Ceph

Okay, this is just bizarre. There are actually 4 different PGs that report a stat mismatch error. 3 of them (including 0.f8) are off by the exact same amount -- 34055KB more in the on-disk state than the in-memory info -- and the other two somehow fix themselves. 0.f8 seems to eventually fix itself, since after doing the repair (which detects the same problem as the initial scrub 3.5 hours later, so the problem can't be JUST a race) it later says "hey, my recorded stats are too high compared to my on-disk state!" and then reduces it by the same amount it initially incremented them by.
But, somehow, between the first repair and the next scrub, the on-disk quantity goes down despite the PG adding 8 objects and apparently only doing write operations (since the repgather commit messages are set to way too high a debug level, this is the one thing I can see).