[ClusterLabs] Low priority colocation

[Ken Gaillot]

On Fri, 2017-11-24 at 11:32 +0000, Alan Birtles wrote:
> Is it possible to setup pacemaker with a resource which will only run
> on the same machine as another resource but that will not trigger a
> failover of the second resource if the first resource becomes
> unrunnable?
> e.g.
> I have 2 resources A and B. Resource A is the main resource and B is
> some monitoring resource which should only be run on the same machine
> as A but if B fails to run on a machine it shouldn't trigger a
> movement of A to another machine. If A fails on a machine it should
> move to another machine and then B should also move to that machine.
> I've tried setting a low score on the colocation constraint which
> does stop A from moving when B fails but this then allows B to start
> on a machine which isn't running A.

Pacemaker will place A on a node first, but it will take B's
preferences into account due to the colocation. If B becomes
unrunnable, its resulting aversion to the current node can cause A to
move.

There's no direct way to say what you want, but I can think of a few
ways to make it happen:

* If you set A's stickiness to INFINITY, that should override any
preference from B. I'd set A's priority meta-attribute higher than B's
as well.

* If you set on-fail=stop for B's monitor, the cluster will stop B
rather than try to restart and/or move it if it fails.

* The ocf:pacemaker:attribute agent, which sets an arbitrary node
attribute to a particular value depending on whether it is started or
stopped, was added for cases like this. You can create a group of A
plus the attribute, then create a location constraint with a rule
allowing B to run only where the attribute is set as started. This way,
A is unaware of the relationship and ignores B's preferences.
-- 
Ken Gaillot <kgaillot at redhat.com>