Kubernetes + autoscaling: gotchas and a clean solution

While implementing **autoscaling** in Kubernetes, I ran into tricky edge cases: - throttling spikes - inconsistent results Here's what I learned and the solution that finally worked.