A municipal AI inference cluster needs fabric congestion visibility per queue. Which telemetry focus best matches that requirement?

Queue and congestion telemetry shows where buffers build on fabric interfaces—matching a municipal AI inference cluster’s need for per-queue congestion visibility. Empty rack-unit placeholders are not congestion telemetry. Cafeteria wait times are unrelated. Disabling queue counters removes the required signal. Exam caveat: East-west AI bursts often show first as queue pressure, not as a simple link-down. Operational check: Identify top congested queues during a job burst and map them to the exact leaf/spine interfaces. In civic data-center operations, record the expected CLI or collector evidence for this concept so the next on-call engineer can validate fabric behavior without re-deriving the design under incident pressure. In civic data-center operations, record the expected CLI or collector evidence for this concept so the next on-call engineer can validate fabric behavior without re-deriving the design under incident pressure. In civic data-center operations, record the expected CLI or collector evidence for this concept so the next on-call engineer can validate fabric behavior without re-deriving the design under incident pressure. In civic data-center operations, record the expected CLI or collector evidence for this concept so the next on-call engineer can validate fabric behavior without re-deriving the design under incident pressure.