> Also, I still think such rollback should be fast - maybe a bug in rbd? The changelog from ceph 0.63 includes: * librbd: parallelize delete, rollback, flatten, copy, resize I guess this will also speedup those operations - already tested that?