[Pacemaker] Stonith & two nodes cluster

[Digimer]

On 17/03/14 11:52 PM, khaled atteya wrote:
> Hi,
> In case of two nodes cluster in case of Active/Active or Active/Passive,
> if split brain happen , which node will STONITH the other ?

"The fast node" is the short answer.

The long answer is that you can give one node a priority over the other 
by setting a 'delay="15"' for the fence method of the node you want to 
live (tells the other node to wait 15 seconds before fencing it).

There are two other notes to be aware of in 2-node clusters;

* Dual fence
* Fence loops

With a dual-fence, both nodes call a fence at the same time and both 
succeed, leaving both nodes powered off. This can happen with separate 
fence devices, like IPMI BMCs. The 'delay' above helps, but alone isn't 
enough. With IPMI, for example, you will also want to disable acpid from 
running. Most servers will react to a power button event by instantly 
shutting off when acpid is stopped, minimizing the window where the 
slower node could get it's fence call out before dieing.

With fence loops, you can avoid this easily but *not* starting the 
cluster on boot. If you allow it to start on boot, then in the case 
where the nodes lose their connection to one another, the fenced node 
will boot back up, timeout waiting for the first node and fence it. That 
node will boot, start the cluster, time out and fence the first. And so 
on...

Even with 3+ nodes, I still recommend not starting the cluster on boot. 
In production, fence events should be *very* rare event and only occur 
when something bad happens. So having the node boot up makes it easier 
to log in and investigate (as opposed to using the 'off' fence action 
directly).

hth

-- 
Digimer
Papers and Projects: https://alteeve.ca/w/
What if the cure for cancer is trapped in the mind of a person without 
access to education?