usually provides higher performance. We can add an EXTERNAL modifier to the CREATE TABLE statement such that the table is created to point to an existing data directory. You’ll need to specify the location of this directory. CREATE EXTERNAL TABLE page_view(viewTime INT, userid BIGINT, page_url STRING, referrer_url STRING, ip STRING) LOCATION '/path/to/existing/table/in/HDFS'; After you’ve created a table, you can ask Hive the table’s schema with the DESCRIBE command: hive> DESCRIBE page_view; You can also change the table structure with the ALTER command. This includes chang- ing the table’s name: hive> ALTER TABLE page_view RENAME TO pv; or adding new columns: hive> ALTER TABLE page_view ADD COLUMNS (newcol STRING); or deleting a partition: hive> ALTER TABLE page_view DROP PARTITION (dt='2009-09-01'); To delete the whole table, use the DROP TABLE command: hive> DROP TABLE page_view; To know what tables are being managed by Hive, you can show them all with hive> SHOW TABLES ; If there are so many tables in use that it becomes unwieldy to list them all, you can narrow down the result with a Java regular expression : hive> SHOW TABLES 'page_.*'; LOADING DATA There are multiple ways to load data into a Hive table. The LOAD DATA command is the workhorse: 258 CHAPTER 11 Hive and the Hadoop herd hive> LOAD DATA LOCAL INPATH 'page_view.txt' > OVERWRITE INTO TABLE page_view; This takes a local file named page_view.txt and loads its content into the page_view table. If we omit the OVERWRITE modifier, the content is added to the table rather than replacing whatever already exists in it. If we omit the LOCAL modifier, the file is taken from HDFS instead of the local filesystem. The LOAD DATA command also allows you to name a specific partition in the table to load the data into: hive> LOAD DATA LOCAL INPATH 'page_view.txt' > OVERWRITE INTO TABLE page_view > PARTITION (dt='2009-09-01', country='US'); When working with data from the local filesystem, it’s useful to know that you can execute local Unix commands from within the Hive CLI. You prepend the command with the exclamation mark (!) and end it with a semicolon (;). For example, you can get a file listing hive> ! ls ; or examine the first few lines of a file hive> ! head hive_result ; Note that the spaces around ! and ; aren’t necessary. We’ve added them for readability. RUNNING QUERIES For the most part, running HiveQL queries is surprisingly similar to running SQL que- ries. One of the general differences is that the results of HiveQL queries are relatively large. You should almost always have an INSERT clause to tell Hive to store your query result somewhere. Often it’s some other table: INSERT OVERWRITE TABLE query_result Other times it’s a directory in HDFS: INSERT OVERWRITE DIRECTORY '/hdfs_dir/query_result' And sometimes it’s a local directory: INSERT OVERWRITE LOCAL DIRECTORY '/local_dir/query_result' The basic queries can look almost identical to SQL: INSERT OVERWRITE TABLE query_result SELECT * FROM page_view WHERE country='US'; Note that the query is over a partition column (country), but the query would look exactly the same if it was a data column instead. One syntax to adjust to is that HiveQL uses COUNT(1) in places where SQL typically would use COUNT(*) . For example, you would use this HiveQL query to find the number of page views from the U.S.: Hive 259 SELECT COUNT(1) FROM page_view WHERE country='US'; Like SQL, the GROUP BY clause allows one to do aggregate queries on groups. This query will list the number of page views from each country: SELECT country, COUNT(1) FROM page_view GROUP BY country; And this query will list the number of unique users from each country: SELECT country, COUNT(DISTINCT userid) FROM page_view GROUP BY country; Table 11.2 shows all the operators supported in HiveQL. These are quite stan- dard in SQL and programming languages and we won’t explain them in detail. The main exception is in regular expression matching. HiveQL provides two commands for regular expression matching—LIKE and REGEXP . (RLIKE is equiv- alent to REGEXP.) LIKE only performs simple SQL regular expression matching, where an underscore (_) character in B matches any single character in A and the percent (%) character matches any number of characters in A. REGEXP treats B as a full Java regular expression.7 Tables 11.3 and 11.4 list the majority of HiveQL functions. Table 11.2 Standard operators in HiveQL Operator type Operators Comparison A = B , A <> B , A < B , A <= B , A > B , A >= B , A IS NULL , A IS NOT NULL , A LIKE B , NOT A LIKE B , A RLIKE B , A REGEXP B Arithmetic A + B , A - B , A * B , A / B , A % B Bit-wise A & B , A | B , A ^ B, ~A Logical A AND B, A && B, A OR B, A || B, NOT A, !A One of the main motivators for users to seek a higher-level language, such as Pig Latin and HiveQL, is the support of joins . Currently HiveQL only supports equijoins (joins on equality) . An example join query is INSERT OVERWRITE TABLE query_result SELECT pv.*, u.gender, u.age FROM page_view pv JOIN user u ON (pv.userid = u.id); 7 The format of Java regular expression is fully explained in the Javadoc http://java.sun.com/j2se/1.4.2/ docs/api/java/util/regex/Pattern.html. 260 CHAPTER 11 Hive and the Hadoop herd Syntactically, in the FROM clause you add the JOIN keyword between the tables and then specify the join columns after the ON keyword. To join more than two tables, we repeat the pattern like this: INSERT OVERWRITE TABLE pv_friends SELECT pv.*, u.gender, u.age, f.friends FROM page_view pv JOIN user u ON (pv.userid = u.id) JOIN friend_list f ON (u.id = f.uid); We can add sampling to any query by modifying the FROM clause. This query tries to compute the average view time, except the average is only taken from data in the first bucket out of 32 buckets: SELECT avg(viewTime) FROM page_view TABLESAMPLE(BUCKET 1 OUT OF 32); The general syntax for TABLESAMPLE is TABLESAMPLE(BUCKET x OUT OF y) The sample size for the query is around 1/y. In addition, y needs to be a multiple or factor of the number of buckets specified for the table at table creation time. For ex- ample, if we change y to 16, the query becomes SELECT avg(viewTime) FROM page_view TABLESAMPLE(BUCKET 1 OUT OF 16); Then the sample size includes approximately 1 out of every 16 users (as the bucket column is userid). The table still has 32 buckets, but Hive tries to satisfy this query by processing buckets 1 and 17 together. On the other hand, if y is specified to be 64, Hive will execute the query on half of the data in one bucket. The value of x is only used to select which bucket to use. Under truly random sampling its value shouldn’t matter. Besides avg, Hive has many other built-in functions. You can see some of the more common ones in tables 11.3 and 11.4. Programmers can also add UDFs to Hive for custom processing. A brief introduction on how to create a UDF is given in http:// wiki.apache.org/hadoop/Hive/AdminManual/Plugins. 11.1.4 Hive Sum-up Hive is a data warehousing layer built on top of Hadoop’s massively scalable architec- ture. By focusing on structured data, Hive has added many performance-enhancing techniques (such as partitions) and usability features (such as a SQL-like language). It makes certain frequent tasks, such as joining, easy. Hive is introducing Hadoop tech- nology to a wider audience of analysts and other nonprogrammers. As of August 2009, Facebook counts 29 percent of its employees as Hive users, more than half of whom are outside of engineering.8 8 This fact is cited in http://www.facebook.com/note.php?note_id=114588058858. An explanation of how Facebook decided to build out its Hadoop infrastructure is given in http://www.facebook.com/note. php?note_id=16121578919. The presentation at http://www.slideshare.net/zshao/hive-data-warehousing- analytics-on-hadoop-presentation gives a detailed description of how Facebook designed its data warehouse and analytics system around Hive. Hive 261 Table 11.3 Built-in functions Function Description concat(string a, string b) Returns the concatenation of string a with string b. substr(string str, int start) substr(string str, int start, int length) Returns the substring of str starting at start. The result goes until the end of str unless an optional length argument is specified. round(double num) Returns the closest integer (BIGINT). floor(double num) Returns the largest integer (BIGINT) that’s equal to or smaller than num. ceil(double num) ceiling(double num) Returns the smallest integer (BIGINT) that’s equal to or bigger than num. sqrt(double num) Returns the square root of num. rand() rand(int seed) Returns a random number (that changes from row to row). The optional seed value can make the random number sequence deterministic. Ln(double num) Returns the natural log of num. log2(double num) Returns the base-2 log of num. log10(double num) Returns the base-10 log of num. log(double num) log(double base, double num) Returns the natural log of num. Or returns the base-base log of num. exp(double a) Raise e (the base of natural logarithm) to the power of a. power(double a, double b) pow(double a, double b) Returns a raised to the power of b. upper(string s) ucase(string s) Returns string s in uppercase. lower(string s) lcase(string s) Returns string s in lowercase. trim(string s) Returns string s with spaces trimmed on both ends. ltrim(string s) Returns string s with spaces trimmed on the left end. rtrim(string s) Returns string s with spaces trimmed on the right end. regexp(string s, string regex) Returns whether the string s matches the Java regular expression regex. regexp_replace(string s, string regex, string replacement) Returns a string where all parts of s that match the Java regular expression regex are replaced with replacement. day(string date) dayofmonth(string date) Returns the day part of a date or timestamp string. month(string date) Returns the month part of a date or timestamp string. year(string date) Returns the year part of a date or timestamp string. To_date(string timestamp) Returns the date part (year-month-day) of a timestamp string. unix_timestamp(string timestamp) Convert a timestamp string to UnixTime. 262 CHAPTER 11 Hive and the Hadoop herd Table 11.3 Built-in functions (continued) Function Description from_unixtime(int unixtime) Convert integer in UnixTime to a timestamp string. date_add(string date, int days) Add a number of days to a date string. date_sub(string date, int days) Subtract a number of days from a date string. datediff(string date1, string date2) Calculate the difference in number of days. Result is negative if date1 is earlier. Table 11.4 Built-in aggregate functions Function Description count(1) count(DISTINCT col) Returns the number of members in the group, or the number of distinct values of the column. sum(col) sum(DISTINCT col) Returns the sum of the values of the column, or the sum of the distinct values of the column. avg(col) avg(DISTINCT col) Returns the average value of the column, or the average of the distinct values of the column. max(col) Returns the maximum value of the column. min(col) Returns the minimum value of the column. 11.2 Other Hadoop -related stuff The Hadoop ecosystem is growing every day. The following are projects or vendors related to Hadoop that we find useful or that have tremendous potential. All of them except Aster Data and Greenplum are open source in some way. 11.2.1 HBase http://hadoop.apache.org/hbase/—HBase is a scalable data store targeted at random read and write access of (fairly-)structured data. It’s modeled after Google’s Bigtable9 and targeted to support large tables, on the order of billions of rows and millions of col- umns. It uses HDFS as the underlying filesystem and is designed to be fully distributed and highly available. Version 0.20 introduces significant performance improvement. 11.2.2 ZooKeeper http://hadoop.apache.org/zookeeper/—ZooKeeper is a coordination service for building large distributed applications. You can use it independently from the Hadoop Core framework. 9 “Bigtable: A Distributed Storage System for Structured Data” by Chang et al., OSDI ‘06—Seventh Symposium on Operating System Design and Implementation. http://labs.google.com/papers/bigtable.html. Other Hadoop-related stuff 263 It implements many of the common services used in large distributed applications, such as configuration management, naming, synchronization, and group services. Histori- cally developers have to reinvent these services for each distributed application, which is time consuming and error prone, as these services are notoriously difficult to imple- ment correctly. By abstracting away the underlying complexity, ZooKeeper makes it easy to implement consensus, leader election, presence protocols, and other primitives, and frees the developer to focus on the semantics of her application. ZooKeeper is often a major component in other Hadoop-related projects, such as HBase and Katta . 11.2.3 Cascading http://www.cascading.org/—Cascading is an API for assembling and executing complex data processing workflows on Hadoop. It abstracts away the MapReduce model into a data processing model consisting of tuples , pipes , and (source and sink) taps. Pipes operate on streams of tuples, where operations include Each, Every, GroupBy, and CoGroup. Pipes can be assembled and nested to create an “assembly.” We create an executable “flow” when we attach a pipe assembly to a (data) source tap and a (data) sink tap . Cascading shares many design similarities and goals with Pig . One difference, though, is that Pig’s Grunt shell makes it easier to execute ad hoc queries. Another difference is that Pig programs are written in Pig Latin, whereas Cascading works more like a Java framework in which you create a data processing flow through instantiating various Java classes (Each, Every, etc.). Using Cascading doesn’t require learning a new language, and the data process flow created can be more efficient because you’ve written it directly yourself. 11.2.4 Cloudera http://www.cloudera.com/—Cloudera is attempting to do for Hadoop what RedHat has done for Linux. It’s supporting and packaging Hadoop to be easy and friendly to en- terprise users. It provides live training sessions in major cities as well as educational videos on their web site. You can simplify your deployment of Hadoop by using their free Hadoop distribution, in either RPM or Ubuntu /Debian Packages. Their Hadoop distribution is based on the most recent stable release of Hadoop, plus useful (and tested) patches from future releases, and additional tools such as Pig and Hive. Cloud- era also offers consulting and support services to help enterprises use Hadoop. 11.2.5 Katta http://katta.sourceforge.net/—As Hadoop can trace its origin to search engines, it should not be surprising to see it being applied to distributed indexing and search. Nutch is a web search engine built on top of Hadoop.10 But as a web search engine, Nutch has many unique requirements. It is often a mismatched solution for specific search applications. 10 It may be more accurate to say that Nutch motivated the creation of Hadoop. See chapter 1 for full history. 264 CHAPTER 11 Hive and the Hadoop herd Katta is a scalable, fault tolerant, distributed indexing system. It’s more lightweight and flexible than Nutch. In some sense it’s adding some extra capabilities (such as replication, redundancy, fault tolerance, and scalability) to Lucene while retaining the basic application semantics. 11.2.6 CloudBase http://cloudbase.sourceforge.net/—CloudBase is an ANSI SQL data warehousing layer on top of Hadoop. Unlike Hive, CloudBase works directly on flat files without any meta- data store. It makes a stricter goal of ANSI SQL adherence, and interaction is primar- ily through a JDBC driver, which makes it easier to connect to business intelligence reporting tools. For the most part, CloudBase is a compiler that takes SQL queries and compiles them into MapReduce programs. As of this writing, CloudBase has a less active developer community than Pig or Hive, and its GPL license is more restrictive than the Apache license. 