Troubleshoot Karpenter problems

Unknown field in Provisioner spec

If you are upgrading from an older version of Karpenter, there may have been changes in the CRD between versions. Attempting to utilize newer functionality which is surfaced in newer versions of the CRD may result in the following error message:

error: error validating "STDIN": error validating data: ValidationError(Provisioner.spec): unknown field "<fieldName>" in sh.karpenter.v1alpha5.Provisioner.spec; if you choose to ignore these errors, turn validation off with --validate=false

If you see this error, you can solve the problem by following the Custom Resource Definition Upgrade Guidance.

Info on whether there has been a change to the CRD between versions of Karpenter can be found in the Release Notes

Unable to schedule pod due to insufficient node group instances

v0.16.0 changed the default replicas from 1 to 2.

Karpenter won’t launch capacity to run itself (log related to the DoesNotExist requirement) so it can’t provision for the second Karpenter pod.

To solve this you can either reduce the replicas back from 2 to 1, or ensure there is enough capacity that isn’t being managed by Karpenter (these are instances with the name<PROVIDER_NAME>) to run both pods.

To do so on AWS increase the minimum and desired parameters on the node group autoscaling group to launch at lease 2 instances.

Node not created

In some circumstances, Karpenter controller can fail to start up a node. For example, providing the wrong block storage device name in a custom launch template can result in a failure to start the node and an error similar to:

2022-01-19T18:22:23.366Z ERROR controller.provisioning Could not launch node, launching instances, with fleet error(s), InvalidBlockDeviceMapping: Invalid device name /dev/xvda; ...

You can see errors like this by viewing Karpenter controller logs:

kubectl get pods -A | grep karpenter
karpenter     karpenter-XXXX   2/2     Running   2          21d
kubectl logs karpenter-XXXX -c controller -n karpenter | less

Node NotReady

There are cases where the node starts, but fails to join the cluster and is marked “Node NotReady”. Reasons that a node can fail to join the cluster include:

  • Permissions
  • Security Groups
  • Networking

The easiest way to start debugging is to connect to the instance and get the Kubelet logs. For an AL2 based node:

# List the nodes managed by Karpenter
kubectl get node -l
# Extract the instance ID (replace <node-name> with a node name from the above listing)
INSTANCE_ID=$(kubectl get node <node-name> -ojson | jq -r ".spec.providerID" | cut -d \/ -f5)
# Connect to the instance
aws ssm start-session --target $INSTANCE_ID
# Check Kubelet logs
sudo journalctl -u kubelet

For Bottlerocket, you’ll need to get access to the root filesystem:

# List the nodes managed by Karpenter
kubectl get node -l
# Extract the instance ID (replace <node-name> with a node name from the above listing)
INSTANCE_ID=$(kubectl get node <node-name> -ojson | jq -r ".spec.providerID" | cut -d \/ -f5)
# Connect to the instance
aws ssm start-session --target $INSTANCE_ID
# Enter the admin container
# Run sheltie
sudo sheltie
# Check Kubelet logs
journalctl -u kubelet

Here are examples of errors from Node NotReady issues that you might see from journalctl:

  • The runtime network not being ready can reflect a problem with IAM role permissions:

    KubeletNotReady runtime network not ready: NetworkReady=false reason:NetworkPluginNotReady message:Network plugin returns error: cni plugin not initialized

    See Amazon EKS node IAM role for details. If you’re using eksctl, the VPC CNI pods may be given permissions through IRSA instead. Verify that this set up is working as intended. You can also look at the logs for your CNI plugin from the aws-node pod:

    kubectl get pods -n kube-system | grep aws-node
    aws-node-?????             1/1     Running   2          20d
    kubectl logs aws-node-????? -n kube-system
  • Not being able to register the node with the Kubernetes API server indicates an error condition like the following:

    Attempting to register node" node="ip-192-168-67-130.ec2.internal"
    Unable to register node with API server" err="Unauthorized" node="ip-192-168-67-130.ec2.internal"
    Error getting node" err="node \"ip-192-168-67-130.ec2.internal\" not found
    Failed to contact API server when waiting for CSINode publishing: Unauthorized

