Prevent instance migration

[bensmrs]

Instance placement scriptlets offer interesting ways to control how instances are placed on cluster nodes. While creation, relocation (due to a node being down), and evacuation are handled, there is to the best of my knowledge no way to block a migration to an unwanted cluster node. So I have two questions:

• How to lock an instance to a specific node in a cluster, without creating a dedicated project attached to a dedicated group?
• Could the instance placement scriptlet be modified so that it includes an optional parameter (or request property) target containing the requested migration target, allowing us to call fail depending on different decision factors? (I can work on this one)

[stgraber]

How to lock an instance to a specific node in a cluster, without creating a dedicated project attached to a dedicated group?

Easiest way to do this would be to use --target during creation and then set cluster.evacuate to stop so it never gets migrated away during evacuation or recovery.

Could the instance placement scriptlet be modified so that it includes an optional parameter (or request property) target containing the requested migration target, allowing us to call fail depending on different decision factors? (I can work on this one)

I think that starting with Incus 6.6 we call the scriptlet even with --target=NAME so that should let you perform your check there.

As for failing, you should be able to either call the fail function or do return "Placement rejected" which will also be treated as an error.

[stgraber]

Marking as incomplete as we may not have anything to add to Incus for that.

[bensmrs]

Ooooh, I didn’t actually see that the candidate_members was actually populated with the --target! Maybe the documentation could be improved, as the behavior is not suggested.
However, set_target doesn’t work as I’d expect:

• incus move instance5 --target node3 --project test
• The scriptlet gets called, with node3 as its candidate_members singleton (first part of my answer; which I didn’t know happened)
• set_target(node5) does nothing; no error, no warning, and the instance is still moved to node3.

I get that set_target might not get called in this scenario, but maybe an error could help admins…

[stgraber]

Hmm, yeah, we should have set_target fail if it's passed a server that's not within the candidate_members.