tested locally:

1.MON=1 OSD=3 ./vstart.sh -l -d -n mon osd -o 'osd_map_cache_size = 20' -o 'mon_min_osdmap_epochs = 30' -o 'paxos_service_trim_min=20' # these options are 200, 500 and 250 by default, the larger these number are, the more osdmaps are kept in cache and objectstore.
2. start three consoles:

• while true; do ./ceph osd pool create test-pool 1 1;./ceph osd pool delete test-pool test-pool --yes-i-really-really-mean-it; done # keep updating the osdmap
• watch -n1 bash -c 'find osd1/current/meta/ | wc -l' # the osdmap are stored in meta collection using objectstore. 1 (incremental) osdmap per object. so we can simply count the number of files in that directory. it increases when more osdmap are stored, but it drops when we are trimming the osdmap in monitor, and hence evicting the osdmap cache in OSD.
• while true; do ./ceph report|grep osdmap_; sleep 3; done # once "osdmap_first_committed" changes, the monitor will trim its osdmap epochs, and OSD should remove them from its objectstore

the fix is wrong. we send NullEvt when a new osdmap is consumed, which also updates all PGs in the pg_map, if a PG's persisted osd map is too stale (osd_pg_epoch_persisted_max_stale), the PG::write_if_dirty() updates the last_persisted_osdmap_ref with the latest osdmap set by PG::handle_advance_map.

but it could still serve as a cleanup.

lgtm