Abstract

• Karpenter - Just-in-time Nodes for Any Kubernetes Cluster. It is one of Node-based autoscaling (adding or removing nodes as needed) beyond cluster autoscaler.

• What does this post provide?

  • Hands-on installing karpenter on EKS Cluster

  • Provide permission for Karpenter to create AWS resources through IAM for service account (IRSA)

  • Create a sample karpenter provisioner to test scaling-out and scaling down nodes by karpenter

  • Finally what Karpenter improvements vs cluster autoscaler.

Table Of Contents

• Pre-requisite

• Create karpenter service account

• Install karpenter using helm chart

• Karpenter provisioner

• Test karpenter scaleout nodes

• Karpenter Deprovisioning

• Troubleshooting

• Karpenter improvements

• Conclusion

🚀 Pre-requisite

• EKS cluster

• IAM Instance profile which will be assigned to EKS nodes and help them to join the EKS cluster

• OIDC provider

🚀 Create karpenter service account

• The best practice to provide AWS permission for Kubernetes service is Using IAM Service Account Instead Of Instance Profile For EKS Pods

• If you already set up OIDC by using IAM identity provider then you can create the IAM role as a service account for karpenter manually or using CDK. The role needs permission on EC2 actions only kapenter-sa.ts

• Then generate service-account yaml based on the output IAM role ARN

• karpenter-sa.yaml

• Create service account by kubectl apply -f karpenter-sa.yaml and then check the result

    ⚡ $ kf get sa -n karpenter
    NAME                   SECRETS   AGE
    karpenter-controller   1         17m
  

🚀 Install karpenter using helm chart

• Use karpenter-values.yaml to disable creating new serviceAccount and point to the one which is created above. Replace clusterName and clusterEndpoint with your EKS cluster

• Installing the Chart

    helm repo add karpenter https://charts.karpenter.sh
    helm repo update
    helm upgrade --install karpenter karpenter/karpenter --namespace karpenter \
      --version 0.8.1 \
      --values yaml/karpenter-values.yaml \
      --wait
  

• Check the karpenter pod created, it is included controller and webhook containers. Describe the pod to ensure it is assigned correct serviceAccount with the proper IAM role

⚡ $ kf get pod -n karpenter
NAME                         READY   STATUS    RESTARTS   AGE
karpenter-67f957c8c4-t752q   2/2     Running   0          2m56s

• Note that Karpenter provides autoscalling nodes for our services but it still needs a node to deploy itself, we run the Karpenter controller on EKS Fargate or on a worker node that belongs to a node group

🚀 Karpenter provisioner

• Provisioner - Provides options for karpenter controller create expected nodes such as instance profile, AMI family such as Bottlerocket, instance type, security group, subnet, tags, capacity type such as spot, etc.

• Sample provisioner:

  • instanceProfile: As described above, this profile just needs enough permissions to join the EKS cluster, eg. eks-node-role-dev with permission 'AmazonEKS_CNI_Policy', 'AmazonEKSWorkerNodePolicy', 'AmazonEC2ContainerRegistryReadOnly', 'AmazonSSMManagedInstanceCore'

  • Use Bottlerocket as AMI family

  • subnetSelector: Specify the tag of the subnet we want to host the nodes, we should use the private subnet of the EKS VPC so ensure the private subnets have a tag EksCluster/EKSVpc/PrivateSubnet*

  • securityGroupSelector: Specify the tag of the security group we want to attach to the nodes

  • Node requirements: spot instances, small size as we just test

  • taints: Taint the nodes if we want to control our services assigned to expected nodes and separate resources

  • dev-provisioner.yaml

• Apply the yaml file by kf apply -f dev-provisioner.yaml and then check the result

    ⚡ $ kf get provisioners.karpenter.sh
    NAME   AGE
    dev    1m
  

🚀 Test karpenter scaleout nodes

• Test the karpenter provisioner we just created above by applying the deployment. First check that there's no provisioned node created yet

    ⚡ $ kf get node -l karpenter.sh/provisioner-name=dev
    No resources found
  

• Apply the deployment test-deployment.yaml, note that the deployment must have tolerations map with the taints if we specify in the provisioner.

