gke-cost-optimization

Start by enabling cost allocation to find waste, then rightsize pod requests using VPA recommendations, move fault-tolerant workloads to Spot VMs, and purchase committed use discounts sized to your steady-state baseline. Scale idle dev cluster node pools to zero.

Add nodeSelector with cloud.google.com/gke-spot set to true on your Deployment. Spot VMs are preemptible with about 30 seconds notice, so run at least 2 replicas, set terminationGracePeriodSeconds to 25, and add a preStop hook for graceful shutdown.

Batch processing, dev/test environments, stateless web services with 2+ replicas, and jobs with checkpointing work well on Spot. Stateful workloads like databases and single-replica critical services should not use Spot due to preemption risk.

Resource-based CUDs give deeper discounts (~55% for 3-year) but lock you to a machine family and region. Flexible CUDs offer lower discounts (~46% for 3-year) with portability. Commit only to your measured baseline usage, never peak.

Over-provisioned CPU and memory requests are the most common cause. Deploy VPA in recommendation mode (updateMode Off), wait 24 hours, then compare requests against P95 actual usage and reduce requests exceeding 2-5x actual consumption.