[Gluster-users] Avoid Split-brain and other stuff

[Brian Candler]

On Fri, Nov 16, 2012 at 08:37:01AM +0100, Martin Emrich wrote:
> > Any multi-master replication suffers from exactly the same split-brain
> > scenarios as you described earlier.
> 
> That would be perfectly acceptable, as long as it would heal deterministically (last one wins, or renamed conflicting files)

Not for me it wouldn't. "Last one wins" means "one set of updates thrown
away", i.e.  definite data loss, which will be compounded when further
updates take place.

Automatic renaming means either that the file vanishes from its original
name (so the application which looks for the file breaks anyway), or that
one version has the original name and the other version is renamed - which
can also result in irrepairable damage.

> > In glusterfs, geo-replication is what you should use for WAN-separated sites.
> > Replicated volumes are for LAN scenarios where partitioning (and hence split-
> > brain) should not be expected to occur.
> 
> Hmm, then I wonder how a high-availability scenario would work, if it is not allowed for a node to go down in a replica-mode setup...

Certainly a node can go down, come up again later, and while out-of-date it
will resync.

What you don't want is both nodes to be up, both reachable only by a subset
of clients, and updates occurring on both.

Regards,

Brian.