[Gluster-devel] Splitbrain Resolution

[gordan at bobich.net]

On Fri, 18 Apr 2008, Joey Novak wrote:

> >   As for AFR, the client takes care of the work. 
> servers are essentially just network
> accessible file systems, that aren't aware of each other, and thus, 
> don't replicate between each others at all.

So for n mirrored servers, the client bandwidth increases n-fold? That's 
quite shocking...

>   SplitBrain...  While it is possible for this to happen, Gluster is not designed like Coda
> for offline operation.  If a client can't hit a server, it is unlikely another client can hit
> it, and thus create a splitbrain scenario...  But to answer your question, I don't think
> there is any conflict resolution method.

OK, say I have 2 nodes. The switch fails, but they are both using their 
own "mirrored" copy to mount the FS. The switch gets fixed. There are now 
two separate copied of the data. What happens?

Alternatively, one node gets shut down. When it comes back up, how does 
GlusterFS tell which is the one that's out of date?

>   Fencing and Quorum...  I don't think you will find those terms anywhere in the
> documentation.

I didn't - hence why I was asking.

>   Metadata change detection...  That is a good question, I hope someone provides a good
> explanation. 

Me too. :-)

Gordan