Kubernetes cost optimization: what 6 months of automated rightsizing taught us

We've been running automated Kubernetes rightsizing across 40 clusters for 6 months. The findings:\n\n**Average CPU request overprovisioning: 340%.** Teams request CPU based on 'what if' scenarios, not observed usage. The actual p99 CPU usage is typically 25-30% of the request.\n\n**Memory is trickier.** JVM applications have complex memory behavior. Rightsizing memory too aggressively causes OOM kills. We're conservative with memory: only downsize if p99 usage over 30 days is below 50% of request.\n\n**Don't trust load tests.** Load test profiles rarely match production traffic patterns. We use 30-day production metrics, not load test results.\n\n**Savings.** After 6 months: 38% reduction in compute spend across the clusters we manage. That's real money.\n\nThe agent's job: watch usage, calculate optimal requests, propose changes, and apply them during low-traffic windows with automatic rollback on OOM.