Pod anti-affinity rules are not set up properly for Kafka Node Pools

Symptoms

You run the kubectl get pods -n <namespace> -o wide command to check the Kafka broker pods and see an output similar to the following:

kubectl get pods -n <namespace> -o wide

NAME                     READY   STATUS    RESTARTS   AGE     IP               NODE                       NOMINATED NODE   READINESS GATES
affinity-controller-0    1/1     Running   0          2m46s   10.144.16.142    worker0.my-cluster.cp      <none>           <none>
affinity-controller-1    1/1     Running   0          2m46s   10.144.28.180    worker0.my-cluster.cp      <none>           <none>
affinity-xhmp6           1/1     Running   0          2m42s   10.144.28.136    worker0.my-cluster.cp      <none>           <none>

Kafka broker pods are scheduled on the same worker node, which reduces fault tolerance and resilience.

Causes

In Event Streams 12.0.x, the default pod anti-affinity rules are not optimized for KRaft-based deployments. If no rules are configured, the Kubernetes scheduler might place multiple Kafka broker pods on the same node.

Resolving the problem

To ensure Kafka broker pods are distributed across multiple nodes, you must configure custom pod anti-affinity rules manually in each Kafka Node Pool.

Update your EventStreams custom resource to include anti-affinity rules under spec.strimziOverrides.nodePools.template.pod.affinity.

For example:

apiVersion: eventstreams.ibm.com/v1beta2
kind: EventStreams
# ...
spec:
  strimziOverrides:
    kafka:
      # ...
    nodePools:
      - name: kafka
        replicas: 2
        roles:
          - broker
        storage:
          # ...
        template:
          pod:
            affinity:
              podAntiAffinity:
                preferredDuringSchedulingIgnoredDuringExecution:
                  - weight: 10
                    podAffinityTerm:
                      labelSelector:
                        matchExpressions:
                          - key: eventstreams.ibm.com/pool-name
                            operator: In
                            values:
                              - kafka
                      topologyKey: kubernetes.io/hostname
                  - weight: 5
                    podAffinityTerm:
                      labelSelector:
                        matchExpressions:
                          - key: eventstreams.ibm.com/cluster
                            operator: In
                            values:
                              - min-prod-scram
                          - key: eventstreams.ibm.com/component-type
                            operator: In
                            values:
                              - kafka
                      topologyKey: kubernetes.io/hostname

The previous example defines two preferred pod anti-affinity rules that improve Kafka pod distribution across nodes:

Rule 1 (weight 10): Aims to place pods from the same node pool on different hosts.
Rule 2 (weight 5): Aims to place Kafka pods from the same Event Streams instance on different hosts.