[Feature Proposal] game server eviction should only take allocated servers into consideration

[swermin]

Is your feature request related to a problem? Please describe.
Our game servers do not support SIGTERM, so we cannot run with eviction=OnUpgrade.

Our fleets are cycling through game servers within an hour, but due to the scale of game servers we have running letting these churn through all of the pods will take days. So this make our node upgrade job a very manual and tedious operation.

Describe the solution you'd like
It would be nice for the eviction=OnUpgrade to only add the PDB for actual Allocated servers, effectively allowing the control plane to reschedule Ready pods to new nodes.
This would allow the fleet to eventually move all game servers to new nodes without having to wait for Ready to go through allocation state.

[markmandel]

Our game servers do not support SIGTERM, so we cannot run with eviction=OnUpgrade.

You could make this work if you wanted to 😁 sidecar process maybe? wrapper around the game server that starts the game server process?

What the issue with this suggestion of only Allocated server is though, is that there is an inherent race condition where labels need to be applied to the GameServer Pod on allocation (which is why the original design of making everything not evictable), before a K8s event marks it for deletion because of an upgrade while it's being potentially allocated at the same time (or just before!).

Same issue with using the cluster autoscaler either. It's a whole can of worms.

Thoughts though? Maybe I'm missing something.