The Auto-Scaling Death Spiral

PagerDuty — cost anomaly on checkout-service

FinOps: Pods scaled 5 → 250 while traffic stayed flat. Estimated burn: $40,000/month.

The trap: HPA targets 70% CPU, but pods cold-start at 95% for 30s. Each new pod triggers more scale-ups → death spiral.

Your task

Fix `hpa.yaml`: add spec.behavior.scaleUp.stabilizationWindowSeconds (≥ 30s, 60s recommended) so the HPA ignores transient cold-start spikes.

Book: scale at 70% CPU · Prod: CPU is a time-series — cold starts are not sustained load.