Usage metering
Every completed inference request is recorded in the state-dir database, tagged with the calling key,
the served model, and the worker that ran it. atlas usage summarizes the ledger (cumulative since it
was created).
atlas usageatlas usage --json # machine-readable- Totals are durable across restarts (they live in the state-dir database).
- A stream interrupted partway (worker drop or client disconnect) still records the output emitted before the cut, so the ledger isn’t systematically short on interrupted requests.
- Cloud-fallback spend, when enabled, is recorded as a distinct usage class —
atlas usageshows it under acloud:<provider>worker, so external spend is visible and separable from local capacity. Cloud-fallback is off by default. See API compatibility.
For live throughput and queue depth rather than cumulative totals, see Observability.