    ⚡ $ kf apply -f yaml/test-deployment.yaml
    deployment.apps/inflate created
  
    ⚡ $ kf get pod
    NAME                      READY   STATUS    RESTARTS   AGE
    inflate-5db4558c8-rtd8g   1/1     Running   0          85s
    inflate-5db4558c8-x8sbn   1/1     Running   0          85s
  
    $ kf get node -l karpenter.sh/provisioner-name=dev
    NAME                                            STATUS   ROLES    AGE   VERSION
    ip-10-0-147-5.ap-northeast-2.compute.internal   Ready    <none>   77s   v1.22.6-eks-b18cdc9
  

• We can check the log of karpenter controller to see how it works

    2022-04-10T10:20:56.558Z        INFO    controller.provisioning Batched 2 pods in 1.029435847s  {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:56.800Z        DEBUG   controller.provisioning Discovered subnets: [subnet-048931802b9fe4d68 (ap-northeast-2a) subnet-03112351e0b6460ab (ap-northeast-2c) subnet-089fe79a0c5c5247d (ap-northeast-2b)] {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:56.903Z        INFO    controller.provisioning Computed packing of 1 node(s) for 2 pod(s) with instance type option(s) [t3.xlarge]     {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:57.057Z        DEBUG   controller.provisioning Discovered security groups: [sg-074f63f39cbb17d7f sg-0de828621f69e459f] {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:57.059Z        DEBUG   controller.provisioning Discovered kubernetes version 1.22      {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:57.155Z        DEBUG   controller.provisioning Discovered ami-09fd6fd7a1729123a for query /aws/service/bottlerocket/aws-k8s-1.22/x86_64/latest/image_id        {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:57.327Z        DEBUG   controller.provisioning Created launch template, Karpenter-eks-dev-1642383919802956422   {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:58.997Z        DEBUG   controller.provisioning Discovered 257 EC2 instance types       {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:59.083Z        DEBUG   controller.provisioning Discovered EC2 instance types zonal offerings   {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:59.495Z        INFO    controller.provisioning Launched instance: i-0c33d36e900713716, hostname: ip-10-0-147-5.ap-northeast-2.compute.internal, type: t3.xlarge, zone: ap-northeast-2b, capacityType: spot    {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:59.518Z        INFO    controller.provisioning Bound 2 pod(s) to node ip-10-0-147-5.ap-northeast-2.compute.internal    {"commit": "78d3031", "provisioner": "dev"}
    2022-04-10T10:20:59.518Z        INFO    controller.provisioning Waiting for unschedulable pods  {"commit": "78d3031", "provisioner": "dev"}
  

• Go to AWS console to check the node

  

  

🚀 Karpenter Deprovisioning

• How Karpenter nodes are deprovisioned

  • Node empty: Node with no pods (non-daemonset) will be deleted after ttlSecondsAfterEmpty

  • Node expired: based on ttlSecondsUntilExpired to delete node no matter any workload. This option is to handle node upgrades with the latest AMI. We can protect the application’s availability by using Disruption budgets

  • Disruption budgets - PodDisruptionBudget

  • By adding a karpenter.sh/do-not-evict annotation to the pod, you are instructing Karpenter to preserve the node until the Pod is terminated or the do-not-evict annotation is removed

• In our provisioner, we specify ttlSecondsAfterEmpty: 30 so let's delete the test-deployment.yaml and check karpenter scale-down the node

    ⚡ $ kf delete -f yaml/test-deployment.yaml
    deployment.apps "inflate" deleted
    ⚡ $ kf get pod
    No resources found in default namespace.
  
