Jittered Backoff Outperforms Fixed Retry in Free Model Server Stress Test

A developer tested four retry strategies — no retry, fixed retry, exponential backoff, and jittered backoff — against a free AI model server, sending 200 requests per strategy. The experiment measured success rate, p95 latency, and total errors to understand how retry behavior affects shared server capacity. Fixed retry showed minimal improvement over no retry and nearly doubled p95 latency, as simultaneous retries from multiple clients overwhelmed the already-struggling server. Exponential backoff performed better by spacing out retry attempts, while jittered backoff — which randomizes wait intervals — achieved the highest success rate of 94% and the lowest error count. The findings highlight that naive or fixed retry logic can worsen server congestion, and randomized backoff is a more considerate strategy for shared infrastructure.