Karpenter Blueprints for Amazon EKS
You can browse all blueprints on the documentation website: https://aws-samples.github.io/karpenter-blueprints/
Motivation
Karpenter, a node provisioning project built for Kubernetes has been helping many companies to improve the efficiency and cost of running workloads on Kubernetes. However, as Karpenter takes an application-first approach to provision compute capacity for the Kubernetes data plane, there are common workload scenarios that you might be wondering how to configure them properly. This repository includes a list of common workload scenarios, some of them go in depth with the explanation of why configuring Karpenter and Kubernetes objects in such a way is important.
Blueprint Structure
Each blueprint follows the same structure to help you better understand what’s the motivation and the expected results:
| Concept |
Description |
| Purpose |
Explains what the blueprint is about, and what problem is solving. |
| Requirements |
Any pre-requisites you might need to use the blueprint (i.e. An arm64 container image). |
| Deploy |
The steps to follow to deploy the blueprint into an existing Kubernetes cluster. |
| Results |
The expected results when using the blueprint. |
How to use these Blueprints?
Before you get started, you need to have a Kubernetes cluster with Karpenter installed. If you’re planning to work with an existing cluster, just make sure you’ve configured Karpenter following the official guide. This project also has a template to create a cluster with everything you’ll need to test each blueprint.
Support & Feedback
[!IMPORTANT]
Karpenter Blueprints for Amazon EKS is maintained by AWS Solution Architects. It is not part of an AWS
service and support is provided as a best-effort by the Karpenter Blueprints community. To provide feedback,
please use the issues templates
provided. If you are interested in contributing to EKS Blueprints, see the
Contribution guide.
Requirements
*NOTE: If you’re planning to use an existing EKS cluster, you don’t need the optional prerequisites.
Preparing to Deploy Blueprints
Before you start deploying and testing blueprints, make sure you follow next steps. For example, all blueprints assume that you have an EKS cluster with Karpenter deployed, and others even required that you have a default Karpenter NodePool deployed.
If you’re planning on using an existing EKS cluster, you can use an existing node group with On-Demand instances to deploy the Karpenter controller. To do so, you need to follow the Karpenter getting started guide.
You’ll create an Amazon EKS cluster using the EKS Blueprints for Terraform project. The Terraform template included in this repository is going to create a VPC, an EKS control plane, and a Kubernetes service account along with the IAM role and associate them using IAM Roles for Service Accounts (IRSA) to let Karpenter launch instances. Additionally, the template configures the Karpenter node role to the aws-auth configmap to allow nodes to connect, and creates an On-Demand managed node group for the kube-system and karpenter namespaces.
To create the cluster, clone this repository and open the cluster/terraform folder. Then, run the following commands:
cd cluster/terraform
export TF_VAR_region=$AWS_REGION
terraform init
terraform apply -target="module.vpc" -auto-approve
terraform apply --auto-approve
Before you continue, you need to enable your AWS account to launch Spot instances if you haven’t launch any yet. To do so, create the service-linked role for Spot by running the following command:
aws iam create-service-linked-role --aws-service-name spot.amazonaws.com || true
You might see the following error if the role has already been successfully created. You don’t need to worry about this error, you simply had to run the above command to make sure you have the service-linked role to launch Spot instances:
An error occurred (InvalidInput) when calling the CreateServiceLinkedRole operation: Service role name AWSServiceRoleForEC2Spot has been taken in this account, please try a different suffix.
Once complete (after waiting about 15 minutes), run the following command to update the kube.config file to interact with the cluster through kubectl:
aws eks --region $AWS_REGION update-kubeconfig --name karpenter-blueprints
You need to make sure you can interact with the cluster and that the Karpenter pods are running:
$> kubectl get pods -n karpenter
NAME READY STATUS RESTARTS AGE
karpenter-5f97c944df-bm85s 1/1 Running 0 15m
karpenter-5f97c944df-xr9jf 1/1 Running 0 15m
You can now proceed to deploy the default Karpenter NodePool, and deploy any blueprint you want to test.
Deploy a Karpenter Default EC2NodeClass and NodePool
Before you start deploying a blueprint, you need to have a default EC2NodeClass and a default NodePool as some blueprints need them. EC2NodeClass enable configuration of AWS specific settings for EC2 instances launched by Karpenter. The NodePool sets constraints on the nodes that can be created by Karpenter and the pods that can run on those nodes. Each NodePool must reference an EC2NodeClass using spec.nodeClassRef.