11.2.7 Aster Data and Greenplum http://www.asterdata.com/; http://www.greenplum.com/—Aster Data Systems and Green- plum are both commercial vendors offering high-performance, scalable data ware- housing solutions tightly combining SQL with MapReduce. Although they support the MapReduce programming model, they were both created independently from Ha- doop and had made many different underlying design choices. Unlike Hadoop, their offerings are architected much more specifically toward enterprise customers looking for higher-performing SQL data warehouses. As they come at the MapReduce para- digm from a different angle than Hadoop, studying them can help understand some of Hadoop’s architectural trade-offs. 11.2.8 Hama and Mahout http://incubator.apache.org/hama/; http://lucene.apache.org/mahout/—Both Hama and Mahout are projects for scientific data processing using Hadoop. Hama is a matrix computation package for calculating products, inverse, eigenvalues, eigenvectors, and other matrix operations. Mahout is targeted more specifically at implementing ma- chine learning algorithms on Hadoop (for more information, see Mahout in Action, Manning Publications). Mahout version 0.1 was released in April 2009 and included implementations such as Naïve Bayes classification, k-means clustering, and collabora- tive filtering. At the time of this writing both projects are relatively new and under the Apache incubator. Interested readers should consider becoming contributors. 11.2.9 search-hadoop.com As a Hadoop programmer, you’ll often have the need to find some piece of documen- tation about Hadoop or its subprojects. Sematext , a company specializing in search and analytics, runs http://search-hadoop.com/, a site that lets you search across all Summary 265 Hadoop subprojects and data sources—mailing list archives, Wikis, issue tracking systems, source code, and so on. The underlying search index is continuously updated. Search results allow filtering by project, data source, and author, and can be sorted by date, relevance, or the combination of the two. 11.3 Summary This chapter covered many of the additional tools you can use with Hadoop. We gave special attention to Hive, a data warehousing package that allows you to process data in Hadoop using an SQL-like language. A rich ecosystem of supporting software has sprung up around Hadoop, and you’ll see some of them in action in the case studies in the next chapter. 12 266 Case studies This chapter covers The New York Times■ China Mobile■ StumbleUpon■ IBM■ We’ve been through many exercises and sample programs by now. The next step is to integrate what you’ve learned about Hadoop into your own real-world applications. To help you in that transition, this chapter provides examples of how other enter- prises have used Hadoop as part of the solution to their data processing problems. The case studies serve two purposes. One is to step back and see the broader systems that utilize Hadoop as a critical part. You’ll discover complementary tools, such as Cascading, HBase, and Jaql. The second purpose is to demonstrate the variety of businesses that have used Hadoop to solve their operational challenges. Our case studies span industries, including media (the New York Times), telecom (China Mobile), internet (StumbleUpon), and enterprise software (IBM). Mining data at China Mobile 267 12.1 Converting 11 million image documents from the New York Times archive In 2007, the New York Times decided to make all their public domain articles between 1851 and 1922 freely available on their website. Doing this required a scalable image con- version system. Because the Times had stored its older articles as scanned TIFF images , they needed image processing to combine different pieces of each article together into a single file in the desired PDF format. Previously, these articles were behind a paid wall and didn’t receive much traffic. The Times could use a real-time approach to scale, glue, and convert the TIFF images. Although that worked well enough for a low volume of requests, it would not scale to handle the significant traffic increase expected from the articles’ free availability. The Times needed a better architecture to handle the opening of its archive. The solution was to pregenerate all the articles as PDF files and serve them like any other static content. The New York Times already had the code to convert the TIFF images to PDF files. It looked like a simple matter of batch processing all the articles in one setting instead of dealing with each individual article as a request came in. The challenging part of this solution came when one realized that the archive had 11 million articles consisting of 4 TB of data. Derek Gottfrid , a software programmer at the Times, thought this was a perfect opportunity to use the Amazon Web Services (AWS ) and Hadoop. Storing and serving the final set of PDFs from Amazon’s Simple Storage Service (S3 ) was already deemed a more cost-effective approach than scaling up the storage back-end of the website. Why not process the PDFs in the AWS cloud as well? Derek copied the 4 TB of TIFF images into S3. He “started writing code to pull all the parts that make up an article out of S3, generate a PDF from them and store the PDF back in S3. This was easy enough using the JetS3t —Open Source Java toolkit for S3, iText PDF Library and installing the Java Advanced Image Extension .”1 After tweaking his code to work within the Hadoop framework, Derek deployed it to Hadoop running on 100 nodes in Amazon’s Elastic Compute Cloud (EC2 ). The job ran for 24 hours and generated another 1.5 TB of data to be stored in S3. At 10 cents per instance per hour, the whole job ended up costing only $240 (100 instances x 24 hours x $0.10) in computation. The storage cost for S3 was extra, but as the Times had decided to archive its files in S3 anyway, that cost was already amortized. Data transfer between S3 and EC2 being free, the Hadoop job didn’t incur any bandwidth cost at all. The whole effort took only a single employee. Thanks to Derek’s work, it has become much easier for people to look up the New York Times’ account of historic events. 12.2 Mining data at China Mobile Contributed by ZHIGUO LUO , MENG XU , and SHAOLING SUN —Research Institute of China Mobile Communication Corporation. 1 http://open.blogs.nytimes.com/2007/11/01/self-service-prorated-super-computing-fun/. 268 CHAPTER 12 Case studies China Mobile Communication Corporation (CMCC) is the largest mobile phone operator in the world. Traded on NYSE under the symbol CHL, China Mobile is seventh in BrandZ’s global brand equity ranking for 2009, behind McDonald’s and Apple but ahead of General Electric. With more than two-thirds of China’s mobile phone market, CMCC serves the communication needs of 500 million subscribers. Even at its size, China Mobile has experienced rapid growth. For example, in 2006 when its subscriber base was only 300 million, its subscriber growth rate was at 22 percent, its voice usage was growing at 30 percent, and its SMS usage was growing at 41 percent annually. As with any telecom operator, China Mobile generates a lot of data in the normal course of running its communication network. For example, each call generates a call data record (CDR) , which includes information such as the caller’s phone number, the callee’s phone number, the start time of the call, the call’s duration, information about the call’s routing, and so forth. In addition to CDR, a phone network also generates signaling data between various switches, nodes, and terminals within the network. At a minimum, we need this data for completing calls and accurately billing customers. We also need it to analyze for marketing, network tuning, and other purposes. At China Mobile, the size of its network naturally leads to large amounts of data created. Every day the network generates 5 TB to 8 TB of CDR data. A branch company of China Mobile can have more than 20 million subscribers, leading to more than 100 GB of CDR data for voice calls and between 100 GB to 200 GB of CDR data for SMS every day. In addition, a typical branch company generates around 48 GB of data per day for General Packet Radio Service (GPRS ) signaling and 300 GB of data per day for 3G signaling. China Mobile looks to data warehousing and mining of this data to extract insights for improving marketing operations, network optimization, and service optimization. Some typical applications include ■ Analyzing user behavior ■ Predicting customer churn ■ Analyzing service association ■ Analyzing network quality of service (QOS ) ■ Filtering spam messages Analyzing signaling data ■ China Mobile has experiences with commercial data mining tools from some well- known vendors. These tools’ architectural design limits China Mobile’s current data mining system because it requires all data to be processed within a single server. Hard- ware capacity thus becomes a performance bottleneck. The current system at one of the branch companies is based on commercial solutions and consists of a Unix server with eight CPU cores, 32 GB memory, and a storage array. It can only perform user behav- ior analysis for up to 1.4 million customers, or only about 10 percent of customer data of that particular branch company. Even within the limitation on the amount of data that can be processed, the current system takes too much time for many applications. Mining data at China Mobile 269 In addition, the high-end Unix servers and storage arrays are expensive, and the com- mercial package software don’t support custom algorithms well. Because of the limitations of the current system, China Mobile initiated an experimental project to develop a parallel data mining tool set on Hadoop and evaluated it against its current system. They named the project Big Cloud–based Parallel Data Mining (BC-PDM ) and it was architected to achieve four objectives : ■ Massive scalability—Using Hadoop for a scale-out architecture ■ Low cost—Built around cheap commodity hardware and free software ■ Customizable—Applications built around specific business requirements ■ Ease of use—Graphical user interface similar to ones in commercial tools BC-PDM implemented many of the standard ETL operations and data mining algo- rithms in MapReduce. The ETL operations include computing aggregate statistics, attribute processing, data sampling, redundancy removal, and others. It implemented nine data mining algorithms from three categories. The categories include clustering (e.g., K-means ), classification (e.g., C4.5), and association analysis (e.g., Apriori ). The MapReduce programs were executed and evaluated within a Hadoop cluster consist- ing of 256 nodes connected to a single 264-port Gbps switch. The hardware for the nodes are ■ Datanode/TaskTracker —1-way 4-core Xeon 2.5 GHz CPU, 8 GB RAM, 4 x 250 GB SATA II disks ■ Namenode/JobTracker —2-way 2-core AMD Opteron 2.6 GHz CPU, 16 GB RAM, 4 x 146 GB SAS disks China Mobile compared BC-PDM against their existing data mining solution using real data from China Mobile’s Business Analysis Support System (BASS) . There were three different data sets. They were all fairly large, and for certain evaluation tasks smaller, sampled subsets were needed. You’ll see the original size (Large scale) as well as the size of the sampled subsets (Middle scale and Small scale) in table 12.1. Table 12.1 Size of data sets used for evaluation Data Large scale Middle scale Small scale Users’ behavior 12 TB 120 GB 12 GB Users’ accessing 16 TB 160 GB 16 GB New service association 120 GB 12 GB 1.2 GB China Mobile evaluated BC-PDM on four dimensions : correctness, performance, cost, and scalability. Correctness was, of course, necessary for the new system to be useful. They verified BC-PDM’s parallel ETL operations by ensuring the same results were generated as the existing serial ETL implementation. The data mining algorithms, on the other hand, were not expected to generate the exact same results as the existing 270 CHAPTER 12 Case studies system. This is because minor implementation and execution details, such as initial conditions and ordering of the input data, can affect the exact output. They examined the results and checked for general consistency. In addition, the UCI data sets were also employed to verify BC-PDM’s parallel data mining algorithms. The UCI data sets are popular among researchers in the machine-learning community and are well un- derstood. China Mobile can verify BC-PDM’s output with known expected models. After establishing the correctness of the MapReduce implementations, the performance of BC-PDM was compared to the current system. The comparison used only a 16-node Hadoop cluster. As we’ll see, this small cluster is cheaper than the monolithic big-iron server of the current system. Figure 12.1 shows the timing between the two setups for ETL operations (left graph) and data mining tasks (right graph). Note that BC-PDM was tasked to process 10 times the amount of data as the current system. BC-PDM was faster for all ETL operations, resulting in general performance improvement of 12 to 16 times. For data mining tasks BC-PDM was further stress- tested with 100 times the amount of data as the current system. Even with data size two orders of magnitude larger, BC-PDM was faster than the current system at the Apriori (association) and C4.5 (classification) algorithms. The K-means clustering algorithm took slightly longer to complete than the current system at 10 times the data size. Full end-to-end testing of three applications from the Shanghai Branch Company showed performance improvement of 3 to 7 times. These real-world applications include channel preference modeling, new service association modeling, and subscriber subdivision modeling. Recall that BC-PDM in this evaluation is based on a relatively small 16-node Hadoop cluster . As we’ll see later, BC-PDM and Hadoop scale well with additional nodes. At the full size of our cluster of 256 nodes, we expect BC-PDM to be able to store, process, and mine data at the 100-TB scale. Not only is the 16-node BC-PDM cluster outperforming the current system, it’s also significantly cheaper. Table 12.2 shows a cost breakdown of the two systems. (As of this writing, one USD converts to a little less than seven RMB.) The 16-node Hadoop/ BC-PDM cluster is roughly one fifth the cost of the current commercial solution. The biggest saving comes from the use of low-cost commodity servers. In fact, the hardware 700 600 500 400 300 200 100 0 AttriDelAttriAddReplace Intervalize MissingReplace Join StatisticSample RedRemove 2000 1500 1000 500 0 Unix&Tools BC-PDM with 10 times data Unix&Tool Papriori PC45 Pkmeans BC-PDM with 100 times data BC-PDM with 10 times data Figure 12.1 Performance comparison of the Hadoop cluster versus existing commercial Unix server. The left graph tests ETL operations whereas the right graph is for data mining algorithms. Mining data at China Mobile 271 cost of the 16-node cluster is less than one sixteenth the hardware cost of the current scale-up solution. ’Till now, we’ve investigated the correctness, performance, and cost of the new BC- PDM system. Let’s examine the scalability of the system as we add more nodes to the cluster. We ran the ETL operations and data mining algorithms on three cluster sizes: 32 nodes, 64 nodes, and 128 nodes. We measured the speed-up in execution on larger clusters, taking the execution time in the 32-node cluster as baseline. You’ll see the results in figure 12.2, with the left graph showing the speed-up for ETL operations and the right graph showing data mining operations. Note that the horizontal axis (representing cluster size) is exponential, with marks doubling from 32 to 64 to 128. As exact linear scalability of the