Abstract

• It often goes smoothly when using Karpenter with the fresh new EKS/kubernetes cluster as nodes are provisioned and able to join the cluster, or it works well on this region but not on others, why???

• This blog describes what issues I faced and how to troubleshoot them.

Table Of Contents

• Pre-requisite

• The issues

• Check IAM instance profile

• Check kube-proxy and aws-node Daemonset

• Check Security group and Elastic network interface

• Debug where the kubelet stuck at

• Check AWS service endpoints

• Conclusion

🚀 Pre-requisite

• Better understand Karpenter and hands-on

🚀 The issues

• The following might be one/more issues:

  • New nodes are failing to communicate with the API server

  • Lease: Failed to get lease: leases.coordination.k8s.io

  • No kube-proxy and aws-node Daemonset Pods are created on the Karpenter nodes.

  • kubelet failed to start so node is stuck in NotReady

• Following is the case-by-case that we need to troubleshoot to find out the rootcause

🚀 Check IAM instance profile

• The condition for an EKS node to join the cluster is IAM instance profile which provides IAM permission for the nodes to access AWS resources and call API to EKS cluster.

• Furthermore, we also need to add the instance profile to aws-auth configmap to grant system:masters permissions in the cluster's role-based access control (RBAC) configuration in the Amazon EKS control plane. Read more Enabling IAM user and role access to your cluster

  aws-auth configmap

    apiVersion: v1
    data:
      mapAccounts: |
        []
      mapRoles: |
        - "groups":
          - "system:nodes"
          - "system:bootstrappers"
          "rolearn": "arn:aws:iam::<accountID>:role/KarpenterNodeInstanceProfile"
          "username": "system:node:{{EC2PrivateDNSName}}"
  

• The right permissions that an instance profile needs

    AmazonEKSWorkerNodePolicy
    AmazonEC2ContainerRegistryReadOnly
    AmazonSSMManagedInstanceCore
    AmazonEKS_CNI_Policy
  

• Instance profile role's Trust Policy:

    {
        "Version": "2012-10-17",
        "Statement": [
            {
                "Sid": "EKSWorkerAssumeRole",
                "Effect": "Allow",
                "Principal": {
                    "Service": "ec2.amazonaws.com"
                },
                "Action": "sts:AssumeRole"
            }
        ]
    }
  

🚀 Check kube-proxy and aws-node Daemonset

• When a new node joins the cluster, 2 Daemonset Pods must be created on it

  • kube-system which provides critical networking functionality for Kubernetes applications

  • aws-node which is the Amazon VPC Container Network Interface (CNI) plugin for Kubernetes If not some checks should be performed

So we need to verify those pods are in Running status on the worker node. If not, perform the following checks.

🚀 Check Security group and Elastic network interface

• We need to ensure traffic between Pods because the coredns Pods are in Deployment type and created on a number of nodes that the new ones must be allowed traffic to

• Checkout Understand Pods communication

🚀 Debug where the kubelet stuck at

• If there are no kube-proxy and aws-node Daemonsets on the node, the kubelet failed to start. Look at the kubelet logs

    Apr 23 05:26:38 ip-10-10-11-100 containerd: time="2022-04-23T05:26:38.814171952Z" level=error msg="failed to load cni during init, please check CRI plugin status before setting up network for pods" error="cni config load failed: no network config found in /etc/cni/net.d: cni plugin not initialized: failed to load cni config"
  

• Karpenter creates the node object, and then Kubelet populates the necessary labels once it can connect to the APIServer.

  • kube-proxy nodeAffinity

      spec:
        affinity:
          nodeAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
              nodeSelectorTerms:
              - matchExpressions:
                - key: beta.kubernetes.io/os
                  operator: In
                  values:
                  - linux
                - key: beta.kubernetes.io/arch
                  operator: In
                  values:
                  - amd64
    

