[Pacemaker] split brain situation

[Andrew Beekhof]

On Feb 10, 2009, at 4:19 AM, Philip Pinto wrote:

> Well one other reason you may want a two node cluster would be  
> running under
> a virtualization engine - VMware or z/VM - just to refrain from  
> wasting
> hypervisor resources

Thats a pretty poor reason.
A minimal third node that does nothing but vote in elections is going  
to be swapped to disk for most of its life and consume very little CPU/ 
RAM.

> - all of the clusters we run are two node clusters for
> this reason - and HA is the only solution we have to allow for  
> monitoring
> and restarting failed resources.  We do automatically restart fenced  
> nodes
> in the event of a failure.  Last time I checked I was sane - at  
> least that's
> what I keep telling myself ;-)

It will work fine right up until the point at which you get a  
persistent failure (one not solved by a reboot).
At which point you'll go into a STONITH loop with each node coming up  
and shooting the other over and over again until you resolve the  
underlying issue.

Being VMs, that might reduce the chance of a persistent failure, but  
the possibility is still there.