Use single Cluster Autoscaler across AZs

Run one Cluster Autoscaler deployment that manages node groups across all Availability Zones — this gives the autoscaler full visibility into cluster-wide capacity and prevents AZ-isolated scaling decisions that lead to imbalanced, over-provisioned clusters.

For the deployment manifest, see the single-ca-multi-az.yaml and verify-ca-az-distribution.sh to verify node distribution across AZs.

Actions

Consolidate multiple per-AZ Cluster Autoscaler deployments into a single instance with --balance-similar-node-groups=true. This prevents one AZ from scaling up while another has idle capacity, eliminating cross-AZ over-provisioning.

Cost impact

Key takeaway: Multiple autoscaler instances per AZ can't see each other's capacity. A single CA with cross-AZ visibility makes globally optimal scaling decisions and avoids redundant nodes sitting idle in one zone while another scales up.