50% CPU across one to four replicas, a 30-second downscale window — and resource requests before any of it works.