[ClusterLabs] Favoured node in priority-fencing-delay

[Andrei Borzenkov]

On 03.11.2021 05:21, Alex Zarifoglu wrote:
> Hello all,
> I have a question about the "priority-fencing-delay" parameter.
> This parameter, although very helpful, it doesn't handle the scenario where 
> nodes have equal priority.

It is intended exactly for the scenario where nodes have equal priority
and fencing decision is based on resources that are active on each node.

> Thus, it is recommended that fence property 
> pcmk_delay_base is still used to avoid nodes fencing each other if they have 
> equal priority.

Where is it recommended?

If you configured priority based fenced, and total weight of resources
on each node is equal, then probably priority based fenced is not
suitable for your case. If one node has clear majority, then there is no
issue.

> I wonder if there is a setting where we can set a node to be favoured in case 
> both nodes have equal priority (So the cluster acts as if one of the nodes has 
> higher priority). I understand that we can already do this by configuring a 
> delay to one of the nodes, and the node with no delay will fence the other node 
> right away... However, this still introduces a delay during fencing and creates 
> a more complicated design and having some setting to avoid the delay parameters 
> altogether would be great.

Actually it is mixing of node and resource priorities that would
complicate design even further.

Priority based fencing is really intended for the case you have clear
winner. Examples where it makes sense

- active-passive cluster where resources are active on one node only.
Any variation thereof where resources on the second node are disposable
(e.g. production and test instances).

- replicated database to keep current primary and avoid fail over. DRBD
in active/passive setup.

If you have a lot of unrelated resources, priority fencing is not
useful. Is resource A with priority 100 on node n1 more important than
resource B with priority 100 on node n2? If you cannot answer this
question, just use standard node based delay.

> Is there some type of variable that can achieve what I described above?
> Please let me know if there is or if you know of any plans to create something 
> similar in future versions.
> Thank you,
> Alex
> *Alex Zarifoglu*
> Software Developer *|* *Db2* pureScale & HA
> *email:* alex.zarifoglu at ibm.com
> 
> 
> 
> _______________________________________________
> Manage your subscription:
> https://lists.clusterlabs.org/mailman/listinfo/users
> 
> ClusterLabs home: https://www.clusterlabs.org/
>