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

[Robert Hajime Lanning]

On 11/15/12 23:37, Martin Emrich wrote:
> 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...

Since replica mode is designed for local replicas, when a node goes down 
all clients write to the other node.

Split-brain is dealt with by determining a winner and all others get 
disabled.

The winner in a split brain is determined by >50% or if 50%, then the 
side with the first brick is the winner.

I have 12 servers, each with 4 bricks, all on the same subnet.
They are accessed by 96 other servers, via the same subnet.
So, if half my gluster servers go down (say all of the replica
number 1 of a 2 replica volume) via some weird switch outage, the
side with storage node 1 will be online, the other side will not.

If you disable the quorum feature, both will be working, but when
then network comes back, the files that conflict will give an IO error.

Also, since writes to replicas are synchronous, your write latency will 
be huge.

-- 
Mr. Flibble
King of the Potato People