Karpenter consolidation without any downtime for pods

Question

I have karpenter deployed in eks and for karpenter provisioner I have consolidation enabled true

  consolidation:
    enabled: true

which means for any underutilized node, the workload (pods) will be deployed in the existing nodes. but during this process I have noticed karpenter terminates the existing pod and then creates a new pod.

I am looking for a soluction where the new pod will the created in the existing node first and then terminate the pod in the underutilized node then terminate the node.

Answer 1

Configure Pod Disruption Budgets, https://kubernetes.io/docs/tasks/run-application/configure-pdb/

Answer 2

Karpenter consolidation works by cordoning a node and draining the pods from the node before deleting it. This causes the pods to be evicted and re-created by their controllers in order to be rescheduled. So, a downtime can get introduced.

The available possibilities are:

• If there is only 1 replica of the pod, then Pod Disruption Budget and annotation karpenter.sh/do-not-evict: true are the ways to stop the node from being terminated.
• Have multiple replicas of the pod and use node affinity, anti-affinity to have pod replicas placed in separate nodes to have minimum disruption during consolidation.