Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
Mesosphere Q4 Sprint 3 - 12/7, Mesosphere Q1 Sprint 1 - 1/23, Mesosphere Q1 Sprint 2 - 2/6, Mesosphere Q1 Sprint 3 - 2/20, Mesosphere Q1 Sprint 4 - 3/6, Mesosphere Q1 Sprint 5 - 3/20, Mesosphere Q1 Sprint 6 - 4/3, Mesosphere Q1 Sprint 7 - 4/17, Mesosphere Q2 Sprint 8 - 5/1, Mesosphere Q1 Sprint 9 - 5/15, Mesosphere Sprint 10, Mesosphere Sprint 11
-
8
Description
Having added a URI flag to CommandInfo messages (in MESOS-2069) that indicates caching, caching files downloaded by the fetcher in a repository, now ensure that when a URI is "cached", it is only ever downloaded once for the same user on the same slave as long as the slave keeps running.
This even holds if multiple tasks request the same URI concurrently. If multiple requests for the same URI occur, perform only one of them and reuse the result. Make concurrent requests for the same URI wait for the one download.
Different URIs from different CommandInfos can be downloaded concurrently.
No cache eviction, cleanup or failover will be handled for now. Additional tickets will be filed for these enhancements. (So don't use this feature in production until the whole epic is complete.)
Note that implementing this does not suffice for production use. This ticket contains the main part of the fetcher logic, though. See the epic MESOS-336 for the rest of the features that lead to a fully functional fetcher cache.
The proposed general approach is to keep all bookkeeping about what is in which stage of being fetched and where it resides in the slave's MesosContainerizerProcess, so that all concurrent access is disambiguated and controlled by an "actor" (aka libprocess "process").
Depends on MESOS-2056 and MESOS-2069.
