Ceph » CephFS

If we rmdir an empty directory whose dirfrag(s) are exports/subtree roots,

- it gets journaled locally, replay is fine
 - the replica is told and subtrees update fine
 - the replica does not journal, so replay will break

We can either
- journal witnesses on replicas the way we do rename (and do the rollback etc.)
- force empty dirfrags to export back to the inode auth, and make the rmdir wait for that.

The latter is simpler, but exports aren't always allowed (e.g. when the mds cluster is degraded), so you could have rmdirs that hang when otherwise unrelated mds's are recovering. The former is more robust but more complex to implement.