In this post, we're going to explore Karpenter, the ultimate solution for Node Autoscaling. Karpenter provides a cost-effective capability to implement your Kong Konnect Data Plane layer using the best EC2 Instances Types options available for your Kubernetes Nodes.
See the previous posts in this series for more on Data Plane Elasticity and Pod Autoscaling with VPA, HPA, and Node Autoscaling with Cluster Autoscaler on Amazon EKS 1.29.
We can summarize Karpenter as a Kubernetes cluster autoscaler that right-sizes compute resources based on the specific requirements of Cluster workloads. In other words, Karpenter evaluates the aggregate resource requirements of the pending pods and chooses the optimal instance type to run them. That improves the efficiency and cost of running workloads.
The Karpenter AWS Provider GitHub repo highlights the main Karpenter capabilities. Karpenter improves the efficiency and cost of running workloads on Kubernetes clusters by:
Please, check the EKS Best Practices Guide for Karpenter provided by AWS to learn more about it.
The following Karpenter Architecture diagram was taken from the AWS Karpenter Introduction blog post.
Our Karpenter deployment is based on the instructions available in its official site. To make it simpler we are going to recreate the Cluster altogether. First, delete the existing one with:
eksctl delete cluster --name kong35-eks129-autoscaling --region us-west-1Get the following environment variables set:
export KARPENTER_NAMESPACE=kube-system
export KARPENTER_VERSION=v0.33.1
export K8S_VERSION=1.29
export AWS_PARTITION="aws"
export CLUSTER_NAME="kong35-eks129-autoscaling"
export AWS_DEFAULT_REGION="us-west-1"
export AWS_ACCOUNT_ID="$(aws sts get-caller-identity --query Account --output text)"Karpenter leverages several AWS technologies to run, including Amazon EventBridge, Amazon Simple Queue Service (SQS) and IAM Roles. All these fundamental components are created with the following CloudFormation template.
curl -fsSL -o cloudformation.yaml https://raw.githubusercontent.com/aws/karpenter-provider-aws/"${KARPENTER_VERSION}"/website/content/en/preview/getting-started/getting-started-with-karpenter/cloudformation.yaml
aws cloudformation deploy \
--stack-name "Karpenter-${CLUSTER_NAME}" \
--template-file ./cloudformation.yaml \
--capabilities CAPABILITY_NAMED_IAM \
--parameter-overrides "ClusterName=${CLUSTER_NAME}"You can check the AWS resources created by the template with:
aws cloudformation describe-stack-resources --stack-name Karpenter-kong35-eks129-autoscaling --region us-west-1 \
--query "StackResources[*].[LogicalResourceId, PhysicalResourceId]" \
--output tableAfter submitting the CloudFormation template, create the actual EKS Cluster with eksctl. Some comments regarding the declaration:
Differently from Cluster Autoscaler, Karpenter uses the new EKS Pod Identities mechanism to access the required AWS Services.
The iam section uses the podIdentityAssociations parameters to describe how Karpenter uses EKS Pod Identities to manage EC2 instances.
The iamIdentityMappings section manages the aws-auth ConfigMap to grant permission to the KarpenterNodeRole-kong35-eks129-autoscaling Role, created by the CloudFormation template, to access the Cluster.
We are deploying Karpenter in the kong NodeGroup again. The NodeGroup will run on a t3.large EC2 Instance.
addons section asks eksctl to install the Pod Identity Agent.eksctl create cluster -f - <<EOF
apiVersion: eksctl.io/v1alpha5
kind: ClusterConfig
metadata:
name: ${CLUSTER_NAME}
region: ${AWS_DEFAULT_REGION}
version: "${K8S_VERSION}"
tags:
karpenter.sh/discovery: ${CLUSTER_NAME}
iam:
withOIDC: true
podIdentityAssociations:
- namespace: "${KARPENTER_NAMESPACE}"
serviceAccountName: karpenter
roleName: ${CLUSTER_NAME}-karpenter
permissionPolicyARNs:
- arn:${AWS_PARTITION}:iam::${AWS_ACCOUNT_ID}:policy/KarpenterControllerPolicy-${CLUSTER_NAME}
iamIdentityMappings:
- arn: "arn:${AWS_PARTITION}:iam::${AWS_ACCOUNT_ID}:role/KarpenterNodeRole-${CLUSTER_NAME}"
username: system:node:{{EC2PrivateDNSName}}
groups:
- system:bootstrappers
- system:nodes
managedNodeGroups:
- instanceType: t3.large
amiFamily: AmazonLinux2
name: nodegroup-kong
labels: { nodegroupname: kong }
desiredCapacity: 1
minSize: 1
maxSize: 10
addons:
- name: eks-pod-identity-agent
EOFCheck the main environment variables:
export CLUSTER_ENDPOINT="$(aws eks describe-cluster --name ${CLUSTER_NAME} --query "cluster.endpoint" --output text)"
export KARPENTER_IAM_ROLE_ARN="arn:${AWS_PARTITION}:iam::${AWS_ACCOUNT_ID}:role/${CLUSTER_NAME}-karpenter"
echo $CLUSTER_ENDPOINT $KARPENTER_IAM_ROLE_ARNNow, we are ready to install Karpenter. By default, Karpenter requests 2 replicas to run itself. For our simple exploration environment, we are changing that to 1.
