Horizontal pod autoscaling

Scale on the thing that hurts

CPU is the default scaling signal and often the wrong one. A service that is slow because it waits on a database shows low CPU while queueing badly — scaling on queue depth or request latency tracks the actual pain.

Autoscaling only helps if a new pod becomes useful quickly. A workload that takes two minutes to warm up cannot respond to a traffic spike, however aggressive the policy.

Stabilisation windows exist to stop flapping. Scaling down too eagerly after a spike means scaling back up moments later, which costs more than staying scaled.