Skip to content

VMAFxScoreLatencyBudgetBurn

Meaning. Too many synchronous Score requests take longer than the latency bound (monitoring.slo.scoreLatencySeconds, by default 30 seconds): the default objective (monitoring.slo.scoreLatency) is 99 % within the bound over 30 days. With the default settings: critical when more than 14.4 % of the requests of the last hour and its last 5 minutes were slower, warning above 6 % over 6 hours and 30 minutes. Recorded ratio: vmafx:score_slow_ratio:rate<window>.

Impact. Callers wait longer than they expect; clients with deadlines turn slow requests into errors.

Diagnose

  1. Are requests queueing behind the concurrency cap? Requests wait for a free slot before scoring; Overview Score request latency rising together with the request rate points at the cap.
  2. Did the work change? Longer or higher-resolution clips take longer: a latency rise with a steady rate and no deploy points at the inputs.
  3. Is the host saturated? CPU-bound scoring on a host with other load slows down; compare the host's CPU use.

Fix

Add replicas or raise VMAFX_MAX_CONCURRENT_SCORES within the host's cores; send long clips through controller jobs instead of synchronous requests.

Dashboard: Overview (Score request latency).