View a markdown version of this page

Cluster Autoscaler optimization - AWS Prescriptive Guidance
Verify autoscaler behaviorActions

Cluster Autoscaler optimization

For clusters still running Cluster Autoscaler, the default configuration prioritizes safety over cost. Tuning these parameters controls how quickly underutilized nodes are reclaimed and how efficiently new capacity is provisioned.

Cost-Optimized Cluster Autoscaler Configuration

The key tuning changes vs defaults:

  • scale-down-delay-after-add: 5m (default 10m): faster reclaim after scale-up

  • scale-down-unneeded-time: 5m (default 10m): detect idle nodes sooner

  • scale-down-utilization-threshold: 0.5: scale down if node is <50% utilized

  • expander: least-waste: choose node group with best bin-packing fit

  • balance-similar-node-groups: true: spread across AZs for Spot availability

For the complete deployment manifest,see the cluster-autoscaler-cost-optimized.yaml

Verify autoscaler behavior

Confirm that scale-down is actually happening, misconfigured flags or pod disruption budgets can silently block node removal. The verification script checks:

  • Autoscaler status ConfigMap for current decisions

  • Nodes with scale-down-disabled label (blocked from removal)

  • Pending pods that may be triggering unnecessary scale-up

  • Node utilization levels (nodes below 50% should be consolidated)

  • CloudWatch node count trend (flat count despite variable traffic = paying for unused capacity)

For the verification commands, see the verify-autoscaler-behavior.sh

For the complete scripts and manifests, see the 04-karpenter-cost-optimization folder in the code repository.

Actions

Set scale-down-utilization-threshold to 0.5, scale-down-unneeded-time to 5m, and use the least-waste expander. Monitor node count trends weekly, if it's flat while pod count fluctuates, your scale-down settings are too conservative.

Key takeaway: The default Cluster Autoscaler settings are designed for safety, not cost. Tuning scale-down aggressiveness and using least-waste bin-packing can reduce node count by 20–40% without impacting workload availability.