GC-Induced HPA Thrash Cycle
📍 Used after slide 8, 15
Loading diagram…
The GC pause → CPU spike → HPA false scale-out feedback loop.
Speaker Notes
Walk-through: Start at the green “JVM Pod” box. A GC pause creates a CPU spike that HPA interprets as load. HPA scales out, but the new pods also GC on startup, creating more CPU spikes. The cycle repeats until you have 20 pods for a workload that needs 3.
Fix: Scale on RPS (requests per second), not CPU. Add stabilizationWindowSeconds: 120 to absorb GC spikes. Set minReplicas: 2 so a single pod GC pause never means 100% downtime.