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 karpenter.sh/provisioner-name=source-provisioner-name karpenter.sh/provisioner-name=destination-provisioner-name --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
ttlSecondsAfterEmptyin 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
ttlSecondsUntilExpiredvalue, 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).
- Consolidation: Karpenter works to actively reduce cluster cost by identifying when nodes can be removed as their workloads will run on other nodes in the cluster and when nodes can be replaced with cheaper variants due to a change in the workloads.
Automated deprovisioning is configured through the ProvisionerSpec
.consolidation.enabledfields. If these are not configured, Karpenter will not default values for them and will not terminate nodes for that purpose.
Keep in mind that a small NodeExpiry results in a higher churn in cluster activity. So, for example, if a cluster brings up all nodes at once, all the pods on those nodes would fall into the same batching window on expiration.
Note that Karpenter does not automatically add jitter to this value. If multiple instances are created in a small amount of time, they will expire at very similar times. Consider defining a pod disruption budget to prevent excessive workload disruption.
Using preferred anti-affinity and topology spreads can reduce the effectiveness of consolidation. At node launch, Karpenter attempts to satisfy affinity and topology spread preferences. In order to reduce node churn, consolidation must also attempt to satisfy these constraints to avoid immediately consolidating nodes after they launch. This means that consolidation may not deprovision nodes in order to avoid violating preferences, even if kube-scheduler can fit the host pods elsewhere.
Node deleted: You could use
kubectlto manually remove a single Karpenter node:
# Delete a specific node kubectl delete node $NODE_NAME # Delete all nodes owned any provisioner kubectl delete nodes -l karpenter.sh/provisioner-name # Delete all nodes owned by a specific provisioner kubectl delete nodes -l karpenter.sh/provisioner-name=$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.
NoteBy adding the finalizer, Karpenter improves the default Kubernetes process of node deletion. When you run
kubectl delete nodeon a node without a finalizer, the node is deleted without triggering the finalization logic. The instance will continue running in EC2, even though there is no longer a node object for it. The kubelet isn’t watching for its own existence, so if a node is deleted the kubelet doesn’t terminate itself. All the pod objects get deleted by a garbage collection process later, because the pods’ node is gone.
Karpenter has two mechanisms for cluster consolidation:
- Deletion - A node is eligible for deletion if all of its pods can run on free capacity of other nodes in the cluster.
- Replace - A node can be replaced if all of its pods can run on a combination of free capacity of other nodes in the cluster and a single cheaper replacement node.
When there are multiple nodes that could be potentially deleted or replaced, Karpenter choose to consolidate the node that overall disrupts your workloads the least by preferring to terminate:
- nodes running fewer pods
- nodes that will expire soon
- nodes with lower priority pods
NoteKarpenter only uses the deletion consolidation mechanism for spot nodes. It will not replace a spot node with a cheaper spot node. Spot instance types are selected with the
capacity-optimized-prioritizedstrategy and often the cheapest spot instance type is not launched due to the likelihood of interruption. Consolidation would then replace the spot instance with a cheaper instance negating the
capacity-optimized-prioritizedstrategy entirely and increasing interruption rate.
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.
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 metadata: name: myapp-pdb spec: minAvailable: 4 selector: matchLabels: app: myapp
Pod set to do-not-evict
If a pod exists with the annotation
karpenter.sh/do-not-evict: 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. Nodes that have pods with a
do-not-evict annotation are not considered for consolidation, though their unused capacity is considered for the purposes of running pods from other nodes which can ber consolidated. 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.
Scheduling Constraints (Consolidation Only)
Consolidation will be unable to consolidate a node if, as a result of its scheduling simulation, it determines that the pods on a node cannot run on other nodes due to inter-pod affinity/anti-affinity, topology spread constraints, or some other scheduling restriction that couldn’t be fulfilled.
Controllerless Pods (Consolidation Only)
Consolidation will not attempt to consolidate a node that is running pods that are not owned by a controller (e.g. a
ReplicaSet). In general we cannot assume that these pods would be recreated if they were evicted from the node that they are currently running on.