Ceph » rbd

Xuehan Xu wrote:

This might be a problem for flattens. If a flatten is interrupted with object maps updated and the objects not actually fully modified, those objects won't be copied up in later flattens, which leads to corrupted data.

Noting that this issue extends beyond an interrupted flatten. If a client running on the clone crashes at the wrong moment, the object map would be left with the same kind of inconsistency. If it's not corrected by that client restarting and writing to the same area as part of filesystem or application level recovery or just by chance in the future, "rbd flatten" would similarly skip the copyup even when running to completion on the first attempt, without ever being interrupted.