operations is the best that we can hope for, the ideal graph would have the speed-up go from 1 to 2 to 4. We see that many ETL operations are close to this linear scalability ideal. In fact, when the cluster size quadrupled, from Figure 12.2 Scalability of ETL (left) and data mining (right) algorithms on Hadoop cluster as extra nodes are added. The horizontal axis represents the number of nodes in the BC-PDM cluster. The vertical axis represents the speed-up, relative to the execution time on the 32-node cluster. Table 12.2 Comparison of cost and configuration between existing solution and a 16-node Hadoop cluster. (As of this writing, one USD converts to a little less than seven RMB.) BC-PDM (16 nodes) Existing commercial Unix server Hardware cost Computing ability CPU: 64 cores memory: 128 GB CPU: 8 cores memory: 32 GB Storage ability 16 TB (4 x 256 GB SATA II each node) storage array Cost 240,000 RMB 4,000,000 RMB Software cost Database 500,000 RMB 1,000,000 RMB Application software 300,000 RMB 500,000 RMB Maintenance cost 200,000 RMB 500,000 RMB Total 1,240,000 RMB 6,000,000 RMB 272 CHAPTER 12 Case studies 32 nodes to 128 nodes, all but two had a speed-up greater than 2.5.2 The data mining algorithms are more complex, yet they achieve respectable scalability as well. Our performance testing earlier (figure 12.1) used only a small 16-node cluster, which we don’t even consider in the current scalability benchmark. Yet the 16-node cluster could handle an order of magnitude more data faster than the existing commercial solution. Together these evaluations demonstrate our BC-PDM cluster’s ability to handle data at the 100-TB level going forward. After the thorough evaluation of the BC-PDM system, we worked with the Shanghai Branch to apply our system to some of their business needs. One application was to characterize their user base to enable precision marketing. More specifically, they wanted to know how their users are segmented, the characteristics and differences of each segment, and to classify each user for targeted marketing. We used the parallel K-means algorithm from our data mining toolset to cluster their user base and created the market segmentation graph in figure 12.3. Further analysis helped to characterize each segment according to the average bill and types of service used. BC-PDM performed this analysis 3 times faster than their existing Unix solution. In conclusion, China Mobile is a large mobile communication provider, and there’s tremendous and growing need to analyze large data sets . Current commercial offerings are expensive and inadequate for analyzing our user data. We investigated the use of Hadoop. We built a data mining system called BC-PDM on top of MapReduce and HDFS and found this system to be accurate, fast, cheap, and scalable. Going forward, we’ll improve BC-PDM’s efficiency as well as expand its scope by implementing more ETL operations and data mining algorithms. More importantly, we intend to establish BC-PDM as a service platform for data analysis across China Mobile’s branch companies. 12.3 Recommending the best websites at StumbleUpon Contributed by KEN MACINNIS and RYAN RAWSON Using a combination of human opinions and machine learning to immediately deliver relevant content, StumbleUpon presents only websites that have been sug- gested by other like-minded Stumblers . Each time you click the Stumble button, you are presented with a high-quality website based on the collective opinions of other like-minded web surfers. StumbleUpon uses ratings of “like” and “dislike” to form collaborative opinions on website quality. When you “stumble,” you’ll only see pages that friends and like-minded 2 The two exceptions were the Join and the Duplicate Removal operations. They ran in roughly constant time irrespective of the cluster size. We are currently investigating the underlying reason for it. One possible explanation for Hadoop running a job in constant time (independent of cluster size) is that the job is not evenly distributed and one task is the bottleneck to the job’s completion. Recommending the best websites at StumbleUpon 273 Stumblers have recommended. This will help you discover great content that is hard to find using a traditional search engine. 12.3.1 Distributed beginnings at StumbleUpon To collect and analyze this stumbling data, StumbleUpon requires its highly available back-end platform to collect, analyze, and transform millions of ratings per day. With nearly 10 million users at present, StumbleUpon fairly quickly surpassed the abilities a traditional LAMP (Linux, Apache, MySQL, PHP) stack afforded us, and we began to build a distributed platform for the following reasons: ■ Scalability —Commodity hardware scales easily in many cases. Twenty Hadoop nodes may cost only as much as a single redundant database slave pair. ■ Freedom of development —Developers have fewer restrictions when compared to de- signing around a carefully architected, somewhat fragile RDBMS. ■ Operational concerns —Removing as many single-point-of-failure cases as possible is crucial to smooth operation of a world-class service. ■ Data processing speed —Many system-wide calculations were simply not possible to perform with a monolithic system. Type General: no special feature, average bill ¥ 46 Type Economic: less service, some basic calling, average bill ¥ 54 Type Life: high mark in life related services, average bill ¥ 70 Type Business: high bill, high roam call, average bill ¥ 291 9.2%55.0% 4.5% 1.8% 1.8% Type White Collar: more MMS, more VAS, high mark in fashion, average bill ¥ 251 27.6% Type Chatting: high SMS, high GPRS, average bill ¥ 120 Figure 12.3 Cluster analysis of user base for China Mobile’s Shanghai Branch using the K-means algorithm. The result can be used for the company’s marketing campaigns. 274 CHAPTER 12 Case studies 12.3.2 HBase and StumbleUpon HBase plays a critical part in StumbleUpon ’s distributed platform. HBase is a distrib- uted, column-oriented database that harnesses the power of the Hadoop and HDFS platform underneath it. But, as with any complex system, there are trade-offs: HBase shelves traditional relational database concepts, such as joins, foreign key relations, and triggers in the pursuit of a system that hosts immensely large, sparsely populated data on commodity hardware in a scalable manner. AN INTRODUCTION TO HBASE HBase is modeled after Google’s Bigtable,3 a distributed storage system. Let’s recap the basics of Bigtable and Bigtable-like systems: ■ Shares concepts of both column- and row-oriented databases. As described by the authors, Bigtable is a “a sparse, distributed multidimensional sorted map.” The basic unit of storage, a table, is split into multiple tablets (regions in HBase parlance). ■ Writes are buffered in memory, then flushed into read-only files after a while. ■ To keep the number of files low, they are merged in a compaction process that rewrites N files into 1. ■ Special tablets or regions are used to track the locations of the data. ■ Due to the column-oriented nature of the datastore, sparse tables—those with a majority of null cell values—are virtually free as null values aren’t stored explicitly. ■ Column families are used to group row columns. All columns in a family are stored together (for locality) and share storage and configuration parameters. ■ Table cells are stored with multiple versions instead of overwriting existing data. ■ Capacity (both storage size and processing speed) can be increased by simply adding machines to the cluster, with no special configuration needed. HBase provides many additional features: ■ REST and Thrift4 gateways allowing for easy access from non-Java development environments ■ Easy integration with Hadoop MapReduce for data processing ■ Harnesses the proven reliability and scalability of Hadoop and HDFS ■ Strong open source community Web-based UIs for management of both the master and region servers ■ 3 Bigtable: A Distributed Storage System for Structured Data. Chang, et al. http://labs.google.com/papers/ bigtable.html. 4 Thrift is a remote procedure call library originally developed at Facebook. It’s now an Apache incubator project at http://incubator.apache.org/thrift/. Recommending the best websites at StumbleUpon 275 Figure 12.4 describes a simplistic version of the data write path in an HBase region server . The write is appended to the server’s write-ahead log: 1 Data is then inserted in to the MemStore . 2 As the MemStore grows beyond a threshold, it’s flushed to a new file on disk. 3 When there are too many files on disk, data store files are compacted into fewer files (minor compaction). Please visit the project site for further information on obtaining, running, and enhanc- ing HBase.5 USING HBASE AT STUMBLEUPON StumbleUpon carefully selected HBase from a host of candidate database and database-like storage and retrieval systems. We value full consistency, where any query subsequent to a write operation is guaranteed to reflect that write. In addition, StumbleUpon is committed to the open source model where we are free to contribute back to the community, and HBase’s strong development community both reflected that commitment and offered a valuable resource with which to drive improvements to the product. Our first large test of HBase was in importing existing, legacy data from our MySQL -based systems. In the past, we undertook this process only when absolutely necessary (such as migrating tables or hosts) and could take days or weeks to complete. COLUMN VERSUS ROW STORE You can see one example of the column -store design pattern in the storage of arbitrary attributes for a user across multiple logical attribute groupings. In this example, we assume a user has Figure 12.4 HBase write operations 5 http://hadoop.apache.org/hbase/. 276 CHAPTER 12 Case studies ■ Contact data—Email and web addresses, instant messaging names, profile photo URLs ■ Statistics—“Signup date,” “last login time,” “last client version seen” ■ Attributes—For remote login credentials, to authenticate to a third-party service ■ Permissions—For access to site features and data In the traditional RDBMS world, we may arbitrarily assign each group to a table. User attributes may be retrieved and associated simultaneously with joins and foreign keys. With careful design6 and a relatively moderate amount of data, such a system is flex- ible and maintainable. However, as access patterns change—for instance, we desire to store multiple credentials per user where we assumed only one to begin—the design is difficult to change. Furthermore, this design suffers its most fatal flaws when the data volume scales past a moderate amount and the schema needs to be refactored. The idea of doing an ALTER TABLE on a production database table containing millions or billions of rows as well as the headache of vetting systemic schema changes for both correctness and completeness is a daunting prospect. Even with a perfect, static, concrete table, data analysis becomes bottlenecked by the selection, input, and output of records. Let’s take a look at listing 12.1. It’s a simple example where our typical user only has an ID and a record per Stumble: Listing 12.1 Determining Stumbles per user, per URL public class CountUserUrlStumbles { public static class Map extends MapReduceBase implements Mapper { @Override public void map(ImmutableBytesWritable key, RowResult value, OutputCollector output, Reporter reporter) throws IOException { byte [] row = value.getRow(); int userid = StumbleUtils.UserIndex.getUserId(row); int urlid = StumbleUtils.UserIndex.getUrlId(row); Text one = new Text("1"); output.collect(new Text("U:" + Integer.toString(userid)), one); output.collect(new Text("Url:" + Integer.toString(urlid)), one); } } public static class Reduce extends MapReduceBase implements Reducer { @Override public void reduce(Text key, Iterator values, OutputCollector output, Reporter reporter) throws IOException { 6 Rarely achieved on the first attempt, since final schemas are rarely known fully a priori! Recommending the best websites at StumbleUpon 277 int count = 0; while (values.hasNext()) { values.next(); count++; } output.collect(key, new Text(Integer.toString(count))); } } public static void main(String []args) throws IOException { if (args.length < 2) { System.out.println("Give the name of the by-userid stumble table"); return; } JobConf job = new JobConf(CountUserUrlStumbles.class); job.setInputFormat(TableInputFormat.class); FileInputFormat.setInputPaths(job, args[0]); job.setMapperClass(Map.class); job.setReducerClass(Reduce.class); job.setOutputFormat(TextOutputFormat.class); TextOutputFormat.setOutputPath(job, new Path(args[1])); job.setNumMapTasks(5000); JobClient jc = new JobClient(job); jc.submitJob(job); } } In this example, we look at a routine StumbleUpon task: counting stumbles per user as well as stumbles per URL. Although this task is not particularly complex or insightful, we provide it here as a concrete example to the reader of a type of analytic task we perform on a daily basis. The most interesting bit is that this trivial example completes in about 1 hour (using twenty commodity nodes) when processing a key count in the tens of billions. The MySQL-based counterpart doesn’t complete in a reasonable amount of time—at least not without special handling and support to dump the data from MySQL , split the lines to a reasonable chunk size, and then combine the results. You may find this series of operations familiar: mapping, then reducing! By using the generalized facilities of both HBase and Hadoop, we are able to conduct similar statistical surveys as needed, without special preparation and runtime handling. To apply this straightforward example to the real world, we are now able to complete all analysis tasks in the same day they’re requested. We can provide the ability to run ad hoc queries at a rate not thought possible before Hadoop and HBase were powering our platform. As a business thrives and dies on the data it can analyze, this decreased turnaround time makes an incredible impact from the front office number crunching to the research engineers doing instant spam analysis on content submissions. One can only imagine the difficulty of refactoring the custom-processing pipeline when the data schema is more complex than this trivial example, if we didn’t have our distributed processing platform to power the extraction, transformation, and analysis. 278 CHAPTER 12 Case studies TRANSCENDING A SINGLE MACHINE As we’ve outlined, one of the most important scalability features of HBase is the ability to (finally) transcend the write limits of a single machine. Typically, scaling a database involves adding read slaves and caching to the system. Read slaves can only help if your application is reading more than writing. Caching only helps if your data set doesn’t change too often. Even so, these architectural features frequently add vast application-side complexity. HBase hosts each region on any one of the machines in the cluster (each is a region server ). Writes touch the region server hosting that region, and the HBase region server writes to three (by default) HDFS data nodes. 