Queue capacity and latency must be validated separately

Increasing queue capacity can eliminate overload rejections while still violating the latency objective. After removing admission errors, rerun the highest supported concurrency and select worker parallelism by service-local p95, not acceptance alone. Keep downstream model concurrency independently bounded so one stage cannot monopolize the accelerator, and verify readiness plus memory after overload. Client-network latency should be reported separately from service processing latency.

Applies when: A multi-stage GPU pipeline uses bounded queues and must survive burst concurrency. Keywords: overload queue capacity p95 latency GPU pipeline concurrency readiness