Question

An order-fulfillment service uses RabbitMQ with a delayed-retry design: the `fulfillment` work queue dead-letters (via DLX) to `fulfillment.wait` which has a per-message TTL of 60s and, on TTL expiry, dead-letters *back* to `fulfillment` to be retried. There is no explicit attempt counter. At 09:10 broker CPU spikes to ~90%, network throughput triples, and the `fulfillment.wait` queue depth oscillates around 80k while `fulfillment` ready-count flickers. The main consumers are healthy and acking normally; effective *useful* throughput has actually dropped. Recent context: a config push 30 min ago made the `inventory` gRPC call time out for a specific SKU family that's ~5% of orders. How do you triage and mitigate?