7 With a large table and a similarly large cluster, writes are spread out to many different machines, inherently avoiding the single machine write problem that master-slave data stores have. This feature can help you scale beyond traditional relational systems at a fraction of the cost. As larger hardware tends to become expensive faster than the actual performance delivered, this is a fairly profound and important ability. For the large work loads at StumbleUpon, the savings could literally be millions of dollars. Some problems simply aren’t approachable on a single machine setup! For highly dynamic data sets, where we frequently read things that were just written, caching in a system, such as memcached, may not help much. HBase holds recently written data in a write buffer. Reads for that data come directly out of memory. This action could completely obviate the need for a caching layer. One example of a highly dynamic data set is event counters. This is a difficult problem because most high-speed solutions tend to be RAM-only for performance (e.g., memcached), while requiring high durability as well. Enter HBase and its incrementColumnValue() call. These mutates are internally treated as any other change, by both logging to disk and buffering up in the write buffer. Reads can come directly from the write buffer, accelerating both and achieving high performance and durability. StumbleUpon harnesses the natural ability of HBase to support event counters in just about every niche of the site—clicks, hits, ads served, and so on. Furthermore, HBase offers a superior choice to typical sharding solutions. Most traditional sharding approaches require a priori assumptions of the key space. This can have surprising performance implications when the hashing function isn’t evenly distributed or when the keys are distributed outside the assumptions of your sharding scheme. HBase takes a data-sized approach to splitting tables into regions; as the data in a region grows and reaches a configured size (currently defaulted to 256 MB), a mid- key is picked from the middle of the data, splitting the region into two roughly equal chunks. Each chunk becomes its own region and now has room to grow. Repeating this procedure thousands of times gives a net result of a table with 2000 roughly equal-sized regions. Figure 12.5 shows the operation of a simplified HBase cluster with three writes and one read concurrently operating on a key space of 256 keys (0x00-0xFF). 7 HDFS writes to multiple data nodes to achieve durability of data, as well as locality-based performance. Recommending the best websites at StumbleUpon 279 ■ The client bootstraps the ROOT table location from ZooKeeper . ■ The ROOT table contains pointers to other META tables that hold user table locations. ■ The request is sent to the region for execution. The client finds the region for the selected operation. The location is cached in the client. ■ As StumbleUpon’s data continues to grow in an uneven manner, we don’t end up with unbalanced shards or regions requiring manual intervention later, a problem frequently encountered by most manual sharding solutions involving RDBMS. LUDICROUS-SPEED HBASE For all the talk about HBase’s advantages, initially we found that the performance of the system was not up to online data servicing. To fix that, Ryan contributed back a large number of internal performance and reliability enhancements to the HBase project. The first major contribution was the HFile format. The previous format had inefficiencies in the index strategy, read paths, and internal APIs. Several issues were identified: ■ Stream-oriented format made caching difficult. ■ Index efficiency was sensitive to the size of the data. ■ Object creation rates were too high. Lots of byte array copies were made. ■ HFile is an immutable file format. Once written, no value in it can change. The reader and writer are separated in code and there are no mutate or update methods available. As most HFiles are hosted on HDFS, it would be impossible anyway because HDFS files are immutable as well. The HFile writer has a straightforward write path, with four elements: ■ Open file, provide compression, block size, and comparator arguments. ■ These never change for the lifetime of the file. ■ Append keys, in comparator sorted order. ■ Optionally append metadata blocks. Any attempt to add keys in nonsorted order results in exceptions. ■ Figure 12.5 HBase easily and automatically “shards.” 280 CHAPTER 12 Case studies ■ Close the file, finalize the index and write the trailers. Useful for additional data, or features, such as Bloom filters . ■ As keys and values are appended to the HFile , the code keeps track of how large the current block is. Once it exceeds the block size specified, it’s flushed and the compres- sion system is reset for the next block. As HFile writer appends a block, an in-memory index of the first key of each block is formed, along with its in-file offset. When the close method is called, the block index is written immediately behind the last block. Optional metadata blocks are appended next, followed by the metadata block index. Finally, a trailer with pointers to the indexes is appended and the file is closed. When a file is opened for reading, the data block index and the meta block index are loaded. They stay in resident for the duration, until the reader object is reclaimed. The index allows for fast seek and reads of blocks of data. To find a key in the file, first the reader does a binary search of the index. Finding a block number, it reads in and decompresses the data block and stores it in the block cache. Code then iterates through the block in memory finding the key, or the closest match. Pointers are then returned, allowing clients a view into the single copy of data. HFile gains its strengths from simplicity in both concept and implementation. The implementation is one file (tests excepted) and is about 1,600 lines for both reader and writer. HFile provided a new internal platform to rewrite the rest of the region server. The internal algorithms for read-merging multiple files into a single-scan result had grown organically over time and needed a fresh look. Jon Gray and Erik Holstad at Streamy.com designed and implemented a brand-new read implementation by adding new delete semantics and restructuring the internal key formats . By using more efficient algorithms and redoing the implementation on the 0-copy HFile, more speed enhancements were gained from the code. Overall speed increases were extremely impressive, ranging from 30 times to 100 times, depending on the particular API call involved. On the low end, scanning a series of rows got a 30 times speed-up. On the high end, single row gets can be up to 100 times faster. With these performance improvements, HBase can truly be labeled as “web ready.” HBASE AND PARALLELISM HBase demonstrates excellent parallel speed-up on read and write workloads. As StumbleUpon has stored so much data in MySQL, insert performance is important. To copy data into HBase, Hadoop jobs with only mappers that read from MySQL and subse- quently write into HBase were written. Running on a 20-node cluster with about 80 times parallelism, aggregate insert performance ranged from 100,000 operations a second up to as much as 300,000 operations a second. The rows involved were about 100 bytes. As impressive as the write performance is, the read performance is exceptional. Using an 80-times-parallel MapReduce read aggregation job, it achieves a total read speed of 4.5 million rows a second. At this rate, reading our largest tables takes less than an hour. The ability to write entire table analytics is a powerful ability that previously didn’t exist. Recommending the best websites at StumbleUpon 281 All the machines involved were dual quad core Intels, with 16 GB of RAM. Each node had two SATA disks , each 1 TB in size. These relatively modest and standard nodes provide an excellent level of performance, and the cluster only performs better with more. 12.3.3 More Hadoop at StumbleUpon At StumbleUpon , we subscribe to the mantra of “Log early, log often, log everything.” No piece of data is too small or too noisy to be useful down the road. Hadoop excels in this traditionally strong area for distributed processing: log-and-click collection com- bined with analysis. StumbleUpon harnesses this natural aptitude of Hadoop for a variety of analysis tasks, including Apache log file collection and user-session analysis. As an example, a basic need for any web product in the days of search engine optimization experts and “black hat” attackers is to look back at a combination of web browser user agent strings combined with the (apparent) originating IP address and action context. Now imagine needing to do this across a fleet of web server frontends, millions of users, and billions of clicks. Scribe, 8 a Facebook project made public, is a platform for aggregating real-time streamed log data in such a context. The service is failure tolerant at both the machine and network level and easily integrates into just about any infrastructure. StumbleUpon uses Scribe to collect data directly into HDFS where it’s reviewed and processed by a number of systems. A combination of Cascading and plain MapReduce- based analysis jobs extract data from the logs for vanilla statistics (such as click counts), while more sophisticated consumers feed data into real-time feedback systems based around BerkeleyDB and TokyoCabinet . A second set of systems use this streamed data for search index updates and thumbnail generation. Figure 12.6 illustrates several data processing modules around Hadoop . 8 http://github.com/facebook/scribe. Figure 12.6 StumbleUpon data collection, analysis, and storage using Hadoop 282 CHAPTER 12 Case studies We obtained an illustrative result by processing 10 GB of standard Apache log files with the Cascading log analysis example.9 Using Hadoop 0.19.1, Cascading 1.0.9, and the previously mentioned node configuration, we obtained the number of Apache hits per minute with this example by bucketing the hits in MapReduce jobs. We wrote a naive single-node Perl hash-based program as an example of a typical quick solution a sysadmin may create. The results shown in table 12.3 confirm that our results easily achieve linear (or better) speed-up with the simple addition of more nodes to the clus- ter. Times are the average of 10 mixed executions, to allow for variances. Table 12.3 Apache log processing with Cascading System Performance measure Result 1 Node Runtime 21m46s Sec/MB 0.127 Sec/MB/Node 0.127 3 Nodes Runtime 8m3s Sec/MB 0.0471 Sec/MB/Node 0.0157 15 Nodes Runtime 1m30s Sec/MB 0.00878 Sec/MB/Node 0.000585 Naive Perl Runtime 42m49s Sec/MB 0.251 Sec/MB/Node 0.251 We see that even the single-node Cascading solution achieves double the throughput of the naive Perl application due to the intelligent segmentation and bucketing built in to the MapReduce framework versus the effect of keeping all data mapped to a single Perl hash. Given familiarity with Cascading, you may also consider the Perl code more complex to optimize (and maintain) to boot! To wit, StumbleUpon uses the native map and reduce functionality in Hadoop and related products, including Nutch and custom-written content surveyors, to perform this data retrieval, analysis, and storage. Keeping the resultant data close to the processing pipeline maximizes our data locality benefits. Putting it all together, StumbleUpon has taken the maximum advantage of the vast power the MapReduce paradigm unlocks by adopting and extending Hadoop, HDFS, and HBase. We’re excited to help lead the future of distributed processing. 12.4 Building analytics for enterprise search —IBM’s Project ES2 Contributed by VUK ERCEGOVAC , RAJASEKAR KRISHNAMURTHY , SRIRAM RAGHAVAN , FREDERICK REISS , EUGENE SHEKITA , SANDEEP TATA , SHIVAKUMAR VAITHYANATHAN , and HUAIYU ZHU In contrast with the radical advances in web search over the last several years, search over enterprise intranets has remained a difficult and largely unsolved problem. Based 9 http://code.google.com/p/cascading/wiki/ApacheLogCascade. Building analytics for enterprise search —IBM’s Project ES2 283 on a study of the IBM intranet, Fagin et. al. [1] highlighted some critical differences between the search problem on the intranet and that on the web. They observed that an overwhelming majority of the queries in the intranet are “navigational.” They have a small set of correct answers [2,3]. For instance, a manual examination of the top 6,500 queries (as of July 2008) on the IBM intranet revealed that more than 90 percent of those queries were navigational. Several enterprise-specific factors complicate the task of finding the “correct” answers for these queries: ■ Absence of an economic incentive for content creators to make their pages easily discoverable (in contrast with the presence of such incentives on the web) ■ The use of enterprise-specific vocabulary, abbreviations, and acronyms, in the search queries and in the intranet pages ■ The fact that the same query can have a different “correct” answer depending on the location and organizational role of the person issuing the query (of particu- lar importance for corporations like IBM with employees and locations in over 80 countries) From earlier efforts at IBM [4], we learned that these problems are difficult to overcome using traditional information retrieval techniques. Subsequently, in [5], we proposed an approach consisting of detailed offline analyses to pre-identify navigational pages and the use of a special-purpose navigational index. We demonstrated the viability of this approach through experiments over a 5.5-million-page corpus from the IBM in- tranet. The system in [5], uses a mix of proprietary platforms and relational databases. We have, since, crawled a much larger portion of the IBM intranet , having discovered over 100 million URLs and indexed over 16 million documents. In order to tackle these sizes and beyond, and having learned from previous efforts [4,5], we have devel- oped ES2 —a scalable, high-quality search engine for the IBM intranet . ES2 is based on the analytics described in [5], but it leverages a number of open source platforms and tools, such as Hadoop, Nutch , Lucene , and Jaql. 10 In principle, the Nutch crawler, the Hadoop MapReduce framework, and the Lucene indexing engine provide a full suite of software components for building a complete search engine. But, to truly address the challenges described earlier, it’s not sufficient to merely stitch these systems together. We describe how to use sophisticated analysis and mining of the crawled pages, and special-purpose navigational indexes in conjunction with intelligent query processing to ensure effective search quality. To understand how these elements come together, we now examine some illustrative search queries and their corresponding results on ES2. See figure 12.7. Figure 12.7 shows the result of running the query idp on ES2 . The IDP is an acronym for Individual Development Plan, a web-based HR application in IBM to assist in tracking employee career development. The first two results returned by ES2 represent two different URLs that both allow the user to launch the IDP web application. The third 10 http://code.google.com/p/jaql/. 284 CHAPTER 12 Case studies result entry is, in fact, a set of pages describing the IDP process, one per country, that have been grouped together (indicated visually through indentation and the presence of a globe icon). We make the following observations: 1 The first result in figure 12.7 doesn’t have the word idp in the title and indeed not even in the content of the page. ES2 was able to associate this page with the word idp by extracting the term from the URL http://w3.ibm.com/hr/idp/. For the second result in figure 12.7, besides idp in the URL, we also extract idp from the title by applying a regular expression pattern that explicitly looks for titles ending in phrases such as “Launch Page,” “Portal,” “Main Page,” and so on. In ES2, we use several hundred such carefully crafted patterns applied to the URL, title, META headers, and various other features of a web page to detect and associate index terms with navigational pages. Section 12.4.3 describes how we execute such analysis, known as Local Analysis , in parallel on a Hadoop cluster. 2 Our current crawl of the IBM intranet has close to 500 pages that have a URL containing idp and over a 1,000 pages that have either idp or individual development plan in the title. To narrow down the result to the two specific URLs shown in figure 12.7, we use a complex set of analysis algorithms as part of a Figure 12.7 Illustrative search results in ES2 Building analytics for enterprise search —IBM’s Project ES2 285 process known as Global Analysis . Section 12.4.3 describes how to implement such analysis in ES2 using the Hadoop framework. 