helm upgrade --install karpenter oci://public.ecr.aws/karpenter/karpenter --version "${KARPENTER_VERSION}" --namespace "${KARPENTER_NAMESPACE}" --create-namespace \
--set "settings.clusterName=${CLUSTER_NAME}" \
--set "settings.interruptionQueue=${CLUSTER_NAME}" \
--set controller.resources.requests.cpu=1 \
--set controller.resources.requests.memory=1Gi \
--set controller.resources.limits.cpu=1 \
--set controller.resources.limits.memory=1Gi \
--set replicas=1 \
--waitYou can check Karpenter Pod's Log with:
kubectl logs -f -l app.kubernetes.io/name=karpenter -n kube-systemWith Karpenter installed we need to manage two constructs:
NodePool: it's responsible to set constraints to the Nodes Karpenter is going to create. You can specify Taints, limit Node creation to certain zones, Instances Types, and Computer Architectures like AMD and ARM.
EC2NodeClass: specific AWS settings for EC2 Instances. Each NodePool must reference an EC2NodeClass using spec.template.spec.nodeClassRef setting.
Let's create both NodePool and EC2NodeClass based on the basic instructions provided via the Karpenter website.
Note we've added the nodegroupname=kong label to it. This is important to make sure the new Nodes will be available for the Konnect Data Plane Deployment. Moreover, the nodeClassRef setting refers to the default NodeClass we create next. Please, check the Karpenter documentation to learn more about NodePool configuration.
cat <<EOF | envsubst | kubectl apply -f -
apiVersion: karpenter.sh/v1beta1
kind: NodePool
metadata:
name: default
spec:
template:
metadata:
labels:
nodegroupname: kong
spec:
requirements:
- key: kubernetes.io/arch
operator: In
values: ["amd64"]
- key: kubernetes.io/os
operator: In
values: ["linux"]
- key: karpenter.sh/capacity-type
operator: In
values: ["on-demand"]
- key: karpenter.k8s.aws/instance-category
operator: In
values: ["c", "m", "r"]
- key: karpenter.k8s.aws/instance-generation
operator: Gt
values: ["2"]
nodeClassRef:
name: default
limits:
cpu: 1000
disruption:
consolidationPolicy: WhenUnderutilized
expireAfter: 720h # 30 * 24h = 720h
EOFThe EC2NodeClass declaration includes specific AWS settings to be used when creating a new Node such as AMI Family, Instance Profile, Subnets, Security Groups, IAM Role, etc. Note we are grating the KarpenterNodeRole-kong35-eks129-autoscaling Role, created by the CloudFormation template, to the new Nodes.
cat <<EOF | envsubst | kubectl apply -f -
apiVersion: karpenter.k8s.aws/v1beta1
kind: EC2NodeClass
metadata:
name: default
spec:
amiFamily: AL2 # Amazon Linux 2
role: "KarpenterNodeRole-${CLUSTER_NAME}"
subnetSelectorTerms:
- tags:
karpenter.sh/discovery: "${CLUSTER_NAME}"
securityGroupSelectorTerms:
- tags:
karpenter.sh/discovery: "${CLUSTER_NAME}"
EOFAs we have Karpenter installed and configured, let's move on and install the Konnect Data Plane. Make sure you use the same declaration we used before and set the same CPU and memory (cpu=1500m, memory=3Gi) resources to it.
Since we are going to use HPA and Karpenter together, install the Metrics Server on your Cluster along with the HPA policy allowing 20 replicas to be created.
Finally, create the new Node for the Upstream and Load Generator as well as deploy the Upstream Service using the same declaration.
Start the same Fortio 60-minute-long load test with 5000 qps.
