Hi all.
I have some questions about how MDS cluster works when crash or operation
failure occur.
1. I read ceph documentation and code that each MDS has its own journal and
some directory operations like rename use distributed transaction mechanism
with ceph-defined events (e.g. EPeerUpdate). What I wonder is when one of
MDSs is crashed and then reboots, if some metadata in other MDS's journal
is not flushed yet, is rebooted MDS read stale metadata?
2. While rename operation is performing on 2 MDSs and one of them is
crashed, Does the other MDS rollback modified in-memory metadata ?
Thanks,
kyujin
Show replies by date