If you create a new EKS cluster following the previous steps, a Karpenter EC2NodeClass “default” and a Karpenter NodePool “default” are installed automatically.
NOTE: For existing EKS cluster you have to modify the provided ./cluster/terraform/karpenter.tf according to your setup by properly modifying securityGroupSelectorTerm and subnetSelectorTerms removing the depends_on section. If you’re not using Terraform, you need to get those values manually. CLUSTER_NAME is the name of your EKS cluster (not the ARN). Karpenter auto-generates the instance profile in your EC2NodeClass given the role that you specify in spec.role with the placeholder KARPENTER_NODE_IAM_ROLE_NAME, which is a way to pass a single IAM role to the EC2 instance launched by the Karpenter NodePool. Typically, the instance profile name is the same as the IAM role(not the ARN).
You can see that the NodePool has been deployed by running this:
kubectl get nodepool
You can see that the EC2NodeClass has been deployed by running this:
kubectl get ec2nodeclass
Throughout all the blueprints, you might need to review Karpenter logs, so let’s create an alias for that to read logs by simply running kl:
alias kl="kubectl -n karpenter logs -l app.kubernetes.io/name=karpenter --all-containers=true -f --tail=20"
You can now proceed to deploy any blueprint you want to test.
Once you’re done with testing the blueprints, if you used the Terraform template from this repository, you can proceed to remove all the resources that Terraform created. To do so, run the following commands:
kubectl delete --all nodeclaim
kubectl delete --all nodepool
kubectl delete --all ec2nodeclass
export TF_VAR_region=$AWS_REGION
terraform destroy -target="module.eks" --auto-approve
terraform destroy --auto-approve
Deploying a Blueprint
After you have a cluster up and running with Karpenter installed, you can start testing each blueprint. A blueprint might have a NodePool, EC2NodeClass and a workload example. You need to open the blueprint folder and follow the steps to deploy the resources needed to test the blueprint.
Why EKS Auto Mode content?
EKS Auto Mode uses the same Karpenter scheduling logic as OSS Karpenter but AWS manages the data plane: AMIs, node IAM role, IMDS settings, bootstrap, and patching. That means a few fields disappear from your manifests and a few API groups change. For every Auto Mode-compatible blueprint below, you’ll find an -automode.yaml alongside the original.
The differences boil down to:
NodeClass replaces EC2NodeClass (apiVersion: eks.amazonaws.com/v1)- Managed fields removed:
amiFamily, amiSelectorTerms, role, userData, blockDeviceMappings, metadataOptions
- Instance label prefix:
eks.amazonaws.com/instance-* replaces karpenter.k8s.aws/instance-*
- Max node lifetime: 336h (14 days) —
expireAfter is capped at this value
- Access Entry required: the node IAM role needs
AmazonEKSAutoNodePolicy via an EKS Access Entry. The reference Terraform template under cluster/automode/ wires this up for you.
Workload manifests (Deployments, Jobs, etc.) are identical between OSS and Auto Mode, with these exceptions called out in their own blueprints:
nvidia-gpu-workload uses the eks.amazonaws.com/instance-gpu-name label key instead of karpenter.k8s.aws/instance-gpu-name for GPU pinning, and skips the NVIDIA device plugin install since Auto Mode includes it.neuron-workload skips the neuron-helm-chart install since Auto Mode includes the AWS Neuron device plugin and drivers automatically.soci-snapshotter collapses three NodePools (AL2023 SOCI, Bottlerocket SOCI, non-SOCI baseline) into a single Auto Mode NodePool because Auto Mode runs Bottlerocket only and SOCI is always enabled.static-nodepool keeps the replicas field but drops the OSS-specific networkInterfaces block — Auto Mode manages EFA networking through other mechanisms.
Here’s the list of blueprints we have so far:
NOTE: Each blueprint is independent from each other, so you can deploy and test multiple blueprints at the same time in the same Kubernetes cluster. However, to reduce noise, we recommend you to test one blueprint at a time.
Supported Versions
The following table describes the list of resources along with the versions where the blueprints in this repo have been tested.
License
MIT-0 Licensed. See LICENSE.
评论