[ClusterLabs] Automatic recover from split brain ?

[Andrei Borzenkov]

11.08.2020 10:34, Adam Cécile пишет:
> On 8/11/20 8:48 AM, Andrei Borzenkov wrote:
>> 08.08.2020 13:10, Adam Cécile пишет:
>>> Hello,
>>>
>>>
>>> I'm experiencing issue with corosync/pacemaker running on Debian Buster.
>>> Cluster has three nodes running in VMWare virtual machine and the
>>> cluster fails when VEEAM backups the virtual machine (I know it's doing
>>> bad things, like freezing completely the VM for a few minutes to make
>>> disk snapshot).
>>>
>>> My biggest issue is that once the backup has been completed, the cluster
>>> stays in split brain state, and I'd like it to heal itself. Here current
>>> status:
>>>
>>>
>>> One node is isolated:
>>>
>>> Stack: corosync
>>> Current DC: host2.domain.com (version 2.0.1-9e909a5bdd) - partition
>>> WITHOUT quorum
>>> Last updated: Sat Aug  8 11:59:46 2020
>>> Last change: Fri Jul 24 07:18:12 2020 by root via cibadmin on
>>> host1.domain.com
>>>
>>> 3 nodes configured
>>> 6 resources configured
>>>
>>> Online: [ host2.domain.com ]
>>> OFFLINE: [ host3.domain.com host1.domain.com ]
>>>
>>>
>>> Two others are seeing each others:
>>>
>>> Stack: corosync
>>> Current DC: host3.domain.com (version 2.0.1-9e909a5bdd) - partition with
>>> quorum
>>> Last updated: Sat Aug  8 12:07:56 2020
>>> Last change: Fri Jul 24 07:18:12 2020 by root via cibadmin on
>>> host1.domain.com
>>>
>>> 3 nodes configured
>>> 6 resources configured
>>>
>>> Online: [ host3.domain.com host1.domain.com ]
>>> OFFLINE: [ host2.domain.com ]
>>>
>> Show your full configuration including defined STONITH resources and
>> cluster options (most importantly, no-quorum-policy and stonith-enabled).
> 
> Hello,
> 
> Stonith is disabled and I tried various settings for no-quorum-policy.
> 
>>> The problem is that one of the resources is a floating IP address which
>>> is currently assigned to two different hosts...
>>>
>> Of course - each partition assumes another partition is dead and so it
>> is free to take over remaining resources.
> I understand that but I still don't get why once all nodes are back
> online, the cluster does not heal from resources running one multiple
> hosts.

In my limited testing it does - after nodes see each other pacemaker
sees resources active on multiple nodes and tries to fix it. This is
with pacemaker 2.0.3. Check logs on all nodes what happens around time
node becomes alive again.

>>
>>> Can you help me configuring the cluster correctly so this cannot
>>> occurs ?
>>>
>> Define "correctly".
>>
>> The most straightforward text book answer - you need to have STONITH
>> resources that will eliminate "lost" node. But your lost node is in the
>> middle of performing backup. Eliminating it may invalidate backup being
>> created.
> Yeah but well, no. Killing the node is worse, sensible services are
> already running in clustering mode at application level so they do not
> rely on corosync. Basically corosync is providing a floating IP for some
> external non critical access and starting systemd timers that are
> pointless to be run on multiple hosts. Nothing critical here.
>>
>> So another answer would be - put cluster in maintenance mode, perform
>> backup, resume normal operation. Usually backup software allows hooks to
>> be executed before and after backup. It may work too.
> This in indeed something I might look at, but again, for my trivial
> needs it sounds a bit overkill to me.
>> Or find a way to not freeze VM during backup ... e.g. by using different
>> backup method?
> 
> Or tweaks some network settings so corosync does not consider the node
> as being dead too soon ? Backup won't last more than 2 minutes and the
> freeze is usually way below. I can definitely leave with cluster state
> being unknown for a couple of minutes. Is that possible ?
> 

You probably can increase token time but that also directly affects how
fast pacemaker will start and also how fast it will react to node failures.

> Removing VEEAM is indeed my last option and the one I used so far, but
> this time I was hoping someone else would be experiencing the same issue
> and could help me fixing that in a clean way.
> 
> 
> Thanks
> 
>> _______________________________________________
>> Manage your subscription:
>> https://lists.clusterlabs.org/mailman/listinfo/users
>>
>> ClusterLabs home: https://www.clusterlabs.org/
> 
>