After some minutes we'll see both HPA and Karpenter in action. Here's one of the HPA results I got:
% kubectl get hpa -n kong
NAME REFERENCE TARGETS MINPODS MAXPODS REPLICAS AGE
kong-hpa Deployment/kong-kong 37%/75% 1 20 15 21hAnd here's the new Nodes Karpenter created:
% kubectl get nodes -o json | jq -r '.items[].metadata.labels | select(.nodegroupname=="kong") | ."node.kubernetes.io/instance-type"'
m3.medium
t3.large
m3.medium
c5a.xlarge
c5a.large
% kubectl top node --selector='karpenter.sh/nodepool=default'
NAME CPU(cores) CPU% MEMORY(bytes) MEMORY%
ip-192-168-110-150.us-west-1.compute.internal 194m 20% 1270Mi 39%
ip-192-168-71-115.us-west-1.compute.internal 696m 74% 386Mi 11%
ip-192-168-82-234.us-west-1.compute.internal 1281m 32% 5742Mi 83%
ip-192-168-91-24.us-west-1.compute.internal 839m 43% 3175Mi 99% One of the most powerful Karpenter capabilities is Cluster Consolidation, that is, the ability to delete or replace Nodes to a cheaper configuration.
You can see it in action if you leave the test load running a little longer. We'll see that Karpenter has consolidated the multiple Nodes into a single one:
% kubectl get hpa -n kong
NAME REFERENCE TARGETS MINPODS MAXPODS REPLICAS AGE
kong-hpa Deployment/kong-kong 32%/75% 1 20 15 21h
% kubectl get nodes -o json | jq -r '.items[].metadata.labels | select(.nodegroupname=="kong") | ."node.kubernetes.io/instance-type"'
t3.large
c5a.2xlarge
% kubectl top node --selector='karpenter.sh/nodepool=default'
NAME CPU(cores) CPU% MEMORY(bytes) MEMORY%
ip-192-168-75-198.us-west-1.compute.internal 2056m 25% 4241Mi 28% From the API consumption perspective, here are the results I got. As you can see the Data Plane layer with all its replicas was able to honor the QPS requested with expected latency time.
The P99 latency: for example, # target 99% 0.0484703
The number of requests sent along with the QPS: All done 18000000 calls (plus 800 warmup) 98.065 ms avg, 4999.8 qps
As a fundamental principle of Elasticity, if we stop the load test, deleting the Fortio Pod, we should see HPA and Karpenter reducing the resources allocated to the Data Plane.
kubectl delete pod fortio
% kubectl get hpa -n kong
NAME REFERENCE TARGETS MINPODS MAXPODS REPLICAS AGE
kong-hpa Deployment/kong-kong 1%/75% 1 20 1 22h
% kubectl get nodes -o json | jq -r '.items[].metadata.labels | select(.nodegroupname=="kong") | ."node.kubernetes.io/instance-type"'
t3.largeKong takes performance and elasticity very seriously. When we come to a Kubernetes deployment, it's important to support all Elasticity technologies available to provide our customers flexibility and a lightweight and performant API gateway infrastructure.
This blog post series described Kong Konnect Data Plane deployment to take advantage of the main Kubernetes-based Autoscaling technologies:
Kong Konnect simplifies API management and improves security for all services infrastructure. Try it for free today!
In the rapidly evolving landscape of API management, understanding the raw performance and reliability of your API gateway is not just an expectation — it's a necessity. At Kong, we're dedicated to ensuring our users have access to concrete, action
After getting our Konnect Data Planes vertically and horizontally scaled, with VPA and HPA , it's time to explore the Kubernete Node Autoscaler options. In this post, we start with the Cluster Autoscaler mechanism. (Part 4 in this series is dedic
In this series of posts, we will look closely at how Kong Konnect Data Planes can take advantage of Autoscalers running on Amazon Elastic Kubernetes Services (EKS) 1.29 to support the throughput the demands API consumers impose on it at the lowest c
In one of our posts, Kong AI/MCP Gateway and Kong MCP Server technical breakdown, we described the new capabilities added to Kong AI Gateway to support MCP (Model Context Protocol). The post focused exclusively on consuming MCP server and MCP tool
Managed Redis cache is a turnkey "Shared State" add-on for Kong Dedicated Cloud Gateways. It is designed to combine the performance of an in-memory data store with the simplicity of a SaaS product. When you spin up a Dedicated Cloud Gateway in Kong
Key Takeaways API testing is crucial for ensuring the reliability, security, and performance of modern applications. Different types of testing, such as functional, security, performance, and integration testing, should be employed to cover all aspe
So, what exactly is Kong Insomnia? Kong Insomnia is your all-in-one platform for designing, testing, debugging, and shipping APIs at speed. Built for developers who need power without bloat, Insomnia helps you move fast whether you’re working solo,