    Check the ConfigMap to check whether or not the correct node role is there. For example:

    $ kubectl get configmaps -n kube-system aws-auth -o yaml
    apiVersion: v1
    mapRoles: |
       - groups:
          - system:bootstrappers
          - system:nodes
          rolearn: arn:aws:iam::973227887653:role/eksctl-johnw-karpenter-demo-NodeInstanceRole-72CV61KQNOYS
          username: system:node:{{EC2PrivateDNSName}}
       - groups:
          - system:bootstrappers
          - system:nodes
          rolearn: arn:aws:iam::973227887653:role/KarpenterNodeRole-johnw-karpenter-demo
          username: system:node:{{EC2PrivateDNSName}}   
    mapUsers: |
    kind: ConfigMap

If you are not able to resolve the Node NotReady issue on your own, run the EKS Logs Collector (if it’s an EKS optimized AMI) and look in the following places in the log:

  • Your UserData (in /var_log/cloud-init-output.log and /var_log/cloud-init.log)
  • Your kubelets (/kubelet/kubelet.log)
  • Your networking pod logs (/var_log/aws-node)

Reach out to the Karpenter team on Slack or GitHub if you are still stuck.

Missing Service Linked Role

Unless your AWS account has already onboarded to EC2 Spot, you will need to create the service linked role to avoid ServiceLinkedRoleCreationNotPermitted.

AuthFailure.ServiceLinkedRoleCreationNotPermitted: The provided credentials do not have permission to create the service-linked role for EC2 Spot Instances

This can be resolved by creating the Service Linked Role.

aws iam create-service-linked-role --aws-service-name

Unable to delete nodes after uninstalling Karpenter

Karpenter adds a finalizer to nodes that it provisions to support graceful node termination. If Karpenter is uninstalled, these finalizers will cause the API Server to block deletion until the finalizers are removed.

You can fix this by patching the node objects:

  • kubectl edit node <node_name> and remove the line that says in the finalizers field.
  • Run the following script that gets all nodes with the finalizer and removes all the finalizers from those nodes.
    • NOTE: this will remove ALL finalizers from nodes with the karpenter finalizer.
kubectl get nodes -ojsonpath='{range .items[*].metadata}{}:{@.finalizers}{"\n"}' | grep "" | cut -d ':' -f 1 | xargs kubectl patch node --type='json' -p='[{"op": "remove", "path": "/metadata/finalizers"}]'

Nil issues with Karpenter reallocation

If you create a Karpenter Provisioner while the webhook to default it is unavailable, it’s possible to get unintentionally nil fields. Related Issue.

You may see some logs like this. +0x688*Controller).Reconcile(0xc000b004c0, 0x23354c0, 0xc000e209f0, 0x235e640, 0xc002566c40, 0x200c786, 0x5, 0xc00259c1b0, 0x1) +0x65*GenericController).Reconcile(0xc000b00720, 0x23354c0, 0xc000e209f0, 0xc001db9be0, 0x7, 0xc001db9bd0, 0x7, 0xc000e209f0, 0x7fc864172d20, 0xc0000be2a0, ...)

This is fixed in Karpenter v0.2.7+. Reinstall Karpenter on the latest version.

Nodes stuck in pending and not running the kubelet due to outdated CNI

If you have an EC2 instance get launched that is stuck in pending and ultimately not running the kubelet, you may see a message like this in your /var/log/user-data.log:

No entry for c6i.xlarge in /etc/eks/eni-max-pods.txt

This means that your CNI plugin is out of date. You can find instructions on how to update your plugin here.

Failed calling webhook “”

If you are not able to create a provisioner due to Internal error occurred: failed calling webhook "":

Webhooks were renamed in v0.19.0. There’s a bug in ArgoCD’s upgrade workflow where webhooks are leaked. This results in Provisioner’s failing to be validated, since the validation server no longer corresponds to the webhook definition.

