Question

To 'improve density', someone lowers the CPU limit on the billing worker Deployment from 2000m to 500m via a config change merged at 17:00 (no image change). Throughput on the billing queue drops 60% and per-job latency triples, but CPU usage dashboards show the pods using only ~480m — well under the new limit — and there are no OOMs or restarts. The nodes have spare CPU. Triage and explain the mechanism, then mitigate.