3 Notice that in all the result pages grouped as part of the third result in figure 12.7, the text Individual Development Plan in the title has been highlighted to indicate a match against the query term idp . To accomplish this match, the following two steps took place during offline analysis: (1) the phrase Individual Development Plan was extracted as part of local analysis using patterns applied to the title and (2) during indexing, the extracted phrase was recognized to be the expansion of the acronym idp and resulted in the term idp being added to the index as well. In general, we employ a process known as variant generation whereby multiple variants of the terms extracted through local analysis are generated and added to the index. ES2 employs a suite of variant generation strategies—from simple n-grams over the extracted phrase to more sophisticated ones. In the interest of space we don’t describe these algorithms in detail. 4 Finally, to enable results to be customized based on the search users’ geography and to support the type of result grouping shown in figure 12.7, we label each page in the intranet with a specific geography (country, region, and/or IBM location). In ES2, this labeling is accomplished using a rule-driven classifier that uses a number of page features extracted during local analysis. The examples in figure 12.7 motivate the vital role of offline analysis and variant gen- eration in ES2 . Each page in the ES2 collection is pushed through multiple logical workflows, each consisting of a local analysis phase, a global analysis phase, and an appropriate variant generation strategy. The output of a workflow is some subset of the input pages along with a set of index terms. Depending on the particular extraction patterns and variant generation rules, the output of two different workflows will have correspondingly different “precision” characteristics. For example, a workflow consist- ing of careful extraction of a person’s name from the title followed by name-specific variant generation is likely to yield much higher-quality answers than a workflow that only generates all possible n-grams of the title of a page. The creation of an index structure consisting of the output of multiple workflows with different precision characteristics is only half the story. To fully leverage such an index, ES2 employs a sophisticated runtime query processing strategy. A discussion of the runtime component of ES2 is beyond the scope of this case study. We restrict our attention to the offline analysis workflows and their implementation on Hadoop. 12.4.1 ES2 architecture We assume that readers are broadly familiar with Hadoop and Nutch.11 Nutch is an open source crawler implemented on the Hadoop MapReduce platform for web crawling. 11 http://nutch.apache.org/. 286 CHAPTER 12 Case studies ES2 also uses Jaql ,12 a data flow language designed for JSON (a popular semi- structured data model). Jaql provides a Unix pipes-like syntax to connect multiple stages of processing over semistructured JSON data. The ES2 workflow involves invoking multiple algorithms for local analysis, global analysis, and variant generation before inserting data into indexes. Without adequate data management support, this complex multistage workflow quickly becomes overwhelming. To address this problem, ES2 uses JSON to represent its data and Jaql to specify the workflow (see figure 12.8). Figure 12.8 shows the architecture of ES2. There are six components in ES2: the crawler, local analysis, global analysis, variant generation and indexing, background mining, and search runtime. ES2 uses an enhanced version of Nutch (version 0.9)—a scalable open source crawler based on the Hadoop platform. In addition, ES2 also gathers information from IBM’s social bookmarking service (called Dogear ). Much like delicious.com, Dogear contains various URLs that have been bookmarked by the IBM community along with a collection of tags associated with each URL. The tags associated with the URLs contain valuable clues about the page, and ES2 uses this information in building its indexes. All the stages use a common distributed filesystem, HDFS, for both input and output. Local analysis processes each page to extract features about the pages and stores the results as JSON objects in HDFS. ES2 uses Jaql to push each page through the rest of the pipeline, transforming the data as needed at each Nutch Crawler IBM Intranet Local Analysis (Low - level analytics on individual page features such as url , title, ..) Global Analysis (Analytics on groups of pages produced by Local analysis) Distributed Data Store (HDFS) Variant Generation + Indexing (Intelligent generation of index terms from page features) Search Runtime (Interpretation & Ranking) Mining (Finding acronyms, regular expressions, and rules to support LA) ES2 Analytics Hadoop map-reduce JAQL Figure 12.8 ES2 Architecture 12 http://code.google.com/p/jaql/. Building analytics for enterprise search —IBM’s Project ES2 287 stage. Jaql queries are used to bring together the results of different local analyses and invoke global analysis. Jaql is also used to invoke the variant generation and indexing workflow using the outputs of local and global analyses. The indexes are periodically copied to a different set of machines that serve user queries. Although not part of the main workflow, ES2 periodically executes several mining and classification tasks. Examples of this include algorithms to automatically produce acronym libraries, regular expression libraries [6], and geo-classification rules. 12.4.2 ES2 crawler ES2 uses Nutch version 0.9. A primary data structure in Nutch is the CrawlDB : a key- value set where the keys are the URLs known to Nutch and the value is the status of the URL. The status contains metadata about the URL, such as the time of discovery, whether it has been fetched, and so on. Nutch is architected as a sequence of three MapReduce jobs: ■ Generate—In this phase, a fetch list is generated by scanning the input key/value pairs (from CrawlDB) for URLs that have been discovered, but not fetched. A common choice in generating this list is to select the top k unfetched URLs using an appropriate scoring mechanism (k is a configuration parameter in Nutch). ■ Fetch—In this phase, the pages associated with the URLs in the input fetch list are fetched and parsed. The output consists of the URL and the parsed representa- tion of the page. ■ Update—The update phase collects all the URLs that have been discovered by parsing the contents of the pages in the fetch phase and merges them with the CrawlDB. The pages fetched in each cycle of generate-fetch-update are referred to as a segment . Out of the box, the first problem we encountered was crawl speed. Nutch’s crawl rate was under three pages per second—far less than the network bandwidth available to the cluster. A deeper problem we encountered after a sample crawl of 80 million pages was that the quality of discovered pages was surprisingly low. In this section, we identify the underlying reasons for both these problems and describe the enhancements made to Nutch to adapt it to the IBM intranet. MODIFICATIONS FOR PERFORMANCE Nutch’s design was aimed at web crawling . When using it to crawl the IBM intranet, we observed multiple performance bottlenecks . We discovered that the reason for the bottlenecks was that the enterprise intranet contains far fewer hosts than the web, and some of the design choices made in Nutch assume a large number of distinct hosts. We describe two ways in which this problem manifests itself, and the approach used in ES2 to adapt Nutch’s design for the enterprise. A major performance bottleneck in the fetch phase, called long tail problem , exhibits the following behavior. The crawl rate in the early part of the fetch phase is relatively high (typically dozens of pages a second). But this deteriorates relatively quickly to less than a page per second, where it remains until completion of the segment. A 288 CHAPTER 12 Case studies quick examination revealed that this behavior is heavily influenced by the host with the largest number of URLs in the fetch list. You can understand this by observing that the fetch rate in Nutch is controlled by two parameters: the number of distinct hosts in the fetch list that Nutch can concurrently crawl from, and the duration for which Nutch waits before making consecutive requests to the same host. A straightforward solution to the long tail problem is to restrict the number of URLs for a particular host in the fetch list. Unfortunately, this is not sufficient because not all host servers are identical, and the time required to fetch the same number of pages from different hosts can be dramatically different. We added a time-shutoff parameter that terminates the fetcher after a fixed amount of time as an engineering fix to this problem. While this terminates the fetch phase early (and fewer pages are retrieved in total in the segment), by avoiding the slow tail phase, we sustain a higher average crawl rate. In practice, we observed that by appropriately setting this shutoff parameter, the average crawl rate could be improved to nearly three times the original crawl rate. Ideally, the current fetch rate should determine such a shutoff; this unfortunately requires pooling information across map tasks and can’t easily be performed in Hadoop today. A main-memory data structure in the fetcher causes a different performance bottleneck. The fetcher works by first creating a set of queues where each queue stores URLs for a particular host—we call this data structure FetchQueues . A fixed amount of memory is allocated to FetchQueues to be shared across the individual queues. The fetcher reads the URLs to be fetched from its input and inserts them into FetchQueues until it exhausts the allocated memory. Worker threads assigned to each queue in FetchQueues concurrently fetch pages from different hosts as long as their queues are non-empty. The bottleneck arises because URLs in the input are ordered by host (this is an artifact of the generate phase) and the fetcher exhausts the memory allocated to FetchQueues with URLs from very few hosts. Such a design is appropriate for crawling a large number of hosts on the web as each host in the fetch list would then have only a few URLs. In the enterprise, host diversity is limited to a few thousand at best. As a result, few worker threads are actively fetching from FetchQueues, leading to severe under-utilization of resources. We address this problem by replacing FetchQueues with a disk-based data structure without any limits on the total size. This allows the fetcher to populate FetchQueues with all the URLs in the input, thereby keeping the maximum possible number of worker threads active. This simple change improved the fetch rate several fold. 12.4.3 ES2 analytics Much of the complexity and power in ES2 lies in its analytics. In this section, we briefly describe the different algorithms, paying special attention to the design choices made in mapping these algorithms onto Hadoop. LOCAL ANALYSIS In local analysis, each page is individually analyzed to extract clues that help decide whether that page is a candidate navigational page . In ES2, five different local analysis Building analytics for enterprise search —IBM’s Project ES2 289 algorithms, namely TitleHomePage, PersonalHomePage, URLHomePage, AnchorHome, and NavLink are used. These algorithms use rules based on regular expression patterns, dictionaries, and information extraction tools [7] to identify candidate navigational pages. For instance, using a regular expression like “\ A\ W*(.+)\s” (Java regular expression syntax), the PersonalHomePage algorithm can detect that a page with a title “G. J. Chaitin’s Home” indicates that this is the home page of G. J. Chaitin. The algorithm outputs the name of a feature (“Personal Home Page”) and associates a value with this feature (“G. J. Chaitin”). The next section describes the impact of redi- rections on local analysis and discusses a solution. REDIRECTION RESOLUTION Many sites in IBM’s intranet employ redirection for updating, load balancing, upgrad- ing, and handling internal reorganizations. Unfortunately, redirections can cause complications in the local analysis algorithms. For instance, URLHomePage uses the text of the URL to detect a candidate navigational page. After redirection, the target URL may not contain the same features as the original URL. As an illustrative ex- ample, consider the URL http://w3.can.ibm.com/hr/erbp. Local analysis algorithms can correctly identify this URL as the home page for the Employee Referral Bonus Program (ERBP) using clues from the URL. But this URL gets redirected to a Lotus Domino server at http://w3-03.ibm.com/hr/hrc.nsf/3f31db8c0ff0ac90852568f7006d 51ea/ac3f2f04ba60a6d585256d05004cef97?OpenDocument, where a Lotus Domino database serves information about the Employee Referral Bonus Program. The clues in the source URL are no longer available in the target, and the local analysis algo- rithm can no longer identify this page as navigational. To prevent this, ES2 resolves all redirections, collects the set of URLs that lead to the target page through redirections, and provides local analysis with the appropriate URLs. To track redirections, we modified Nutch to tag every page that was a target of redirection with the source URL. Consider figure 12.9. The crawler follows redirections from a page A to page B, and from page B to arrive at page C. We track these redirections by tagging pages B and C with the source URL, A. This tag is stored as a metadata field in the segment file . A segment file is a key/value set Figure 12.9 Resolving redirections URL: A URL: B Source: A URL: C Source: A Redirect Redirect URL: C Sources: A, B After Redirection Resolution 290 CHAPTER 12 Case studies where the key is the URL of a page and the value is contents of the page (along with additional metadata fields). Listing 12.2 (called ResolveSimple ) outlines the map and reduce functions that are used to resolve redirections on a segment and invoke local analysis. The map phase outputs the source URL and the page contents. The reduce phase brings all the pages with the same source URL into a single group. In the preceding example of figure 12.9, the common source URL for pages A, B, and C is A. The target page in this group (C) is then passed to local analysis along with the other URLs in the group—A and B. Listing 12.2 ResolveSimple Map (Key: URL, Value: PageData) if PageData.SourceURL exists then Ouput [PageData.SourceURL, PageData] else Output [URL, Pagedata] end if End Reduce (Key: URL, Values: Pageset) Let URLset = Set of all URLs in Pageset Let page = Target of redirection in Pageset result = LocalAnalysis(page, URLset) output [page.URL, result] End HADOOP IMPLEMENTATION In ResolveSimple, local analysis is invoked in the reducer . This requires Hadoop to pass along the contents of each page from the map phase to the reduce phase. This involves sorting and moving a large amount of data across the network. To avoid this, we modify ResolveSimple (listing 12.2) and separate the task of redirection resolution and the local analysis so that the algorithms in local analysis are run in the map phase. This allows the local analysis computation to be colocated with the data, and therefore results in significant performance improvement. We have outlined the modified algorithm, called Resolve2Step, in listing 12.3. In the map phase of this algorithm, we only pass the metadata along and the page content (which accounts for a majority of the data volume) is projected out. In the reduce phase of ResolveSimple, we