[DRBD-user] Avoiding data inconsistencies after a split-brain

Andre Nathan andre at digirati.com.br
Fri Apr 16 19:02:22 CEST 2010

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


Hello

I'm setting up a two-node active/active cluster with DRBD and OCFS2.
When the nodes lose communication with one another, a split-brain
happens, and both machines and up in the stand-alone/active state.

One possible solution to minimize the amount of time where writes can be
done to both nodes without synchronization would be to remount the
filesystems as read-only with the split-brain hook. This, however, won't
avoid inconsistencies with writes happening before the hook has a chance
to run.

Is there a better way to avoid inconsistencies in a situation like this?
I was thinking that a new protocol that would try to sync writes with
the peer before doing a local write could solve this, refusing to do the
local write (and possibly setting the device as read-only -- can drbd do
that?) if the remote write isn't successful. Such a protocol would
surely result in a performance penalty, but maybe it could be worth it
if that level of safety could be achieved.

Does this idea make sense? Is there a different solution for this
problem that doesn't require something as radical as writing a new
protocol?

Thanks in advance,
Andre




More information about the drbd-user mailing list