While we're spitballing, from my recollection, it seems to do a lot of small writes; probably different write cache policies, possibly including the VM not passing through fsync/write barriers.

And, you know, organizational failure to have this issue lingering.