Ceph

[0,1] is the acting set
Object foo is present on 0, missing on 1.
We accept a write on foo submitting the write to both 0 and 1.
0 restarts before persisting the write, 1 remembers the log entry.
foo is now unfound.

This isn't actually a bug as such, even without degraded writes, there is a similar window when writing to a size one acting set (min_size 1, size 2), allowing the single osd to propogate the log to another osd, and then killing the first osd. It might be worth changing find_best_info to prefer a log with an old last_update if the newer last_update holder happens to have a lower last_complete.