Ceph

lxo reported that on one monitor, after seeing #2031 and bringing the monitor back up (much later), the monitor failed to update its PGMap.

This version is the one the monitor crashed at, and the referenced version no longer exists on disk (nor do its surroundings). The only way for it to have been trimmed past the latest_stashed is as part of slurp(), which should have also stashed a newer version. How this failed to happen is beyond me.

mon.2@2(peon).pg v283678 update_from_paxos: error parsing incremental update: buffer::end_of_buffer

I have made a bunch of related changes to this so that if it happens again it will be more obvious, though.