Delete the stale webhooks.

kubectl delete mutatingwebhookconfigurations
kubectl delete validatingwebhookconfiguration

Failed calling webhook “”

If you are not able to create a provisioner due to Error from server (InternalError): error when creating "provisioner.yaml": Internal error occurred: failed calling webhook "": Post "https://karpenter-webhook.karpenter.svc:443/default-resource?timeout=10s": context deadline exceeded

Verify that the karpenter pod is running (should see 2/2 containers with a “Ready” status)

kubectl get po -A -l
NAME                       READY   STATUS    RESTARTS   AGE
karpenter-7b46fb5c-gcr9z   2/2     Running   0          17h

Karpenter service has endpoints assigned to it

kubectl get ep -A -l
NAMESPACE   NAME        ENDPOINTS                               AGE
karpenter   karpenter,   16d

Your security groups are not blocking you from reaching your webhook.

This is especially relevant if you have used terraform-eks-module version >=18 since that version changed its security approach, and now it’s much more restrictive.

DaemonSets can result in deployment failures

For Karpenter versions 0.5.3 and earlier, DaemonSets were not properly considered when provisioning nodes. This sometimes caused nodes to be deployed that could not meet the needs of the requested DaemonSets and workloads. This issue no longer occurs after Karpenter version 0.5.3 (see PR #1155).

If you are using a pre-0.5.3 version of Karpenter, one workaround is to set your provisioner to only use larger instance types that you know will be big enough for the DaemonSet and the workload. For more information, see Issue #1084. Examples of this behavior are included in Issue #1180.

Unspecified resource requests cause scheduling/bin-pack failures

Not using the Kubernetes LimitRanges feature to enforce minimum resource request sizes will allow pods with very low or non-existent resource requests to be scheduled. This can cause issues as Karpenter bin-packs pods based on the resource requests.

If the resource requests do not reflect the actual resource usage of the pod, Karpenter will place too many of these pods onto the same node resulting in the pods getting CPU throttled or terminated due to the OOM killer. This behavior is not unique to Karpenter and can also occur with the standard kube-scheduler with pods that don’t have accurate resource requests.

To prevent this, you can set LimitRanges on pod deployments on a per-namespace basis. See the Karpenter Best Practices Guide for further information on the use of LimitRanges.

Missing subnetSelector and securityGroupSelector tags causes provisioning failures

Starting with Karpenter v0.5.5, if you are using Karpenter-generated launch template, provisioners require that subnetSelector and securityGroupSelector tags be set to match your cluster. The Provisioner section in the Karpenter Getting Started Guide uses the following example:

kind: AWSNodeTemplate
  subnetSelector: ${CLUSTER_NAME}
  securityGroupSelector: ${CLUSTER_NAME}

To check your subnet and security group selectors, type the following:

aws ec2 describe-subnets --filters,Values=${CLUSTER_NAME}

Returns subnets matching the selector

aws ec2 describe-security-groups --filters,Values=${CLUSTER_NAME}

Returns security groups matching the selector

Provisioners created without those tags and run in more recent Karpenter versions will fail with this message when you try to run the provisioner:

 field(s): spec.provider.securityGroupSelector, spec.provider.subnetSelector

Terraform fails to create instance profile when name is too long

In the Getting Started with Terraform instructions to Configure the KarpenterNode IAM Role, the name assigned to the aws_iam_instance_profile cannot exceed 38 characters. If it does, it will fail with a message similar to:

Error: expected length of name_prefix to be in the range (1 - 38), got with module.eks.aws_iam_role.cluster[0],
on .terraform/modules/eks/ line 131, in resource "aws_iam_role" "cluster":
131: name_prefix = var.cluster_iam_role_name != "" ? null : var.cluster_name

Note that it can be easy to run over the 38-character limit considering that the example includes KarpenterNodeInstanceProfile- (29 characters) and -karpenter-demo (15 characters). That leaves only four characters for your user name. You can reduce the number of characters consumed by changing KarpenterNodeInstanceProfile- to something like KarpenterNode-.

Karpenter Role names exceeding 64-character limit

If you use a tool such as AWS CDK to generate your Kubernetes cluster name, when you add Karpenter to your cluster you could end up with a cluster name that is too long to incorporate into your KarpenterNodeRole name (which is limited to 64 characters).

