Fabryka · OperationsRolling windows · UTC
Public service status
Service status.
Measured in public.
How the API's executions have actually ended: success, failures and time, per model and hour. Read from the inference gateway's own records, not from a test probe and not from a promise.
Reading the measurements from the gateway…
06 · Method
How we measure.
- Execution
- A request the gateway accepted and passed to a model. Each one leaves a record of how it ended and how long it took, never its content.
- Success rate
- Completed executions out of completed and failed. Executions the client abandoned are counted separately; without executions there is no rate.
- Time
- From admission to the last byte of the answer, streaming included, of completed executions. It grows with the answer's length; it is not the time to the first token.
- What it does not show
- There are no synthetic probes, so a quiet hour says nothing. Refusals before generation and the HTTP codes clients received are not recorded. The gateway recomputes the measurements at most every 30 seconds; the time of the measurement is shown under the banner.