Retry queues sound safe until surge traffic turns them into a deadline killer. When AI calls hit provider rate limits, the queue silently swallows hours you budgeted for debugging. I now plan for burst capacity up front, or shed load deliberately, instead of trusting retries to absorb the spike.