output a table with two columns: the first column is the URL of the target page in the group of pages, and the second column is the set of URLs to be associated with the page when it’s submitted to local analysis. Listing 12.3 Resolve2Step 1: Resolve Redirections Map (Key: URL, Value: Page) if PageData.SourceURL exists then Ouput [PageData.SourceURL, PageData.metadata] else Building analytics for enterprise search —IBM’s Project ES2 291 Output [URL, Pagedata.metadata] end if End Reduce (Key: URL, Values: Pageset) Let URLset = Set of all URLs in Pageset Let page = Target of redirections in Pageset output [page.URL, URLset] End 2: Run Local Analysis Map (Key: URL, Value: Page) Load resolveTable from output of previous step if needed Let URLSet = resolveTable[URL] result = LocalAnalysis(page, URLset) output [page.URL, result] End If a URL results in a redirection, we don’t add an entry for it in this table. Table 12.4 shows an example of such a table. The chain of redirections shown in the preceding figure 12.9 results in the first row in table 12.4. In a subsequent map-only job for local analysis, the map tasks read the redirection table into memory. This table is fairly small for typical segments and easily fits in memory. For each URL in the input segment, the mapper looks up the table if it finds a non-empty entry. It passes these URLs on to local analysis. By invoking local analysis in the map phase, Resolve2Step avoids the transfer of the page contents over the network to the reducers as in ResolveSimple. We executed both algorithms for local analysis on a segment of around 400,000 pages on the cluster using eight nodes. Whereas ResolveSimple completed in about 22 minutes, Resolve2Step took only 7 minutes. In order to understand how Resolve2Step scales, we ran this algorithm on the same segment (400,000 pages) and varied the cluster size from one to eight servers. The times are shown in figure 12.10. The speed- up graph shows that for the early part of the curve, we get linear scaling; the benefits of adding more nodes decreases after this point. This is because the input consists of only a single segment of 400,000 pages. Hadoop is unable to efficiently divide this task at a finer granularity. We’ll see in the next section that with larger input data sets, Hadoop can efficiently divide the task and provide linear scaling. GLOBAL ANALYSIS The local analysis tasks described in the previous section identify candidate navi- gational pages by extracting relevant features from each page. But as described in section 12.4.1, the same navigational feature can be associated with multiple pages. Consider the case where homepage authors use the same title for many of their web pages. For example, “G. J. Chaitin home page” is the title for many pages on G. J. Chaitin’s website. Local analysis for personal home pages considers all such pages to be candidates. ES2 uses global analysis to determine an appropriate subset of pages Table 12.4 Resolution table in Resolve2Step URL Sources C {A,B} … … X {Y} … … 292 CHAPTER 12 Case studies as navigational. [5] describes two algorithms: site root analysis and anchor text analysis . We briefly review these algorithms and describe how Jaql is used to implement these algorithms on large data sets. GLOBAL ANALYSIS ALGORITHMS Each global analysis task takes as input a set of pages and the associated features discov- ered during local analysis. Listing 12.4 shows an example JSON record corresponding to the page titled “G. J. Chaitin home page.” The fields produced by local analysis are in the record titled “LA” and those produced by global analysis are in “GA.” Listing 12.4 Example JSON output from global analysis and Jaql query [..., {docid: 1879495641814943578, url: "http://w3.watson.ibm.com/~chaitin/index.html", title: "G J Chaitin Home Page", ... LA: { personalHomepage: {name: G J Chaitin, begin: 0, end: 11}, geography: {countries: "USA", ...} ...}, GA: { personalHomepageSiteRootAnalysis: {marked: true, ...}, ...} }, ...] $alldocs = file "laDocs.json"; $results = file "phpGADocs.json"; $alldocs -> filter not isnull($.LA.personalHomepage.name) -> partition by $t = $.LA.personalHomepage.name |- SiteRootAnalysis($t, $) -| -> write $results; Figure 12.10 Resolve2Step for one segment (400K Pages) has a performance speed-up linear in cluster size.Cluster Size (# Servers) 1 2 3 4 5 6 7 8 Elapsed Time (mins) 0 5 10 15 20 25 30 35 40 45 Speed-up 0 1 2 3 4 5 6 7 8 Time Speed-up Building analytics for enterprise search —IBM’s Project ES2 293 The two algorithms of global analysis are ■ Site root analysis —Both algorithms are used to group candidate pages and iden- tify a set of representative pages. Given a collection of candidate pages, it’s first partitioned by the feature of interest, for example, PersonalHomePage. For each group, a forest of pages is constructed where each URL is a node in the forest, relating the two URLS A and B as parent and child if A is the longest prefix of B. (Shorter prefixes are higher ancestors.) The forest is pruned using some com- plex logic that may involve inputs from other local analysis algorithms, the de- tails of which are beyond the scope of this case study. We use site root analysis not only for the output from PersonalHomePage, but also for TitleHomePage (e.g., pages titled “Working at Almaden Research Center” or “IT Help Central”). ■ Anchor text analysis —This algorithm collects all the anchor text for each page by examining all the pages that point to it. The aggregated anchor text is processed to pick a set of representative terms for that URL. For further details on this algorithm, see [5]. HADOOP IMPLEMENTATION In global analysis, first, a merge step joins together the results of local analysis on the main crawl and the tags for the URLs collected from Dogear. This is followed by a de- duplication step where duplicate pages are eliminated. Each global analysis task then involves some standard data manipulation (e.g., partitioning, filtering, joining) in con- junction with some task-specific user-defined function, such as URL forest generation and pruning. Jaql is used to specify these tasks at a high level, and execute them in parallel using Hadoop. Consider the Jaql query in listing 12.4 used for the global analysis on PersonalHomePage data. The first two lines specify the input and output files. The input is assumed to be a JSON array—in this case, an array of records—each record representing a page and the associated results from local analysis. The third line is the start of a Jaql pipe: pages flow from the input file, referred to by $allDocs , to subsequent operators. The connection between pipe operators is denoted by ->. Following the input, the “filter” operator produces a value when its predicate evaluates to true. In the example, only pages that have a local analysis (LA) field, a PersonalHomePage field, and a non-null name are output to the next operator. The $ is a variable that refers to the current value in the pipe. The filtered pages are partitioned according to name. For each partition, the user-defined function SiteRootAnalysis is evaluated. The function takes as input the partitioning field $t (a variable for name), and all pages in the partition ($). Finally, the annotated pages are written to $results output file. Jaql evaluates the query shown in the preceding listing 12.4 by translating to a MapReduce job and submitting the job to Hadoop for evaluation. In this example, the map stage filters pages and extracts the partitioning key. The reduce stage evaluates the SiteRootAnalysis function per partition and writes the output to a file. In general, 294 CHAPTER 12 Case studies Figure 12.11 Global analysis times Jaql automatically translates a collection of pipe definitions into a directed acyclic graph of MapReduce jobs. In figures 12.11 and 12.12, we present how the global analysis task scaled on a collection of 16 million documents using Jaql and Hadoop as the cluster size was increased from two to eight servers. Figure 12.11 shows detailed elapsed times for each stage involved after local analysis through the end of global analysis. Figure 12.12 shows that as servers were added to the cluster, the total time to evaluate merge, dedup, and global analysis improved proportionally. MINING TASKS ES2 builds acronym libraries , regular expression patterns , and geo-classification rules automatically using the crawled data in background mining tasks. Recall that 0 1 2 3 4 5 6 7 8 9 2 4 6 8 C l uster S i ze (# S e rvers ) Speed-up Figure 12.12 Global analysis speed-up Building analytics for enterprise search —IBM’s Project ES2 295 local analysis algorithms use these resources. Periodically, the local analysis is rerun on all the pages after updating these resources. As an example, we provide a brief description of the acronym mining algorithm and the geo-classification algorithm used in ES2 below. Acronym mining is a computationally intensive task that benefits from a parallel implementation on Hadoop. The algorithm used in ES2 is adapted from [8]. It works by examining the text of each page to first identify patterns of the form “longForm (shortForm)” or “shortForm (longForm)”. After identifying acronyms and their candidate longForms, the map function outputs [shortForm, longForm] as a key/value pair. The reduce function gathers all the possible longForms together for a given shortForm and ranks them by frequency before producing the output. The reduce function merges together longForms that are nearly identical, such as “Individual Development Plan” and “Individual Development Plans” as longForms for “idp .” You can see the map and reduce functions for the mining task in listing 12.5. This task is easily parallelized on a cluster. Listing 12.5 Mining for acronyms Map (Key: URL, Value: PageText) Identify all (shortForm, longForm) pairs in the text For each instance, output [shortForm, longForm] End Reduce (Key: shortForm, Values: longForms) Canonicalize longForms that differ slightly Compute frequency of each longForm Output longForms in sorted order End Figure 12.13 shows the running time of this algorithm on a sample of 10 million docu- ments as the number of nodes is increased from two to eight. As can be seen, the overall task completes in less than 25 minutes even with two nodes. But we see that this task doesn’t scale linearly with the size of the cluster. We suspect that this is because the input data is fragmented over several segments, and Hadoop chooses to split this job into a large number of tasks in the map phase which imposes a large, fixed over- head independent of the cluster size. We’re investigating methods to overcome these performance issues. The goal of the geo-classification task is to label each page on the intranet with the country, region, and/or IBM location for which the page is most relevant. Numerous factors make this task particularly nontrivial. For instance, many new business processes and web applications within IBM are first deployed in the U.S. before being extended to other countries and regions across the world. Site administrators responsible for developing the content for the subsequent rounds of deployment often use the U.S. page as a starting point and make appropriate edits to tailor the pages to their respective countries. But when performing this customization, quite often the administrators don’t edit the corresponding HTML meta headers that 296 CHAPTER 12 Case studies convey the locale, language, and other geo-specific information.13 A simple-minded classifier that directly exploits information from the meta headers is therefore prone to errors. In ES2, we’re currently employing a complex rule-driven classifier consisting of a robust set of manually created rules over a small set of page features (e.g., presence of a country name in the title, a country code in the URL, etc.). While the rules have been hand-tuned for high precision, the recall—the number of pages for which the classifier is able to assign non-empty geo labels—is limited. Improvement in recall requires the use of significantly more features from a page than is used by our current classifier. But manually developing accurate rule sets over these larger feature sets is extremely laborious. We’re now in the process of developing a scalable mining algorithm to automatically “induce” additional classification rules over these new features, given the high-quality rule set already available today. The use of a platform like Hadoop is critical to scale our mining algorithms to millions of pages, each with several hundred features. 12.4.4 Conclusions We described the architecture of ES2—a scalable enterprise search system developed at IBM using open source components, such as Nutch, Hadoop, Lucene, and Jaql. We also outlined the changes we needed to make to Nutch for the purposes of crawl- ing the enterprise. We mapped the local and global analysis algorithms from [5] on to Hadoop. In implementing a complex workflow involving crawling, local analysis, global analysis, and indexing, we found JSON to be a convenient data format and Jaql to be an extremely powerful tool. In summary, we believe that Hadoop, Nutch, Lucene, and Jaql constitute a powerful set of tools with which sophisticated, scalable systems like ES2 can be built. Cluster Size (# Servers) 0 2 4 6 8 Elapsed Time (mins) 5 10 15 20 25 30 Speed-up 0 1 2 3 4 5 6 7 8 Time Speed-up Figure 12.13 Acronym mining has a speed-up linear in cluster size. 13 Note that the meta headers are intended for consumption by browsers and crawlers and not visible when the page is rendered. Building analytics for enterprise search —IBM’s Project ES2 297 12.4.5 References Fagin, R., R. Kumar, K. S. McCurley, J. Novak, D. Sivakumar, J. A. Tomlin, and D. P. Williamson. “Searching the workplace web.” In WWW, pages 366–375, 2003. Broder, A. “A taxonomy of web search.” SIGIR Forum, 36(2):3–10, 2002. Hawking, D. “Challenges in enterprise search.” In ADC, pages 15–24, 2004. Fontura, M., E. J. Shekita, J. Y. Zien, S. Rajagopalan, and A. Neumann. “High performance index build algorithms for intranet search engines.” In VLDB, pages 1158–1169, 2004. Zhu, H., S. Raghavan, S. Vaithyanathan, and A. Löser. “Navigating the intranet with high precision.” In WWW, pages 491–500, 2007. Li, Y., R. Krishnamurthy, S. Raghavan, S. Vaithyanathan, and H. V. Jagadish. “Regular expression learning for information extraction.” In EMNLP, 2008. Reiss, F., S. Raghavan, R. Krishnamurthy, H. Zhu, and S. Vaithyanathan. “An algebraic approach to rule-based information extraction.” In ICDE, pages 933–942, 2008. Schwartz, A. S., and M. A. Hearst. “A simple algorithm for identifying abbreviation definitions in biomedical text.” In Pacific Symposium on Biocomputing, pages 451– 462, 2003. 