Kubernetes Autoscaling with Karpenter: Cut Your AWS Bill in Half

Why Cluster Autoscaler isn't enough

Cluster Autoscaler (CAS) was designed for a world where Kubernetes node groups were fixed. It scales node groups up and down, but it's slow (1–2 minutes to provision a new node) and makes poor decisions about which instance type to use.

Karpenter changes this fundamentally. Instead of managing node groups, Karpenter provisions individual nodes in response to pending pods — often in under 30 seconds, and always with the right instance type for the workload.

What Karpenter does differently

1. Direct EC2 provisioning — Karpenter calls the EC2 API directly, bypassing Auto Scaling Groups for new nodes
2. Just-in-time sizing — Analyzes pending pod requirements and picks the optimal instance type
3. Consolidation — Actively moves pods to fewer, fuller nodes and terminates the empty ones
4. Spot-aware — Understands Spot interruptions and handles them gracefully

Installing Karpenter on EKS

# Terraform: Karpenter IAM resources
module "karpenter" {
  source  = "terraform-aws-modules/eks/aws//modules/karpenter"
  version = "~> 20.0"

  cluster_name = module.eks.cluster_name

  node_iam_role_additional_policies = {
    AmazonSSMManagedInstanceCore = "arn:aws:iam::aws:policy/AmazonSSMManagedInstanceCore"
  }
}

helm install karpenter oci://public.ecr.aws/karpenter/karpenter \
  --version "1.0.0" \
  --namespace kube-system \
  --set settings.clusterName=my-cluster \
  --set settings.interruptionQueue=my-cluster-karpenter \
  --set controller.resources.requests.cpu=1 \
  --set controller.resources.requests.memory=1Gi

NodePool configuration

NodePool replaces the old Provisioner CRD:

apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: default
spec:
  template:
    spec:
      requirements:
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["spot", "on-demand"]
        - key: node.kubernetes.io/instance-category
          operator: In
          values: ["c", "m", "r"]
        - key: kubernetes.io/arch
          operator: In
          values: ["amd64", "arm64"]
        - key: karpenter.k8s.aws/instance-generation
          operator: Gt
          values: ["2"]
      nodeClassRef:
        group: karpenter.k8s.aws
        kind: EC2NodeClass
        name: default
  limits:
    cpu: 1000
    memory: 4000Gi
  disruption:
    consolidationPolicy: WhenEmptyOrUnderutilized
    consolidateAfter: 1m

Cost analysis: Cluster Autoscaler vs Karpenter

A real-world migration we performed for a fintech client:

The key difference: Karpenter automatically picks Spot instances when they're available and cheaper, falling back to on-demand only when necessary. CAS requires manual node group configuration for each instance type.

Handling Spot interruptions

Karpenter handles Spot interruptions via the SQS interruption queue:

spec:
  disruption:
    # Karpenter will gracefully terminate Spot nodes 2 minutes
    # before the interruption, giving pods time to drain
    budgets:
      - nodes: "10%"  # Max 10% of nodes disrupted at once

For stateless workloads, this is transparent. For stateful workloads, combine with Pod Disruption Budgets.

Key takeaways

• Karpenter is faster, smarter, and cheaper than Cluster Autoscaler for most teams
• Expect 30–50% cost reduction from better instance selection and Spot adoption
• Consolidation actively reduces waste by packing workloads onto fewer nodes
• Migration from CAS is straightforward: install Karpenter, create NodePools, remove CAS