- HTTP 5xx on any endpoint: safe to retry with backoff (30s, 60s, 120s). Job creation is idempotent on identical input.
- Job stuck in
queued: the queue reports position and estimated wait; sustained stalls beyond 2× the postedetaSecondswarrant a retry or a support ping. - HTTP 429: you’ve hit a concurrency limit or rate limit; back off and retry, don’t tight-loop.
Setting up
Infrastructure errors
5xx, timeouts, and retries: the failures that are on us.
Distinct from refusals (facts about your material),
these are faults in our infrastructure. Failed jobs are not charged.