Bad scheduling checker

[ArchangelSDY]

Background

Sometimes pods can be scheduled in a non-HA way. For example, all pods in a deployment scheduled to a single machine, or a single zone. Authoring a deployment with HA is hard for beginners. We should be able to point out such risks and guide users to configure pod spread topology properly.

Goals

• Detect bad scheduling issues
  • All pods in a deployment scheduled to a single machine.
  • All pods in a deployment scheduled to a single zone, if cluster supports AZ.
• Give general guidance about configuring pod anti-affinity and spread topology

Non-goals

• No need to detect theoretical issues. Just check current status.

References

• https://kubernetes.io/docs/concepts/scheduling-eviction/assign-pod-node/#affinity-and-anti-affinity
• https://kubernetes.io/docs/concepts/scheduling-eviction/assign-pod-node/#pod-topology-spread-constraints

[zackliu]

A test spec

apiVersion: apps/v1
kind: Deployment
metadata:
  name: nginx-deployment
  labels:
    app: nginx
spec:
  replicas: 3
  selector:
    matchLabels:
      app: nginx
  template:
    metadata:
      labels:
        app: nginx
    spec:
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
            - matchExpressions:
              - key: kubernetes.io/hostname
                operator: In
                values:
                - <node-name>
      containers:
      - name: nginx
        image: nginx:1.14.2
        ports:
        - containerPort: 80