    ⚡ $ kf get node -l karpenter.sh/provisioner-name=dev
    No resources found
  

• See log for more detail

    2022-04-10T10:33:31.785Z        INFO    controller.node Added TTL to empty node {"commit": "78d3031", "node": "ip-10-0-147-5.ap-northeast-2.compute.internal"}
    2022-04-10T10:34:01.000Z        INFO    controller.node Triggering termination after 30s for empty node {"commit": "78d3031", "node": "ip-10-0-147-5.ap-northeast-2.compute.internal"}
    2022-04-10T10:34:01.031Z        INFO    controller.termination  Cordoned node   {"commit": "78d3031", "node": "ip-10-0-147-5.ap-northeast-2.compute.internal"}
    2022-04-10T10:34:01.289Z        INFO    controller.termination  Deleted node    {"commit": "78d3031", "node": "ip-10-0-147-5.ap-northeast-2.compute.internal"}
  

🚀 Troubleshooting

• Error we might get and how to troubleshoot

  1. I got the following error due to the wrong setup the ServiceAccount

      2022-03-15T14:59:22.984Z        ERROR   controller.provisioning Could not launch node, launching instances, getting launch template configs, getting launch templates, creating launch template, UnauthorizedOperation: You are not authorized to perform this operation. Encoded authorization failure message: u3BZrZUSziC6V7RjV0uPSiwFLaqJMFHwZIuPvJPp-p7sYQMP0EUtzbldDO7xSZD5XrdTLdFz_zxjzNhOPtnmenOgb-sEDxiMuClKprCuB4Gl2QxOJz97UYEFx5FB1Mks0qC-aoPrnMnDzKjdMlNdBsnj81UuFn9tnYBo3pIZG7A_Mwk7g6VotuOLDZl3sjZi15dyWQ6Roe1Htf_8uT2Wzw9n2WTF815us6SEA_3yh71xu4dEuPoT2CECj-RdbSgpcH6ZUi0UKDK3SRNvwLlibg8-dmvQriGC5eXRNRhQuMridNGIV4XmH1j8BN887nmgl023iZQ7XzV3Ge_5NY-MI0-Wsyk1gL57tm87n3CM23vGhTe1GIp2LvZMolVdE4IOgDn00DydMfI9VxD1KqJLNFuxwhzT0pv7JUYKR_i-H8ibmRgNAc-NXwDQBoEqk0OuSiTYBzYDsKyCgrtV4jwPHgOe9sC93GUegNJBrErc0sXSSxSR4pTqYu-Fw8CAPLKFzwph30rChTxlaONWfi-LyDpkpbV0i30soH-_vqJ244k49bs
     

     • The error of authorization status of a request was encoded so we need to decode to get more info on the error by using decode-authorization-message

