Single consumer starves queue

A RabbitMQ classic queue image-thumbnail is consumed by 10 worker pods. At 15:30 queue depth (ready) climbs from ~200 to 60k and keeps rising. Dashboards: total consumer throughput dropped to roughly one tenth of normal even though all 10 pods are up and connected. Per-pod metrics show ONE pod at high CPU pulling a large unacked count pinned at its prefetch (which was recently raised to 5000), while the other 9 pods sit at near-zero unacked and almost idle. No errors, no redeliveries. Recent context: yesterday someone raised prefetch (QoS) from 50 to 5000 to 'improve throughput.' How do you triage and mitigate?