The kube-scheduler can panic with the error "Observed a panic: "integer divide by zero" (runtime error: integer divide by zero)" when a Pod specifies a Node selector that has no matching nodes.

This issue is only present in OCP 4.14, likely due to this bug being introduced and later fixed upstream [1].

More specifically, I could note reproduce the issue on 4.12, 4.13, or 4.15; it seems to only occur on 4.14, which is likely post-regression and pre-bugfix.

OCP 4.14

Always

apiVersion: v1
kind: Pod
metadata:
  name: break-kube-scheduler
spec:
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchFields:
          - key: metadata.name
            operator: In
            values:
            - invalid-node # a node that doesn't exist
  containers:
  - name: main
    image: image-registry.openshift-image-registry.svc:5000/openshift/ubi:latest
    command: ["cat"]
    stdin: true

E0701 15:35:25.419149 1 schedule_one.go:158] "Error selecting node for pod" err="nodeinfo not found for node name \"invalid-node\"" pod="openshift-kube-scheduler/break-kube-scheduler"
E0701 15:35:25.419196 1 schedule_one.go:891] "Error scheduling pod; retrying" err="nodeinfo not found for node name \"invalid-node\"" pod="openshift-kube-scheduler/break-kube-scheduler"