This experiment simulates a simple autoscaler responding to a spiky 10-minute load pattern. Compare two policies:
- Thrashy – scales replicas to the instantaneous desired count with no smoothing or cooldown.
- Stable – uses a smoothed utilization signal and a cooldown window before adjusting replica counts.
Use the toggle below to see how replica thrashing affects CPU, memory, and overall capacity tracking.
Aggressive Autoscaling
Simulating autoscaling behavior…
What to watch
- How often replica count changes under each policy.
- CPU and memory volatility.
- Whether capacity closely tracks load.
Autoscaling is a control system. Thrashy policies look responsive but oscillate badly. A few guardrails turn chaos into something the on-call engineer can actually predict.