298 Appendix HDFS file commands This appendix lists the HDFS commands for managing files. They’re in the form of hadoop fs -cmd where cmd is the specific file command and is a variable number of argu- ments. You can see the command usage in the following convention. Parameters inside brackets ([]) are optional and ellipsis (. . .) means the optional parameter can be repeated. FILE is for filenames whereas PATH can be either filenames or directory names. SRC and DST are path names but they function specifically as source and destination, respectively. LOCALSRC and LOCALDST are further required to be on the local filesystem. Command Usage and description cat hadoop fs –cat FILE [FILE ...] Displays the files’ content. For reading compressed files, you should use the text command instead. chgrp hadoop fs –chgrp [-R] GROUP PATH [PATH ...] Changes the group association for files and directories. The -R option applies the change recursively. The user must be the files’ owner or a superuser. See section 8.3 for more background information on the HDFS file permission system. APPENDIX HDFS file commands 299 Command Usage and description chmod hadoop fs –chmod [-R] MODE[,MODE ...] PATH [PATH ...] Changes the permissions of files and directories. Similar to its Unix equivalent, MODE can be a 3-digit octal mode, or {augo}+/-{rwxX}. The -R option applies the change recursively. The user must be the files’ owner or a superuser. See section 8.3 for more background information on the HDFS file permission system. chown hadoop fs –chown [-R] [OWNER][:[GROUP]] PATH [PATH ...] Changes the ownership of files and directories. The -R option applies the change recursively. The user must be a superuser. See section 8.3 for more background information on the HDFS file permission system. copyFromLocal hadoop fs –copyFromLocal LOCALSRC [LOCALSRC ...] DST Identical to put (copy files from the local file system). copyToLocal hadoop fs –copyToLocal [-ignorecrc] [-crc] SRC [SRC ...] LOCALDST Identical to get (copy files to the local file system). count hadoop fs –count [-q] PATH [PATH ...] Displays the number of subdirectories, number of files, number of bytes used, and name for all files/directories identified by PATH. The -q option displays quota information. cp hadoop fs –cp SRC [SRC ...] DST Copies files from source to destination. If multiple source files are specified, destination has to be a directory. du hadoop fs –du PATH [PATH ...] Displays file sizes. If PATH is a directory, the size of each file in the directory is reported. Filenames are stated with the full URI protocol prefix. Note that although du stands for disk usage, it should not be taken literally, as disk usage depends on block size and replica factors. dus hadoop fs –dus PATH [PATH ...] Similar to du, but for a directory, dus reports the sum of file sizes in aggregate rather than individually. expunge hadoop fs –expunge Empties the trash. If the trash feature is enabled, when a file is deleted, it is first moved into the temporary .Trash/ folder. The file will be permanently deleted from the .Trash/ folder only after a user-configurable delay. The expunge command forcefully deletes all files from the .Trash/ folder. Note that as long as a file is in the .Trash/ folder, it can be restored by moving it back to its original location. get hadoop fs –get [-ignorecrc] [-crc] SRC [SRC ...] LOCALDST Copies files to the local filesystem. If multiple source files are specified, local destination has to be a directory. If LOCALDST is -, the files are copied to stdout. HDFS computes a checksum for each block of each file. The checksums for a file are stored separately in a hidden file. When a file is read from HDFS, the checksums in that hidden file are used to verify the file’s integrity. For the get command, the -crc option will copy that hidden checksum file. The -ignorecrc option will skip the checksum checking when copying. 300 APPENDIX HDFS file commands Command Usage and description getmerge hadoop fs –getmerge SRC [SRC ...] LOCALDST [addnl] Retrieves all files identified by SRC, merges them, and writes the single merged file to LOCALDST in the local filesystem. The option addnl will add a newline character to the end of each file. help hadoop fs –help [CMD] Displays usage information for the command CMD. If CMD is omitted, it displays usage information for all commands. ls hadoop fs –ls PATH [PATH ...] Lists files and directories. Each entry shows name, permissions, owner, group, size, and modification date. File entries also show their replication factor. lsr hadoop fs –lsr PATH [PATH ...] Recursive version of ls. mkdir hadoop fs –mkdir PATH [PATH ...] Creates directories. Any missing parent directories are also created (like Unix mkdir –p). moveFromLocal hadoop fs –moveFromLocal LOCALSRC [LOCALSRC ...] DST Similar to put, except the local source is deleted after it’s been successfully copied to HDFS. moveToLocal hadoop fs –moveToLocal [-crc] SRC [SRC ...] LOCALDST Displays a “not implemented yet” message. mv hadoop fs –mv SRC [SRC ...] DST Moves files from source(s) to destination. If multiple source files are specified, destination has to be a directory. Moving across filesystems is not permitted. put hadoop fs –put LOCALSRC [LOCALSRC ...] DST Copies files or directories from local system to destination filesystem. If LOCALSRC is set to -, input is set to stdin and DST must be a file. rm hadoop fs –rm PATH [PATH ...] Deletes files and empty directories. rmr hadoop fs –rmr PATH [PATH ...] Recursive version of rm. setrep hadoop fs –setrep [-R] [-w] REP PATH [PATH ...] Sets the target replication factor to REP for given files. The -R option will recursively apply the target replication factor to files in directories identified by PATH. The replication factor will take some time to get to the target. The -w option will wait for the replication factor to match the target. stat hadoop fs –stat [FORMAT] PATH [PATH ...] Displays “statistical” information on files. The FORMAT string is printed exactly but with the following format specifiers replaced. %b Size of file in blocks %F The string “directory” or “regular file” depending on file type %n Filename %o Block size %r Replication %y UTC date in yyyy-MM-dd HH:mm:ss format %Y Milliseconds since January 1, 1970 UTC APPENDIX HDFS file commands 301 Command Usage and description tail hadoop fs –tail [-f] FILE Displays the last one kilobyte of FILE. test hadoop fs –test –[ezd] PATH Performs one of the following type checks on PATH: -e PATH existence. Returns 0 if PATH exists. -z Empty file. Returns 0 if file length is 0. -d Returns 0 if PATH is a directory. text hadoop fs –text FILE [FILE ...] Displays the textual content of files. Identical to cat if files are text files. Files in known compressed format (gzip and Hadoop’s binary sequence file format) are uncompressed first. touchz hadoop fs –touchz FILE [FILE ...] Creates files of length 0. Fails if files already exist and have nonzero length. 302 index A Access Key Id 195–196, 198 acronym libraries 295 acronyms, mining for 295 ad hoc queries 247 ad networks 124 aggregated 13 Aggregate package 90–94, 99 aggregating operation 219 algebraic property 86 algorithms computational complexity 158 data mining 272 ES2 296 global analysis 293 allDocs 293 ALTER 257 ALTER TABLE, limitations 276 Amazon AWS 101 Amazon Machine Image 194 Amazon Web Services. See AWS AMI. See Amazon Machine Image anchor text analysis 291, 293 anomaly detection 131 Apache log file collection 281 log files 42 log processing with Cascading 282 Lucene 19 Top Level Project 20 Apriori 269 vs. BC-PDM 270 arrays 256 association analysis 269 Aster Data Systems 264 authentication 195 authority 39 average, computing 86–87, 93–98 AVG 225 AWS 194 account ID 195 command line tools 198 Import/Export 210 service recommendations 211 New York Times 267 regional support 199 setting up 194–201 AWS Account Number 197–198 B Bag 222 bags grouping operators 226 Pig and 213 basic arithmetic 224 BASS. See Business Analysis Support System BC-PDM 269 cost breakdown 270 four-dimensional evaluation 269 Hadoop cluster 270 vs. Apriori 270 BerkeleyDB 281 Big Cloud–based Parallel Data Mining. See BC-PDM BIGINT 256 bigram 11 Bigtable 262 basic details 274 BitSet, Bloom filter 125 blocks 52 replica placement 184 Bloom filter 122–131, 158, 279 applications of 123–124 end-of-split output 128 implementing 124 overview 122 using Hadoop 128 BloomFilter 126, 129 Hadoop v 0.20 131 Boolean expressions 224 bottleneck 10 buckets 202 Hive 255 specified number 256 Business Analysis Support System 269 bytearray 220, 222 C caching 278–279 call data record 268 candidate navigational page 288 Capacity Scheduler 189 Carnegie Mellon University 101 Cascading 263 Apache log processing 282 MapReduce 281 case studies China Mobile 267, 269, 271–272 ES2 282–296 New York Times 267 StumbleUpon 272–273, 275, 277, 279, 281–282 cast 224 casting 236 cat command 39 cat, command 298 CDR. See call data record certificate file 197 chaining 102 driver 105 jobs in sequence 103 MapReduce jobs 131 pre- and postprocessing steps 104 chaining MapReduce jobs 103–107 INDEX 303 ChainMapper 104–107, 131 ChainReducer 104, 131 chararray 222 checksums, in HDFS files 139 chgrp, command 298 China Mobile 4, 267–272 chmod, command 299 chown, command 299 chunks 51 citation graph 65 cite_count 254 cite_grpd 242 cite table 251 classification 269 CLI 247 Hive 250 close 57 cloud 5 CloudBase 264 cloud computing 194 EC2 194 multiple reducers 49 S3 194 S3 data storage 205 Cloudera 263 ClueWeb09 101 cluster ID 142 clustering 269 clusters accessing data from 204 busy 182 launching 200 metadata handling 181 moving code to 204 multicluster setup 186 of one 29 private key, public key 26 production, monitoring 140 SSH setup for Hadoop 25–27 topology 25 type configuration 202 CMU 4, 101 cocitation, symmetric 242 cocite 242 cocite_bag 241 cocite_flat 244 codecs 153 COGROUP 227 collaborative filtering 240 columns 254, 256 column store 275 combiner 76, 95–98, 152 command line interface. See CLI command line tools 198 comments 237 commutative property 99 comparison 224 compiler, Pig 233 complex statistical model 8 compression 153–155 Hive 248 CONCAT 225 conditional 224 -conf 70 Configuration 42, 161–162 configuration object, state information 163 configuration properties dfs.permissions.supergroup 178 Hadoop 28 configure 50, 161 Configured 70 constant 224 copyFromLocal, command 299 copyToLocal, command 299 core-site.xml 28 COUNT 220, 225, 253 COUNT(*) 258 count, command 299 counters 98, 136, 146–148 missing values 147 names 146 summary 141 tracking records 146 counting 72–80, 86, 91–92 cp, command 299 CrawlDB 287 crawling 287 cross-product 110 flattening 242 record combinations 112 Cutting, Doug 4, 19 cygwin 14 D daemons DataNode 22 Hadoop 22 JobTracker 24 logs 34 NameNode 22 Secondary NameNode 23 shut down 31 TaskTracker 24 data, semistructured 223 databases accessing 170 bulk loading into 171 input and output 169 data block index 280 data flow language 213 DataInput 47 DataInputStream 52 data joining 103 cross-product 114 from different sources 107 DataJoinMapperBase 112, 163 datajoin package 131 DataJoinReducerBase 112 DataJoinReducerBase. combine() 114 DataNode 140, 142 adding 180 daemon 180 dfsadmin 177 multiple drives 174 NameNode interaction 23 offline 180 overview 22 RAM 175 removing 180 DataOutput 47 data processing model 8 data sets 64 –67, 101 Amazon AWS Public Data Sets 101 analysis, China Mobile 272 collaborative filtering 101 development 140 human genome 101 maximum and minimum computation 157 multiple 164 Netflix 101 patent citation 64 patent country codes 90 reducing 152 sparse 240 test 140 UCI 270 US Census 101 data skew 83, 94–95 data source, reduce-side join 108 data types 213 nested 223 data warehousing 268 DayOfWeek 227 DBConfiguration 170 DBOutputFormat 170 304 INDEX DBWritable 170 Debian 263 debugger 136, 151 debugging 70, 145–152 arithmetic mistakes 137 local mode 135 modes 31 pre- and postprocessing 104 decommissioning 180 deflate 154 departureNode 49 Derby 249 DESCRIBE 219, 251, 257 development data set 64 DIFF 225 diff’ing data 138 directories, adding 39 distributed applications 4 distributed cache 118, 136 reduce-side join 121 DistributedCache 118 replicated joins 117 distributed system 6 HDFS 22 distributive functions 97 distributive property 84, 86, 95, 100 document similarity 99 Dogear 286 dot notation 223 dot operator 224 double 222 DOUBLE 256 DoubleValueSum 90 DRBD 184 driver 103, 161 MapReduce program 128 DROP TABLE 253, 257 du, command 299 DUMP 217 dus, command 299 E EC2 4, 101 command line tools 195, 198 concurrent instances 204 GUI tools 199 images 203 instance handling 209 MapReduce programs and 203, 205–209 New York Times 267 SSH key pair 195 supported OSs 194 Edge 47, 49 EditLog 182 Elastic Compute Cloud. See EC2 Elastic MapReduce. See EMR emit() function 14 Employee Referral Bonus Program 289 EMR 209 equijoins 259 Ercegovac, Vuk 282 ES2 282–297 analytics 288 architecture 285 crawler 287 global analysis algorithms 292 Hadoop implementation 293 IBM 283 idp query 283 JSON 286 local analysis 288 mining tasks 295 offline analysis 285 redirection resolution 289 variant generation 285 ETL operations 269, 270 speed-up 271 evaluation mechanisms 213 Excite query log 217 exercises advanced MapReduce techniques 131–132 MapReduce paradigm 98–100 expressions, constant value 223 expunge, command 299 EXTERNAL 257 F Facebook 4 Hadoop and 247 Hive and 260 Scribe 281 failed tasks recovering 183 rerunning 151 Fair Scheduler 187, 189 false negatives, Bloom filter 122 false positives, Bloom filter 122 Fast Fourier Transform 99 fetch list 287–288 FetchQueues 288 FFT 99 field, referencing 223 field delimiter 257 FIFO scheduler 186 file commands 38, 298–301 file management tasks 38 FileOutputFormat 57 -files 83, 120 files adding 39 deleting 41 I/O 128, 129 merging 42 multiple 42 retrieving 41 storing multiple 267 file splits 154 FileStatus 43 filesystem block replication 176 checking 176–178 corrupt block 176 fsck 176 metadata, handling 181 S3, default 208 unbalanced 181 FileSystem 42 operations 44 FILTER 237 filter and transform 12 FIR filter 99 FLATTEN 229 flattening 229 cross-product 242 float 222 FOREACH 219, 228 output/input schema 230 fsck 176 FSDataInputStream 43, 52 FSDataOutputStream 43 FsImage 182 fully distributed mode 31, 135, 145–152 FuncSpec 236 functions in Aggregate package 90 G garbage collection, Java 75 General Packet Radio Service 268 GenericOptions 70 GenericOptionsParser 70, 82, 120, 149 INDEX 305 geo-classification 295 geographical data 64 get 39 get() 236 getAll() 236 getCollector 168 get, command 300 getmerge 42 getmerge, command 300 getPartition 50 getPos 57 getProgress 57 getRecordReader 55 getSplits 55 get(String) 162 getter methods 163 GFS. See Google File System global analysis algorithms 293 ES2 291 Jaql and Hadoop 294 Global Analysis 284 Google File System 19 Gottfrid, Derek 267 GPRS. See General Packet Radio Service graph data 64 Gray, Jon 280 Greenplum 264 GROUP 242 GROUP BY 219, 253, 259 group key join key 109 reduce-side join 108 Grunt managing shell 216 Pig Latin and 217 shell 215 GUI tools 199 H Hadoop 4–5 building blocks 21–25 cluster performance 270 clusters, key pairs 200 command line utilities 38 configuration 173 configuration directory 27 datajoin package 112 data processing modules 281 data types 46 default settings 28 distributed systems 6 EC2 setup 201–203 EMR 209 ES2 293 file API 42 file commands 38 file operations 40–41 fully distributed cluster 216 history 19, 20 kill command 216 modes 135, 140 linear scalability 152 production cluster properties 174 related projects and vendors 262–265 ResolveSimple 290 running 27–33 SQL databases 7–8 subprojects list 246 HADOOP_CLASSPATH 187 Hadoop cluster 5, 25 Hadoop Core 212 hadoop-default.xml 28 hadoop-env.sh 28 Hadoop-related projects Aster Data 264 Cascading 263 CloudBase 264 Cloudera 263 Greenplum 264 Hama 264 HBase 262 Katta 263 Mahout 264 ZooKeeper 262 Hadoop site properties dfs.balance.bandwidth PerSec 181 dfs.block.size 181 dfs.hosts.exclude 180 dfs.http.address 182 dfs.name.dir 183 fs.checkpoint.dir 183 fs.trash.interval 179 mapred.job.tracker 186 mapred.jobtracker. taskScheduler 187 topology.script.file. name 185 topology.script.number. args 185 hadoop-site.xml 28, 135 Hama 264 hash partitioner 172 HashPartitioner 49 hash table 10 HBase 262 added features 274 HFile 279 introduction 274 parallelism 280 region server 275, 278 SATA disks 280 StumbleUpon 274–275 table splitting 278 transcending single machine 277 ZooKeeper 263 HDFS 6, 129, 170, 248 balancer 181 blocks 52 data nodes 278 dataset storage 38 directory creation 39 file permission 178 filesystem 216 free space 175 HBase 262 metadata snapshots 23 moving data into 204 NameNode 22, 214 quota 179 reading and writing to 42 replication factor 32 security 178 shell commands 83 web interface 34 working directory 39 working with files 38–44 hdfs-site.xml 28 heap memory 175 help, looking up 41 help, command 300 heuristics 242 HFile 279 internal platform 280 keys and values 280 write path 279 histogram 73, 93 Hive 