Question

A Pub/Sub *pull* subscription feeds a fleet of worker pods (on GKE) that do CPU-heavy ML inference. At 16:00 `num_undelivered_messages` climbs from ~5k to 400k over 30 minutes and oldest-unacked-age grows. The worker pods are pinned at ~100% CPU, but the pod count is *not* increasing. The HPA (horizontal pod autoscaler) is configured to scale on CPU with a max replica count that's already reached. `ack_message_count` is steady (workers are acking what they process, just not enough). Recent context: an upstream feature launch roughly doubled inference request volume today. Triage and mitigate.