clarify behavior of node affinity in the context of evictable pods

[irl-segfault]

Hi

It is unclear based on the documentation what happens when a pod with a preferredDuringSchedulingIgnoredDuringExecution affinity has a choice between

• evicting a lower priority pod on the preferred node type, and taking its place
• not doing that, and scheduling onto a node that does not match the preferred criteria

I believe the latter is the default behavior. It would be nice to explicitly call this out in documentation, or better yet, make it configurable. I am trying to run a buffer pod to reserve space on preferred nodes to hedge against scale up delays, and would prefer to always evict such a pod instead of schedule the deploying workload onto non-preferred nodes.

Thanks! :)

[MaciekPytel]

This seems to be an issue related to scheduler and not Cluster Autoscaler. As such it should probably be opened in kubernetes/kubernetes repo and tagged to sig-scheduling.

Cluster Autoscaler only cares about making each pod schedulable, it completely ignores scheduler priority functions (ie. it respects requiredDuringScheduling, it doesn't care whatsoever about preferredDuringScheduling).