Kubernetes topologySpreadConstraints doesn't distribute 3 pod evenly to 3 zone ( a b c), but places 1 to a, and 2 to b)

Question

I'm trying to spread my pods to different availability zones. I used topologySpreadConstraints feature with 3 replicas and such config:

  topologySpreadConstraints:
    - maxSkew: 1
      topologyKey: topology.kubernetes.io/zone
      whenUnsatisfiable: DoNotSchedule
      labelSelector:
          matchLabels:
              release: {{ .Release.Name }}

Worker nodes are available in 3 zones. The problem is when I deploy it, scheduler palces one pod to A-zone , and 2 pods in B-zone . Instead of placing one into each zone. I understand that maxSkew: 1 , can't be set to 0 , and it allows scheduler to schedule unevenly with +1 skewness. But shouldn't the scheduler respect domain first? Like shouldn't it do best effort to spread 3 pods by different zones (domains) first? Can it be done with help of topologySpreadConstraints ? Or I should use podAntiAffinity ?

Answer 1

Like you mentioned, maxSkew can't be set to 0. My guess is that it's done to handle odd number of pods. In your case: one pod in A and two pods in B still respects the maxSkew of one so the scheduler is doing its job properly. If you know for a fact that your replica count will always be 3 (maybe on a statefulset), you can use a podAntiAffinity rule instead of your topologySpreadContraints but this will not work with more than one pod per zone. If you use a deployment, the simplest solution would be to increase the replica count. Something like 5 would do the trick.