[ClusterLabs] Fencing on 2-node cluster

[Klaus Wenninger]

On 06/21/2018 05:52 AM, Andrei Borzenkov wrote:
> 21.06.2018 00:50, Digimer пишет:
>> On 2018-06-20 05:46 PM, Jehan-Guillaume de Rorthais wrote:
>>> On Wed, 20 Jun 2018 17:24:41 -0400
>>> Digimer <lists at alteeve.ca> wrote:
>>>
>>>> Make sure quorum is disabled. Quorum doesn't work on 2-node clusters.
>>> It does with the "two_node" parameter enabled in corosync.conf...as far as I
>>> understand it anyway...
>> It doesn't, that option disables quorum in corosync.
>>
> This option does not disable quorum - this option fakes quorum so
> corosync continues to report "in quorum" even when one node is lost. it
> is quite possible that pacemaker quorum does not map one-to-one to
> corosync quorum though.
>
>> Quorum is floor(($nodes / 2) + 1). So in a 3-node, that is 3 -> 1.5 ->
>> 2.5 -> 2 votes needed for quorum. In a 2-node, that is 2 -> 1 -> 2 -> 2
>> votes needed for quorum, meaning you can't lose a node to operate (which
>> is kinda not HA :) ).
>>

Let me try to shed some light on this:
With Corosync 2.x.x and up Pacemaker gets quorum from Corosync so
setting two_node there directly gives you the behavior desired.
It was with Corosync 1.x.x when Pacemaker just used the messaging
from there and had it's own quorum logic built in.
As far as I'm aware SBD is the only HA-component that - given
Corosync 2 and up is used - in case of two_node configured with
Corosync doesn't take over the quorum information. (It actually
doesn't need the two_node-logic as it either doesn't work at
all on two-node-clusters if it is running without disk or having
the disk it has something like a 3rd instance.)

Klaus
 
> Yes, but that assumes normal, non two_node, configuration. As said,
> two_node makes corosync to always pretend quorum is available (after
> initial implicit wait_for_all).
> _______________________________________________
> Users mailing list: Users at clusterlabs.org
> https://lists.clusterlabs.org/mailman/listinfo/users
>
> Project Home: http://www.clusterlabs.org
> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> Bugs: http://bugs.clusterlabs.org