Node role names for Karpenter are created in the form KarpenterNodeRole-${Cluster_Name} in the Create the KarpenterNode IAM Role section of the getting started guide. If a long cluster name causes the Karpenter node role name to exceed 64 characters, creating that object will fail.

Keep in mind that KarpenterNodeRole- is just a recommendation from the getting started guide. Instead using of the eksctl role, you can shorten the name to anything you like, as long as it has the right permissions.

Node terminates before ready on failed encrypted EBS volume

If you are using a custom launch template and an encrypted EBS volume, the IAM principal launching the node may not have sufficient permissions to use the KMS customer managed key (CMK) for the EC2 EBS root volume. This issue also applies to Block Device Mappings specified in the Provisioner. In either case, this results in the node terminating almost immediately upon creation.

Keep in mind that it is possible that EBS Encryption can be enabled without your knowledge. EBS encryption could have been enabled by an account administrator or by default on a per region basis. See Encryption by default for details.

To correct the problem if it occurs, you can use the approach that AWS EBS uses, which avoids adding particular roles to the KMS policy. Below is an example of a policy applied to the KMS key:

        "Sid": "Allow access through EBS for all principals in the account that are authorized to use EBS",
        "Effect": "Allow",
        "Principal": {
            "AWS": "*"
        "Action": [
        "Resource": "",
        "Condition": {
            "StringEquals": {
            "kms:ViaService": "ec2.${AWS_REGION}",
            "kms:CallerAccount": "${AWS_ACCOUNT_ID}"
        "Sid": "Allow direct access to key metadata to the account",
        "Effect": "Allow",
        "Principal": {
            "AWS": "arn:aws:iam::${AWS_ACCOUNT_ID}:root"
        "Action": [
        "Resource": "*"

Pods using Security Groups for Pods stuck in “ContainerCreating” state for up to 30 minutes before transitioning to “Running”

When leveraging Security Groups for Pods, Karpenter will launch nodes as expected but pods will be stuck in “ContainerCreating” state for up to 30 minutes before transitioning to “Running”. This is related to an interaction between Karpenter and the amazon-vpc-resource-controller when a pod requests resources. More info can be found in issue #1252.

To workaround this problem, add the "false" label in your Karpenter Provisioner spec and ensure instance-type requirements include instance-types which support ENI trunking.

kind: Provisioner
  name: default
  labels: "false"
  ttlSecondsAfterEmpty: 30

Stale pricing data on isolated subnet

The following pricing-related error occurs if you are running Karpenter in an isolated private subnet (no Internet egress via IGW or NAT gateways):

ERROR  updating on-demand pricing, RequestError: send request failed
caused by: Post "": dial tcp i/o timeout; RequestError: send request failed
caused by: Post "": dial tcp i/o timeout, using existing pricing data from 2022-08-17T00:19:52Z  {"commit": "4b5f953"}

This network timeout occurs because there is no VPC endpoint available for the Price List Query API.. To workaround this issue, Karpenter ships updated on-demand pricing data as part of the Karpenter binary; however, this means that pricing data will only be updated on Karpenter version upgrades. To disable pricing lookups and avoid the error messages, set the AWS_ISOLATED_VPC environment variable (or the --aws-isolated-vpc option) to true. See Environment Variables / CLI Flags for details.

Helm Error When Pulling the Chart

If Helm is showing an error when trying to install Karpenter helm charts:

  • Ensure you are using a newer Helm version, Helm started supporting OCI images since v3.8.0.
  • Verify that the image you are trying to pull actually exists in
  • Sometimes Helm generates a generic error, you can add the –debug switch to any of the helm commands in this doc for more verbose error messages
  • If you are getting a 403 forbidden error, you can try docker logout as explained here
  • If you are receiving this error: Error: failed to download "oci://" at version "0.17.0", then you need to prepend a v to the version number: v0.17.0. Before Karpenter moved to OCI helm charts (pre-v0.17.0), both v0.16.0 and 0.16.0 would work, but OCI charts require an exact version match.