Karpenter — Just-in-Time Kubernetes Node Autoscaler for AWS

What it is

Karpenter is an open-source Kubernetes node autoscaler that originated at AWS and is now a CNCF project. It watches for pending pods and launches the right-sized EC2 instance type in approximately 30 seconds, bypassing the traditional Auto Scaling Group approach used by the Cluster Autoscaler.

Karpenter is designed for platform engineers and DevOps teams running EKS or self-managed Kubernetes on AWS who want faster scaling and lower compute costs.

How it saves time or tokens

The traditional Cluster Autoscaler adds nodes by scaling pre-defined Auto Scaling Groups, which forces you to choose instance types in advance. Karpenter selects the optimal instance type at scheduling time based on actual pod requirements. It also consolidates underutilized nodes, moving workloads to fewer, better-utilized instances. Teams report 30-60% compute cost reduction compared to static ASG-based scaling.

How to use

1. Install Karpenter on an existing EKS cluster:

helm upgrade --install karpenter oci://public.ecr.aws/karpenter/karpenter \
  --namespace kube-system --version 1.0.6 \
  --set "settings.clusterName=${CLUSTER}" \
  --set "serviceAccount.annotations.eks\.amazonaws\.com/role-arn=${ROLE_ARN}"

1. Create a NodePool and EC2NodeClass:

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: default
spec:
  template:
    spec:
      requirements:
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: [c, m, r]
        - key: karpenter.sh/capacity-type
          operator: In
          values: [on-demand, spot]
      nodeClassRef:
        group: karpenter.k8s.aws
        kind: EC2NodeClass
        name: default
  disruption:
    consolidationPolicy: WhenEmptyOrUnderutilized

1. Deploy workloads normally; Karpenter handles node provisioning automatically.

Example

When a deployment scales up and pods are pending:

# Scale a deployment
kubectl scale deployment my-app --replicas=50

# Karpenter detects pending pods and selects optimal instances
# Watch nodes appearing:
kubectl get nodes -w
# NAME                          STATUS   ROLES    AGE
# ip-10-0-1-42.ec2.internal     Ready    <none>   32s

Related on TokRepo

• DevOps tools — infrastructure automation and Kubernetes management
• Automation tools — automated scaling and provisioning

Common pitfalls

• Karpenter requires IRSA (IAM Roles for Service Accounts) to be configured; without it, the controller cannot launch EC2 instances
• Setting overly broad instance-category requirements can lead to Karpenter choosing unusual instance types; constrain to c, m, r families for general workloads
• Consolidation can disrupt stateful workloads; use the do-not-disrupt annotation on pods that should not be moved