247–250 buckets 255 complex types 256 database 249 directory structure example 254 installing and configuring 248–250 loading data 257 MapReduce 247, 252 metadata 249 metastore 247 queries 250–253 relational database 247 306 INDEX running queries 258 structured data 247 tables, managing 255 table storage 254 tarball 248 HiveQL 247 built-in aggregate functions 262 built-in functions 260–262 data model 254 joins 259 queries 258 SELECT COUNT 252 standard operators, list 259 statements 250 usage 254–260 hive-site.xml 249 holistic property 86 Holstad, Erik 280 I IBM 4 IBM intranet 282–283 idempotent 156 identity 243 IdentityReducer 83, 85, 88, 104 idp query 285, 295 ES2 283 ILLUSTRATE 220, 241 indexes 254 Individual Development Plan. See IDP information retrieval 104 INNER 227 inner join 108, 112, 115 inner product of vectors 99 InputFormat 52 classes 53 custom class 54 joining 117 KeyValueTextInput Format 75 SequenceFileInput- Format 155 InputSplit 54 input splits 51 INSERT 258 INSERT OVERWRITE TABLE 253 int 222 INT 250, 256 interactive mode 217 intermediate data 11 intermediate records 166 in total 172 intranet vs. web issues 283 int type 140 IntWritable 17 inverted index 67, 73, 103, 136 IsEmpty 225 IsolationRunner 151 isSplitable 55 J Jaql 285 ES2 283 global analysis 292 jar files 15 Java 14 generics 47 regular expression 257 Java Advanced Image Extension 267 JAVA_HOME 28 java.io.DataInputStream 52 java.lang.Comparable 46 JBOD 181 jdb 151 JDBC, interface 247 JetS3t 267 joining data 107–122 Job 103 JobClient 70 JobClient.runJob() 103 JobConf 54, 70, 103, 161 chaining 106 object 129 set() 70 setCombinerClass() 97 setCompressMapOutput() 153 setInputFormat() 72, 155 setJobName() 70, 143 setMapOutputCompressor- Class() 153 setMaxMapAttempts() 149 setMaxReduceAttempts() 149 setNumTasksToExecute PerJvm() 156 setOutputKeyClass() 72 setOutputValueClass() 72 JobConf object 106 JobConf properties keep.failed.tasks.files 151 key.value.separator.in.input. line 75 mapred.job.name 143 mapred.job.reuse.jvm.num. tasks 155 mapred.local.dir 151 mapred.map.max.attempts 149 mapred.output.dir 129 mapred.reduce.max. attempts 149 mapred.reduce.tasks 70, 138 table 150 JobConf.setOutputFormat 128 JobControl 103 job ID 142, 151 killing jobs 145 jobs chaining in sequence 103 EMR 209 ID 142 kill 145, 216 maximum per user 188 name 142 pools 187 scheduling 186–189 tracking 142 JobTracker 70, 140, 142, 269 administration page 144 counter information 147 multiple 186 overview 24 TaskTracker interaction 24 Web UI 143 join datajoin package 112 HiveQL 259 inner 112 outer 112 reduce-side 109, 121 repartitioned 110, 111 join, reduce-side 102, 108 join key, group key 109 join process 107 jpox.properties 249 JSON, ES2 286 JVM mapred.job.reuse.jvm.num. tasks 156 reusing 155 TaskTracker 24 JVM usage 69 K Katta 264 ZooKeeper 263 key pair, RSA 26 key pairs 200 INDEX 307 KeyValueLineRecordReader 55, 57 key/value pairs 7, 12, 38 Configuration 42 data flow 44 list 46 splits and 55 KeyValueTextInputFormat 53, 55, 72–73 key/value types 72 in data flow 71 klass 106 K-means 269 Shanghai Branch 272 Krishnamurthy, Rajasekar 282 L LAMP 273 LIKE 259 LIMIT 218 linear scalability 152, 158 LineRecordReader 55 LinkedIn 4 Linux 14 lists 12 listStatus 43 LOAD DATA 257 Local Analysis 284 local mode 135–140 local reduce 50 log files 141–142 logging 141 messages 142 login, validating 27 log-normal distribution 76 long 222 LongSumReducer 51 long tail problem 287 long type 140 LongValueMax 90 LongValueMin 90 LongValueSum 90 LongWritable 17, 54 ls, command 300 lsr 40 lsr, command 300 Lucene 264 ES2 283 Lucene engine 19 Luo, Zhiguo 267 M machine learning 100 MacInnis, Ken 272 Mac OS X 14 Mahout 264 map 54 Map 222 MapClass 71, 161 map_input_file 164 -mapper 81 Mapper 47, 69, 160 class list 48 close() 71 configure() 71 map() 71 model 106 Mapper.configure() 119 Mapper.map() 107 mappers 8, 47 configuration properties description 153 mapred.compress.map. output 153 mapred.map.output. compression. codec 153 with Bloom 102 mapping 12 mapred.join package 117 mapred.min.split.size 55 mapred-site.xml 28 MapReduce 8–14 algorithms, efficient 157 chaining jobs 103 configuration details 35 data flow 45 EC2 and 203, 205 framework 19 key/value pairs 38 partitioning and shuffling 50 program anatomy 44–51 reading and writing 51–58 tasks 156 tuning for performance 152 web interface 35 writing to database 170 MapReduceBase 47, 71, 119 maps 256 map-side join 117 master files 32 master/slave architecture 22 copying public key 27 Hadoop cluster 25 MAX 225 maximum 84, 86, 90, 93–95, 98 median 86, 93–94 memcached 133 MemStore 275 messaging queues 8 meta block index 280 metadata block size 181 partitioning 167 metastore 247 MIN 225 minimum 86, 93–94 missing values 65, 88 mkdir 40 mkdir, command 300 Modulo 224 Moore’s law 6 move 6 moveFromLocal, command 300 moveToLocal, command 300 moving average 99 multiple machines 11 MultipleOutputFormat 164, 166 MultipleOutputs 166 filename generation 168 naming structure 169 multiple parameter files 238 multiple relations 227 MultipleTextOutputFormat 164 multiquery 239 multiset 10 mutates 278 mv, command 300 MySQL database at StumbleUpon 277 legacy data 275 metadata store configuration 250 N NameNode 140, 142 backup node 184 DataNode interaction 23 dfsadmin 178 failure, recovering 183–184 multiple hard drives 184 name quotas 179 overview 22 RAID 181 name quotas 179 National Bureau of Economic Research 64 natural language processing 11, 99, 118 308 INDEX navigational 283 NBER 64, 90 nested data types 223 nested schema 223 nesting 222 Netflix 101 network topology 184–185 New York Times 4, 267 next 55, 57 nodes 50 null 224 NullOutputFormat 58 nulls 251 NullWritable 58 numSplits 55 Nutch 19, 263 ES2 283 generate, fetch, update 287 StumbleUpon 282 web crawling 287 O offline analysis 285 offline processing 8 open source framework 4 operators, grouping 226 org.apache.hadoop.fs 42 org.apache.hadoop. io.compress 153 outer join 112 datajoin package 114 output files 57 in total sorting 172 sorting 171 OutputCollector 48, 71, 107, 128–129, 166 OutputCollector.collect (K k, V v) 107 OutputFormat 57 classes 58 FileOutputFormat 154 NullOutputFormat 129 SequenceFileOutput Format 154 TextOutputFormat 75, 87 OutputFormat class 164 outputs, multiple 164 P page_view 255 pair-wise combinations 242 patents example 240 pair-wise computations 240 parallelization, granular 52 PARALLEL n 231 parameters for practical use 174–176 multiple files 238 passing custom 160 security 201 substitution 238 partitioner 87, 166, 172 hash 172 Mapper output redirecting 49 TotalOrderPartitioner 172 partitioning 11 data 254 Hive 248 into multiple output files 164 LOAD DATA 258 metadata 165 pass by reference, in output collector 107 pass by value, in output collector 107 patent citation data 64–65, 107, 131 patent description data 64–65, 107 patents cocitation 240 cocitation counts 244 path 39 Path 43 pattern matching 224 performance improvement 107 Perl 282 permissions, setting 178 petabytes 3 Pig and Cascading 263 Cygwin 214 data types 213 Hadoop cluster 214 installing 214–215 interactive shell 215 Java API 235 JAVA_HOME 214 JobTracker 214 JVM 215 like SQL queries 215 multiquery execution 239 NameNode 214 parameters, setting 216 philisophy 213, 215 running 215–217 similar patents 240–244 UDFs and 214 PIG_CLASSPATH 214 PiggyBank 233 UDF 236 piggybank.jar 234 PIG_HADOOP_VERSION 214 Pig Latin 213, 217–233 APIs 233 atomic data types 222 built-in expressions 225 comments 237 complex data types 222 data flow language 213 data read/write operators 218 data types and schemas 221–223 diagnostic operators 221 execution optimization 233 expressions 224 expressions and functions 223–225 Hadoop mode 215 Java equivalent classes 235 like SQL query 219 local mode 215 relational operators 225–233 schema 219 script 213 scripts 237–239 type 235 UDF statements 234 user-defined functions 233, 235–237 pig.properties file 214 PigServer 215 pipelines 8 pipes, Cascading 263 pirate-speak 100 pools 187 POSIX 178 pound operator 224 power law distribution 76 prefix 238 primitives 8 private key 197, 198 SSH 26, 200 production cluster 174 pseudo-code 9 pseudo-distributed mode 29, 39, 41, 64, 135, 140 INDEX 309 public key 197 cryptography 26 distribution 27 EC2 200 SSH 200 put 39 put, command 300 PutMerge 42, 43 Q QOS 268 quadratic complexity 240 quality of service. See QOS queries, column 254 query processing Hive 248 Hive examples 250 quotas, managing 179 R rack awareness 184–185 Raghavan, Sriram 282 RAID 181 RAM storage 10 Rawson, Ryan 272 RDBMS 276 readFields 47 read-many-times 8 read-slaves 278 RecordReader 54–55 joining 117 RecordReader 55 records processing guaranteed 52 skipping bad 148 RedHat, Cloudera 263 reduce 48, 139 Reduce, class 71 reducer 8, 48, 139 multiple 49 partitioner and 172 ResolveSimple 290 Reducer 48, 69, 97, 160 class list 49 reduce() 71, 97 reduce-side join 108–115 reducing 12 Reflection 236 REGEXP 259 region server 275 HBase 278 REGISTER 234 regression testing 138–140 regular expression patterns 295 Reiss, Frederick 282 relational data 64–65, 108 aggregate analysis 227 model 223 nonmatching 227 relational database 7, 152 interfacing with 169 remote storage 184 repartitioned join 108 replicated join 118–121 replication factor 40 Reporter 48, 142, 146 incrCounter 146 setStatus 142, 145 Resolve2Step 290 resolution table 291 ResolveSimple 290 results 293 rm 41 rm, command 300 rmr, command 300 ROW FORMAT 256 row store 275 RPM 263 S S3 194 accessing data directly 207 dataflow model 206 datapath 207 GUI tools 199 Hadoop options 207 moving data through 205 New York Times 267 storage service 202 S3 Block FileSystem 206 S3 bucket 202, 206 s3n 208 sampling data 64, 82, 84, 135, 152 sanity checking 137–138 scheduler 189 schema 7 data source records 108 fixed 219 nested 223 scheme 39 Scribe 281 search engine optimization 281 search-hadoop.com 264 Secondary NameNode 23, 181–183 confusing name 182 location specification 30 server 183 snapshots 23, 182 SecondaryNameNode 140, 142 Secret Access Key 198 slashes and 195 slash problem 208 Secure Shell. See SSH security, parameters 201 segment 287 segment file 289 SELECT 251 Sematext 264 semijoin 108, 121–122, 131 semistructured data 223 SequenceFileInputFormat 54 SequenceFileOutputFormat 58 sequence file 54, 154–155 configuration properties 155 SequenceFileOutput- Format 154 set command 216 SETI@home 6 setOutputFormat 57 setrep, command 301 setSpaceQuota command 179 Shanghai Branch 270 K-means 272 Shekita, Eugene 282 SHOW TABLES 251, 257 shuffle 45 shuffle, MapReduce 87 shuffling 11, 50 sign 224 signal processing 99 Simple Storage Service. See S3 site root analysis 291, 293 SIZE 225 SkipBadRecords 149, 150 skipping bad records 148–150 slave machines 25 slave nodes 186 location specification 30 slaves files 32 SMALLINT 255 SNN. See Secondary NameNode Snoop Dogg 100 social networks 64 Solaris 14 310 INDEX sorted order 171 spam filter 10 sparse data 240 sparse vector representation 99, 132 spatial join 132 speculative execution 156 configuration properties 157 SPLIT 225, 237 splits Bloom filters and 128 input 52 InputFormat class 55 splittable file formats 154 SQL 7 and Hive 213 engine 8 vs. Pig Latin 213 Squid, Web proxy server 124 SSH channels 30 installation 26 SSH login 200 SSN. See Secondary NameNode standalone mode 28, 64, 135 standalone node 25 standard deviation 93 Stanford 4 stat, command 301 status messages 142, 145 STDERR 142, 148 STDOUT 142 stemming 104 stop words 104 STORE 217 Streaming 80–94 configuration properties 164 outputting logs 142 PHP scripting 84 Python scripting 82, 89 rewriting for Hadoop Java 157 setting job name 143 skipping bad records 150 split processing 102 Unix commands 81–82 updating status 142 use of combiner 96 use of compressed files 155 use of counters 148, 150 Streamy.com 280 String 161 STRING 256 StringValueMax 90 StringValueMin 90 structs 256 structured query language. See SQL Stumblers 272 Stumbles, per user 276 StumbleUpon 272–273, 275, 277, 279, 281–282 architecture 273 data handling 281 HBase 274–275 ratings 272 SUM 225 summarization 247 summing 86, 90 Sun, Shaoling 267 supergroup 178 superuser 178 symbolic links 33 sync markers in files 154 system malfunction 182 T tab character 58 tables, splitting, HBase 278 tag, reduce-side join 108 TaggedMapOutput 112, 115 TaggedWritable datajoin 113 tail 41 tail, command 301 taps 263 tarball, Hive 248 task attempt ID 151 task ID 145 task initialization 155 tasks CPU-intensive loads 175 failed, rerunning 156 parallel running 156 passing parameters 160 placement 185 reduce, side effects 157 retrieving information 163 TaskTracker 140, 269 bad record tracking 148 configuration object 161 default tasks 175 JobTracker interaction 24 overview 24 RAM 175 Web UI 145 Tata, Sandeep 282 template, for MapReduce program 67, 118 terabytes 3 test, command 301 Text 17, 54, 72 datajoin 112 text, command 301 TEXTFILE 257 TextInputFormat 52 TIFF, image processing 267 time series 64, 99 TimeUrlLineRecordReader 56 TINYINT 255 TokenCountMapper 51 tokenization 16 TOKENIZE 225 TokyoCabinet 281 Tool 70, 162 Tool.run() 70 ToolRunner 70, 162 toString 58 touchz, command 301 trash 179 trigram 11 Tuple 222 tuples Cascading 263 Pig and 213 Twitter 4 type casting 224 U Ubuntu 263 UCI data sets 270 UDFs 213 eval 234 filter functions 237 jar file 234 load/store 234 Pig Latin and 233 simple scalar 236 unique values, computing 92–93 UniqValueCount 90 Unix commands, parameter values 238 cost breakdown 270 Hive CLI 258 pipes 8, 41, 103 server 268 server performance 270 wc -1 136 Unix command 137 UPPER UDF 235 URI, format 39 INDEX 311 user-defined functions. See UDFs username 178 V Vaithyanathan, Shivakumar 282 ValueHistogram 90, 93 variance, computing 86, 158 variant generation 285 visualization 65, 77, 92 void close 47 void configure(JobConf job) 47 W web interfaces 34 search engine 19 Web GUI 247 web server, log analysis 99, 131 Web UI 98, 142, 147, 189 cluster 34–36 Hadoop 216 whitespace characters 16 Windows 14 word count 9, 72 predefined classes 51 with Hadoop 14–19 Writable 72, 170 Bloom filter 126 datajoin 113 IntWritable 73, 140 LongWritable 140 Text 73 WritableComparable 72 example 47 WritableComparable 46 write-once 8 X X.509 Certificate 195, 198 Xu, Meng 267 Y Yahoo 4, 212 Z Zhu, Huaiyu 282 ZooKeeper 262, 278 B ig data can be diffi cult to handle using traditional data- bases. Apache Hadoop is a NoSQL applications frame- work that runs on distributed clusters. Th is lets it scale to huge datasets. If you need analytic information from your data, Hadoop’s the way to go. Hadoop in Action introduces the subject and teaches you how to write programs in the MapReduce style. It starts with a few easy examples and then moves quickly to show Hadoop use in more complex data analysis tasks. Included are best practices and design patterns of MapReduce programming. Th is book requires basic Java skills. Knowing basic statistical concepts can help with the more advanced examples. What’s Inside Introduction to MapReduce Examples illustrating ideas in practice Hadoop’s Streaming API Other related tools, like Pig and Hive Chuck Lam is a Senior Engineer at RockYou! He has a PhD in pattern recognition from Stanford University. For online access to the author and a free ebook for owners of this book, go to manning.com/HadoopinAction $44.99 / Can $51.99 [INCLUDING eBOOK] Hadoop IN ACTION DATABASES/PROGRAMMING Chuck Lam “A guide for beginners, a source of insight for advanced users.” —Philipp K. Janert Principal Value, LLC “A nice mix of the what, why, and how of Hadoop.” —Paul Stusiak Falcon Technologies Corp. “Demystifi es Hadoop. A great resource!” —Rick Wagner, Acxiom Corp. “Covers it all! Plus, gives you sweet extras no one else does.” —John S. Griffi n, Overstock.com “An excellent introduction to Hadoop and MapReduce.” —Kenneth DeLong BabyCenter, LLC MANNING SEE INSERT




需要 15 金币 [ 分享pdf获得金币 ] 30 人已下载





下载需要 15 金币 [金币充值 ]
亲,您也可以通过 分享原创pdf 来获得金币奖励!