On Fri, Feb 01, 2013 at 07:24:52PM -0500, Digimer wrote:
> The need for a resync is because you didn't have a fence device setup so
> it split-brained. If you use it with the 'crm-fence-peer.sh' fence
> handler this would not happen.

Even if DRBD "split-brained" (experienced data divergence),
there won't be a need for "full sync".

As each node tracks which block was modified without being replicated,
at a granularity of 4kB, even after data-divergence (split brain),
you just have to tell them the direction of the resync.

> Even without DRBD you will need fencing in a cluster to maintain sanity
> and avoid losing data.

That is true.

> On 02/01/2013 06:39 PM, Paul Archer wrote:
> > Thanks so much for the quick, detailed, and informative answer!
> > I'll definitely go with building the latest and greatest of corosync
> > and pacemaker. 
> > As far as the cluster I already have setup, I'll have to get those
> > configs on Monday. Oh, and FWIW, I tried using DRBD with KVM, but
> > found it too fragile. If I lost a node, the whole block device
> > needed to be synced.

That would be *very* unexpected.
"Whole block device", or "full" resyncs should be really rare with DRBD,
and usually are a result of either a conceptually broken setup,
or indication of a bug in DBRD.

Sometimes, it was just a misunderstanding, because /proc/drbd shows
something about some 100% ...  but those are the 100% of the
to-be-resynced blocks in this particular resync,
not the 100% of the device...

> > I settled on the latest version of gluster instead. Resyncing is
> > faster, as the files for the individual VMs can be synced.

Whatever works best for you.

-- 
: Lars Ellenberg
: LINBIT | Your Way to High Availability
: DRBD/HA support and consulting http://www.linbit.com
_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems

Reply via email to