Xinference model request duration histogram

Whole-request wall time (queue + inference), not a phase breakdown. 13 finite buckets plus float('inf'). Because the tail is dense in the sub-second range and the finite ladder stops at 120 s, a slow-but-not-pathological request lands in the overflow bucket and the histogram cannot distinguish a 3-minute request from a 30-minute one - a percentile above 120 s is not computable from this ladder. Xinference main (xorbitsai/inference), read 2026-10-05. NOT ON BY DEFAULT on the worker: the /metrics endpoint is only mounted when XINFERENCE_DISABLE_METRICS is unset AND a metrics exporter host or port is supplied; otherwise launch_metrics_export_server() is never called and the endpoint does not exist. Supervisor-side metrics are deregistered from the worker's registry on purpose (launch_metrics_export_server removes _SUPERVISOR_ONLY_METRICS), so the worker endpoint carries only worker metrics and the supervisor endpoint only supervisor metrics. aioprometheus, not prometheus_client: counters do not auto-suffix _total the way prometheus_client does, so two of these names end in _counter and three do not - match on the literal name.