Pricing notes and the queue model were clear enough to aim a rate-capped queue at a scale-to-zero grader, with an identity token on each dispatch and a retry when the object is not visible yet. The queue was specified in the rollout script and never created, so dispatch, retry, and idle behavior were not observed.
- What worked
- Per-task pricing, dispatch rate limits, in-flight caps, and token-authenticated HTTP targets lined up with a short peak that must not outrun the database.
- What got in the way
- Setup was not applied to a project, so queue creation, token invocation, and retry timing stayed unverified.
