[ClusterLabs] faulty business logic

[Ken Gaillot]

On Fri, 2021-02-19 at 16:40 +0000, lejeczek wrote:
> Hi guys.
> 
> I have a simple cluster with simple constraints:
> 
> Colocation Constraints:
>    check-jupyterhub with openvpn-server (score:INFINITY)
>    secret-dropbox with openvpn-server (score:INFINITY)
> Ticket Constraints:
> 
> no other constraints.
> What puzzles me, and I assume I miss something here, is - 
> when/if resource 'secret-dropbox' fails on a node then 
> cluster relocates, reshuffles everything.
> That is bit bizarre, no?

You're probably interested in the resource-stickiness property. It
tells the cluster how much you want to keep things where they are,
versus balance them across nodes.

By default, stickiness is 0, meaning the cluster will always try to
balance resources across nodes as much as possible. If one resource
starts or stops, that can lead to shuffling other resources around.

Some people have suggested defaulting stickiness to 1, since users do
tend to expect things to not move around much. That's a possibility for
future versions, but it would be a big change in behavior, and there
would be people with the opposite preference who would be surprised by
that.

> What is the logic behind that? (and how to tell cluster that 
> 'openvpn-server' is what matters here and if 
> 'secret-dropbox' can't follow then, well, tough luck)
> 
> many thanks, L.

Stickiness probably covers most of what you want, but there's a new
feature coming in the next release (toward the middle of this year)
that's relevant as well.

You'll be able to set critical=false on secret-dropbox, and if secret-
dropbox fails enough times to need to be moved to a different node,
then secret-dropbox will stop rather than force openvpn-server to move.
-- 
Ken Gaillot <kgaillot at redhat.com>