The problem:

We have nodes in kubernetes, which will occasionally become tainted with an effect of PreferNoSchedule. When this happens, we would like our pods to completely avoid scheduling on these nodes (in other words, act like the taint's effect were actually NoSchedule). The taints aren't applied by us - we're using GKE and it's an automated thing on their end, so we're stuck with the PreferNoSchedule behaviour.

What we can control is the spec of our pods. I'm hoping this might be possible using a nodeAffinity on them, however the documentation on this is fairly sparse: see e.g. here. All the examples I can find refer to an affinity by labels, so I'm not sure if a taint is even visible/accessible by this logic.

Effectively, in my pod spec I want to write something like:

spec:
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: kubernetes.io/taints *I've invented this, it doesn't work
            operator: NotIn
            values:
            - DeletionCandidateOfClusterAutoscaler

where DeletionCandidateOfClusterAutoscaler is the taint that we see applied. Is there a way to make this work?

The other approach we've thought about is a cronjob which looks for the PreferNoSchedule taint and adds our own NoSchedule taint on top... but that feels a little gross!

Any neat suggestions / workarounds would be appreciated!

The long version (if you're interested):

The taint gets applied by the autoscaler to say the node is going to go away in 30 minutes or so. This issue describes in some more detail, from people having the same trouble: https://github.com/kubernetes/autoscaler/issues/2967