• Due to kubelet failed to start so the node does not contain those labels

    # kubectl describe node ip-10-10-11-198.ap-southeast-1.compute.internal
    Name:               ip-10-10-11-198.ap-southeast-1.compute.internal
    Roles:              <none>
    Labels:             karpenter.sh/capacity-type=spot
                        karpenter.sh/provisioner-name=karpenter-test
                        lifecycle=spot
                        node.kubernetes.io/instance-type=t3.large
                        role=karpenter
                        topology.kubernetes.io/zone=ap-southeast-1b
                        type=test
    Annotations:        node.alpha.kubernetes.io/ttl: 0
    CreationTimestamp:  Fri, 22 Apr 2022 15:37:34 +0000
    Taints:             node.kubernetes.io/unreachable:NoExecute
                        dedicated=karpenter:NoSchedule
                        karpenter.sh/not-ready:NoSchedule
                        node.kubernetes.io/unreachable:NoSchedule
    Unschedulable:      false
    Lease:              Failed to get lease: leases.coordination.k8s.io "ip-10-10-11-198.ap-southeast-1.compute.internal" not found
    Conditions:
      Type             Status    LastHeartbeatTime                 LastTransitionTime                Reason                   Message
      ----             ------    -----------------                 ------------------                ------                   -------
      Ready            Unknown   Fri, 22 Apr 2022 15:37:34 +0000   Fri, 22 Apr 2022 15:38:34 +0000   NodeStatusNeverUpdated   Kubelet never posted node status.
      MemoryPressure   Unknown   Fri, 22 Apr 2022 15:37:34 +0000   Fri, 22 Apr 2022 15:38:34 +0000   NodeStatusNeverUpdated   Kubelet never posted node status.
      DiskPressure     Unknown   Fri, 22 Apr 2022 15:37:34 +0000   Fri, 22 Apr 2022 15:38:34 +0000   NodeStatusNeverUpdated   Kubelet never posted node status.
      PIDPressure      Unknown   Fri, 22 Apr 2022 15:37:34 +0000   Fri, 22 Apr 2022 15:38:34 +0000   NodeStatusNeverUpdated   Kubelet never posted node status.
  

• Now we get into the node by using SSM or SSH to check kubelet logs, get the kubelet bootstrap command to run directly inside the node

    # /etc/eks/bootstrap.sh dev-d1 --apiserver-endpoint https://111111111111111111111.yl4.ap-southeast-1.eks.amazonaws.com --b64-cluster-ca xxx= --container-runtime containerd --kubelet-extra-args '--node-labels=role=karpenter,type=test,karpenter.sh/provisioner-name=karpenter-test,lifecycle=spot,karpenter.sh/capacity-type=spot --register-with-taints=dedicated=karpenter:NoSchedule'
    sed: can't read /etc/eks/containerd/containerd-config.toml: No such file or directory
    Exited with error on line 469
  

• Ooh, we're missing the configuration file for containerd containerd-config.toml. Check the status of sandbox-image.service and we see it failed to start

    # systemctl status sandbox-image -l
    ● sandbox-image.service - pull sandbox image defined in containerd config.toml
      Loaded: loaded (/etc/systemd/system/sandbox-image.service; enabled; vendor preset: disabled)
      Active: activating (start) since Sat 2022-04-23 08:52:45 UTC; 12min ago
    Main PID: 2965 (bash)
        Tasks: 2
      Memory: 17.0M
      CGroup: /system.slice/sandbox-image.service
              ├─2965 bash /etc/eks/containerd/pull-sandbox-image.sh
              └─3691 sleep 170
  
    Apr 23 09:05:42 ip-10-10-11-162.vc.p pull-sandbox-image.sh[2965]: github.com/containerd/containerd/vendor/github.com/urfave/cli.(*App).RunAsSubcommand(0xc00032c540, 0xc0000ecf20, 0x0, 0x0)
    Apr 23 09:05:42 ip-10-10-11-162.vc.p pull-sandbox-image.sh[2965]: /builddir/build/BUILD/containerd-1.4.13-2.amzn2.0.1/src/github.com/containerd/containerd/vendor/github.com/urfave/cli/app.go:404 +0x8f4
    Apr 23 09:05:42 ip-10-10-11-162.vc.p pull-sandbox-image.sh[2965]: main.main()
    Apr 23 09:05:42 ip-10-10-11-162.vc.p pull-sandbox-image.sh[2965]: github.com/containerd/containerd/cmd/ctr/main.go:37 +0x125
  

• Try to run the script for debuging and we see the script stuck at AWS ECR login.

    # bash -x /etc/eks/containerd/pull-sandbox-image.sh
    ++ awk '-F[ ="]+' '$1 == "sandbox_image" { print $2 }' /etc/containerd/config.toml
    + sandbox_image=123456789012.dkr.ecr.ap-southeast-1.amazonaws.com/eks/pause:3.1-eksbuild.1
    ++ echo 123456789012.dkr.ecr.ap-southeast-1.amazonaws.com/eks/pause:3.1-eksbuild.1
    ++ cut -f4 -d .
    + region=ap-southeast-1
    ++ aws ecr get-login-password --region ap-southeast-1
  

• Now we are near to the root cause.

🚀 Check AWS service endpoints

• We know that our node is getting stuck at AWS ECR login. For security, the Amazon ECR which holds the production repository are often put in the private network so we need to verify that our worker node can access API endpoints for Amazon ECR (and even Amazon EC2 and S3 if we don't pubic the service endpoints)

• Check the service endpoint (go to VPC -> Endpoints) to allow traffics from our worker node to the VPC endpoint.

Conclusion

• There are many usecase for the issue of worker nodes failing to join the EKS/kubernetes cluster so this blog just guides you the way and some clues to troubleshoot.

References:

• How do I resolve kubelet or CNI plugin issues for Amazon EKS?