Skip to main content
HTTP 502 · type api_error. The gateway could not reach the extraction engine, so your request — a session create or an estimate — never started anything. Nothing was billed: no reservation was written, no run exists, and there is nothing to clean up. This is our infrastructure, not your request.

​
The shape

​
How to fix

  • Retry with backoff. The response carries no Retry-After header today — start around 10–30 seconds and back off; outages observed in practice cleared within about 10 minutes. A create is safe to retry with the same Idempotency-Key; an estimate is free either way.
  • Don’t misread partial availability. During an engine outage, paths that don’t need the engine (auth, reads, some parse-only declines) can keep succeeding while creates and estimates 502 — that’s the engine being down, not your queries being rejected.
  • If it persists beyond ~15 minutes, include request_id in a support message.

​
Where it sits in the catalog

engine_unavailable is an api_error like internal, but at 502 and with a stronger guarantee: the failure happened before anything started, so there is never a reservation to reconcile. Distinct from provider_unavailable, which is a stream-terminal code on a run that did start and then lost a non-model upstream.