Why more attempts, sooner, makes recovery slower, and why spacing retries out belongs in a library rather than at each call site.
8 March 2027
Exponential backoff does not improve any single attempt’s chance of succeeding. It changes when the attempt arrives. Each wait roughly doubles — 100 ms, 200 ms, 400 ms — so a failing dependency sees a thinning stream of requests instead of a constant flood, and gets idle time to work through the queue that is already there. A fixed one-second delay has none of that property: ten thousand clients retrying every second is still ten thousand requests a second, arriving a beat later. The schedule belongs in the client library, because a call site that writes its own loop picks its own numbers, and once every call site has done that nobody can state the service’s worst-case attempt rate. Jitter is the other half and it is next week: a fleet backing off on the same doubling clock stays in step.
Outline#
- why more attempts, sooner, makes recovery slower
- exponential backoff, and why a fixed delay is not enough on its own
- what belongs in a library instead of at the call site
Teaching team
- Marisol Quaye — convenor