         ~ $ aws sts decode-authorization-message --encoded-messag u3BZrZUSziC6V7RjV0uPSiwFLaqJMFHwZIuPvJPp-p7sYQMP0EUtzbldDO7xSZD5XrdTLdFz_zxjzNhOPtnmenOgb-sEDxiMuClKprCuB4Gl2QxOJz97UYEFx5FB1Mks0qC-aoPrnMnDzKjdMlNdBsnj81UuFn9tnYBo3pIG7A_Mwk7g6VotuOLDZl3sjZi15dyWQ6Roe1Htf_8uT2Wzw9n2WTF815us6SEA_3yh71xu4dEuPoT2CECj-RdbSgpcH6ZUi0UKDK3SRNvwLlibg8-dmvQriGC5eXRNRhQuMridNGIV4XmH1j8BN887nmgl023iZQ7XzV3Ge_5Y-MI0-Wsyk1gL57tm87n3CM23vGhTe1GIp2LvZMolVdE4IOgDn00DydMfI9VxD1KqJLNFuxwhzT0pv7JUYKR_i-H8ibmRgNAc-NXwDQBoEqk0OuSiTYBzYDsKyCgrtV4jwPHgOe9sC93GUegNJBrErc0sXSSxSR4pTqYu-FwCAPLKFzwph30rChTxlaONWfi-LyDpkpbV0i30soH-_vqJ244k49bs
         {
             "DecodedMessage": "{\"allowed\":false,\"explicitDeny\":false,\"matchedStatements\":{\"items\":[]},\"failures\":{\"items\":[]},\"context\":{\"principal\":{\"id\":\"AROAZUFR7JW2VHL4MT3MS:i-0fa071d72e68e7a33\",\"arn\":\"arn:aws:sts::123456789012:assumed-role/eks-node-role-dev/i-0fa071d72e68e7a33\"},\"action\":\"ec2:CreateLaunchTemplate\",\"resource\":\"arn:aws:ec2:ap-northeast-2:123456789012:launch-template/*\",\"conditions\":{\"items\":[{\"key\":\"aws:Region\",\"values\":{\"items\":[{\"value\":\"ap-northeast-2\"}]}},{\"key\":\"aws:ID\",\"values\":{\"items\":[{\"value\":\"*\"}]}},{\"key\":\"aws:Service\",\"values\":{\"items\":[{\"value\":\"ec2\"}]}},{\"key\":\"aws:Resource\",\"values\":{\"items\":[{\"value\":\"launch-template/*\"}]}},{\"key\":\"aws:Type\",\"values\":{\"items\":[{\"value\":\"launch-template\"}]}},{\"key\":\"aws:Account\",\"values\":{\"items\":[{\"value\":\"123456789012\"}]}},{\"key\":\"aws:ARN\",\"values\":{\"items\":[{\"value\":\"arn:aws:ec2:ap-northeast-2:123456789012:launch-template/*\"}]}}]}}}"
         }
       

  2. Invalid instance profile

     • Error

         2022-03-24T17:16:36.430Z        ERROR   controller.provisioning Could not launch node, launching instances, with fleet error(s), InvalidParameterValue: Value (eks-node-role-dev) for parameter iamInstanceProfile.name is invalid. Invalid IAM Instance Profile name {"commit": "78d3031", "provisioner": "dev"}
       

     • The reason in my case is that the IAM role attached to EC2 was created by CDK and the instance profile named eg. eks-febfdbb4-6ab2-a865-acf8-2884edf78fdc, not the role name.

     • To fix this issue we can create a new instance profile with a role name and add it to the role

         ⚡ $ aws iam create-instance-profile --instance-profile-name  eks-node-role-dev
         ⚡ $ aws iam add-role-to-instance-profile --instance-profile-name eks-node-role-dev --role-name eks-node-role-dev
       

🚀 Karpenter improvements

• Karpenter improvements

  • Designed to handle the full flexibility of the cloud: Karpenter has the ability to efficiently address the full range of instance types available through AWS. Cluster autoscaler was not originally built with the flexibility to handle hundreds of instance types, zones, and purchase options.

  • Group-less node provisioning: Karpenter manages each instance directly, without using additional orchestration mechanisms like node groups. This enables it to retry in milliseconds instead of minutes when capacity is unavailable. It also allows Karpenter to leverage diverse instance types, availability zones, and purchase options without the creation of hundreds of node groups.

  • Scheduling enforcement: Cluster autoscaler doesn’t bind pods to the nodes it creates. Instead, it relies on the kube-scheduler to make the same scheduling decision after the node has come online. A node that Karpenter launches has its pods bound immediately. The kubelet doesn’t have to wait for the scheduler or for the node to become ready. It can start preparing the container runtime immediately, including pre-pulling the image. This can shave seconds off of node startup latency.

🚀 Conclusion

• At the time of writing this post, Karpenter is still on its way to add/improve many features, check more here

• In practice, we can create multiple karpenter provisioners in following usecases

  • When different teams are sharing a cluster and need to run their workloads on different worker nodes with specific request resources or AMI such as Bottlerocket

  • when creating multiple vclusters and would like to separate resources between vclusters.

  • In general, using multiple provisioners makes sure that the most appropriate assets are available to each team

References:

• Introducing Karpenter – An Open-Source High-Performance Kubernetes Cluster Autoscaler

• Configure the Node Termination Handler to use queue processor mode