Cluster Autoscaler vs Karpenter: Node-Level Scaling in Kubernetes

HPA scales pods, but pods need nodes to run on. When HPA requests 30 replicas and the cluster has room for 3, 27 pods stay Pending. Node-level autoscaling is the other half of the elasticity story.

Cluster Autoscaler

• Open source, works on all clouds, operates on pre-defined node groups (ASGs on AWS, MIGs on GCP).
• Watches Pending pods; when a pod cannot schedule on any existing node, CA picks a matching node group and tells the cloud to scale that ASG up.
• Scale-up takes 3 to 5 minutes — EC2 boot + AMI + kubelet join; Pending pods back up during the wait.
• Scale-down drains under-utilized nodes (below 50% requests for 10 minutes by default) respecting PodDisruptionBudgets.

Karpenter

• Open source, originally AWS, now Azure too; replaces Cluster Autoscaler for those clouds.
• Node-group-less — looks at the exact resource shape of pending pods and queries the EC2 catalog for the cheapest instance type that fits; provisions via EC2 Fleet API.
• Scale-up is 30 to 60 seconds vs CA's 3 to 5 minutes — 5 to 10 times faster because there is no ASG indirection.
• Consolidation continuously re-evaluates whether the fleet can be replaced with fewer or cheaper nodes; drains and replaces respecting PDBs.
• First-class Spot support with mixed Spot/on-demand in one NodePool; subscribes to EC2 interruption events so it can drain and provision replacement before AWS reclaims the instance.

Cost impact

• On variable workloads, Karpenter typically runs 20 to 50 percent cheaper than Cluster Autoscaler because it picks right-sized instances and leverages Spot aggressively.
• For stable workloads the savings are smaller but still positive due to continuous consolidation removing idle capacity.

How they compose with HPA and KEDA

• HPA or KEDA scales the Deployment replica count — replicas become Pods — Pods Pending — the cluster autoscaler adds a node.
• With KEDA scaling from zero and Karpenter provisioning nodes in under a minute, end-to-end wake-up from zero to serving a burst is measurable in one to two minutes.

When to pick each

• Karpenter on AWS or Azure with variable workloads and cost optimization pressure — the default recommendation now for new EKS clusters.
• Cluster Autoscaler for multi-cloud, on-prem, or stable workloads where the operational simplicity of node groups is valuable.
• Both require PDBs and working graceful shutdown on apps — consolidation drains nodes; apps that don't handle SIGTERM drop requests.

Gotchas

• PodDisruptionBudgets must be set correctly — without them consolidation and scale-down cause availability hits.
• Pod density limits per instance type (AWS VPC CNI IP limits) — small instances only run 10 pods; autoscaler bin-packing must respect this.
• Karpenter's consolidation can be aggressive — consolidateAfter: 30s is fine for cost optimization but moves pods frequently; raise it if churn is a problem.