Source: Exponential Backoff and Jitter (AWS Architecture Blog)
Marc Brooker's 2015 simulation comparing no backoff, plain exponential backoff, and Full/Equal/Decorrelated jitter; jitter cuts total calls by more than half under contention; Full and Decorrelated Jitter both beat Equal Jitter.
Contents
- What: engineering blog post with simulation results
- Author: Marc Brooker, AWS Architecture Blog
- Date: 2015-03-04 (updated May 2023: most AWS SDKs now implement backoff with jitter)
- Original: Exponential Backoff And Jitter
Summary
Simulates many clients contending for one resource and measures total calls and time to completion under different retry delays.
Key claims
- With N contending clients, total work grows roughly with N². Plain capped exponential backoff helps little because retries still arrive in synchronized clusters, and jitter does not change the N² nature of the work, though it reduces it substantially at realistic contention.
- Adding randomness spreads retries to a roughly constant rate; with 100 contending clients, jitter reduced the call count by more than half compared with un-jittered backoff.
- Variants: Full Jitter sleeps a random time across the whole capped backoff window; Equal Jitter keeps half the window fixed and randomizes the other half; Decorrelated Jitter grows the random window from the previous sleep.
- Equal Jitter is the weakest jittered option (similar work, longer completion). Full and Equal Jitter make about the same number of calls and Decorrelated makes more; Full Jitter uses less work but slightly more time than Decorrelated, so Brooker does not crown either. The fix is cheap, so it should be standard for remote clients.