evenly distribute pods over nodes

Question

I've got an on-premise cluster with 1 master and 3 workers. Now there's a deployment with 3 replica's and they MUST be evenly distributed over all the nodes.

Now we can use "topologySpreadConstraints" on kubernetes/hostName label for that but it doesn't always work.

apiVersion: apps/v1
kind: Deployment
metadata:
 name: revproxy
 labels:
  app: proxy
spec:
 replicas: 3
 selector:
   matchLabels:
     role: proxy
 template:
   metadata:
     labels:
       role: proxy
   spec:
     topologySpreadConstraints:
     - maxSkew: 1
       topologyKey: kubernetes.io/hostname
       whenUnsatisfiable: ScheduleAnyway
       labelSelector:
         matchLabels:
           app: revproxy

There's many times 2 pods running on the same node and 1 pod on another and zero pods on the 3rd. But when manually deleting a second pod then it ALWAYS gets scheduled on the empty node.

How can I use topologySpreadConstraints to have pods evenly distributed over the nodes?

Answer 1

Found the answer.

When the deployment has a yaml with the following labels.

kind: Deployment
metadata:
 name: revproxy
 labels:
  app: proxy
spec:
 selector:
   matchLabels:
     role: proxy
 template:
   metadata:
     labels:
       role: proxy

The matchLabels in topologySpreadConstraints with matchLabels 'revproxy' is wrong.

topologySpreadConstraints:
     - maxSkew: 1
       topologyKey: kubernetes.io/hostname
       whenUnsatisfiable: ScheduleAnyway
       labelSelector:
         matchLabels:
           app: revproxy

When pointing to just 'proxy' in the labelSelector the pods get distributed over the 3 nodes evenly.

     matchLabels:
       app: proxy

Am still confused about the 3x the same labels in a description, changing any of these breaks the deployment, but at least it works.