Queue-Based Retry Design for OpenAI Moderation API Rate Limits

A 429 rate-limit response should be treated as scheduling input rather than a reason for every layer to retry at once. The article recommends a bounded retry inside a queue worker that carries the candidate and rubric revisions through each attempt. Operators should log HTTP status, Retry-After, rate-limit headers, request IDs, and attempt numbers, and alert on queue age and retry exhaustion. Retries should use exponential backoff with jitter, honoring Retry-After where provided, under capped delays and attempt counts, while reusing the same decision ID to prevent duplicate scores. Transport failures, malformed output, and rate limits should be handled as distinct states, ending in a visible delayed state or terminal review rather than unbounded retries.