While checking logs for SERVER-4100 causes (this case is unrelated, mentioned for information purposes only), I got some strange reconds on the primary:

Sat Jan 14 02:15:37 [initandlisten] connection accepted from 10.2.1.200:48055 #108356
Sat Jan 14 02:15:48 [conn108356] update models_stats.show_log query:

{ _id: ObjectId('4ebf7eb76b8dda061480d539') }

update: { $unset:

{ h: 1 }

, $set:

{ e: null, s: 1321172663 }

} keyUpdates:1 1271310315ms
Sat Jan 14 02:15:53 [conn108356] end connection 10.2.1.200:48055

This part of log shows the update running for something around 14 days, but entire connection existed less than a minute, looks like there's something amiss.
We have several (in the tens range) of such queries in logs with the same time (1271310315ms) on different connections with different operations.