Understand different ways Karpenter deprovisions nodes

Karpenter sets a Kubernetes finalizer on each node it provisions. The finalizer specifies additional actions the Karpenter controller will take in response to a node deletion request. These include:

  • Marking the node as unschedulable, so no further pods can be scheduled there.
  • Evicting all pods other than daemonsets from the node.
  • Terminating the instance from the cloud provider.
  • Deleting the node from the Kubernetes cluster.

How Karpenter nodes are deprovisioned

There are both automated and manual ways of deprovisioning nodes provisioned by Karpenter:

  • Provisioner Deletion: Nodes are considered to be “owned” by the Provisioner that launched them. Karpenter will gracefully terminate nodes when a provisioner is deleted. Nodes may be reparented to another provisioner by modifying their labels. For example: kubectl label node -l --overwrite.

  • Node empty: Karpenter notes when the last workload (non-daemonset) pod stops running on a node. From that point, Karpenter waits the number of seconds set by ttlSecondsAfterEmpty in the provisioner, then Karpenter requests to delete the node. This feature can keep costs down by removing nodes that are no longer being used for workloads.

  • Node expired: Karpenter requests to delete the node after a set number of seconds, based on the provisioner ttlSecondsUntilExpired value, from the time the node was provisioned. One use case for node expiry is to handle node upgrades. Old nodes (with a potentially outdated Kubernetes version or operating system) are deleted, and replaced with nodes on the current version (assuming that you requested the latest version, rather than a specific version).

  • Node deleted: You could use kubectl to manually remove a single Karpenter node:

    # Delete a specific node
    kubectl delete node $NODE_NAME
    # Delete all nodes owned any provisioner
    kubectl delete nodes -l
    # Delete all nodes owned by a specific provisioner
    kubectl delete nodes -l$PROVISIONER_NAME

Whether through node expiry or manual deletion, Karpenter seeks to follow graceful termination procedures as described in Kubernetes Graceful node shutdown documentation. If the Karpenter controller is removed or fails, the finalizers on the nodes are orphaned and will require manual removal.

What can cause deprovisioning to fail?

There are a few cases where requesting to deprovision a Karpenter node will fail. These include Pod Disruption Budgets and pods that have the do-not-evict annotation set.

Disruption budgets

Karpenter respects Pod Disruption Budgets (PDBs) by using a backoff retry eviction strategy. Pods will never be forcibly deleted, so pods that fail to shut down will prevent a node from deprovisioning. Kubernetes PDBs let you specify how much of a Deployment, ReplicationController, ReplicaSet, or StatefulSet must be protected from disruptions when pod eviction requests are made.

PDBs can be used to strike a balance by protecting the application’s availability while still allowing a cluster administrator to manage the cluster. Here is an example where the pods matching the label myapp will block node termination if evicting the pod would reduce the number of available pods below 4.

apiVersion: policy/v1
kind: PodDisruptionBudget
  name: myapp-pdb
  minAvailable: 4
      app: myapp

You can set minAvailable or maxUnavailable as integers or as a percentage. Review what disruptions are, and how to configure them.

Pod set to do-not-evict

If a pod exists with the annotation true on a node, and a request is made to delete the node, Karpenter will not drain any pods from that node or otherwise try to delete the node. This annotation will have no effect for static pods, pods that tolerate NoSchedule, or pods terminating past their graceful termination period.

This is useful for pods that you want to run from start to finish without interruption. Examples might include a real-time, interactive game that you don’t want to interrupt or a long batch job (such as you might have with machine learning) that would need to start over if it were interrupted.

If you want to terminate a node with a do-not-evict pod, you can simply remove the annotation and the deprovisioning process will continue.

Last modified August 10, 2022 : v0.14.0 release (51a360f7)