# What is Exponential backoff?

Also called backoff with jitter, backoff.

Exponential backoff is a retry timing strategy in which the wait between attempts grows multiplicatively, for example one second, then two, then four. Random jitter is added so that many clients failing at once do not retry in synchronized waves. The approach gives an overloaded dependency time to recover instead of adding load during its worst moment.

The delay for each attempt is computed from a base value multiplied by a growth factor raised to the attempt number, then capped at a maximum. A random component, commonly called jitter, is applied so that clients that failed together do not return together. Full jitter, which picks a random delay between zero and the computed value, is a widely used variant.

The purpose is to reduce load on a dependency that is already struggling. Fixed-interval retries from many clients produce synchronized waves, sometimes called a thundering herd, that keep a recovering service saturated. Growing, randomized delays spread the traffic out, which shortens total recovery time even though each individual client waits longer than it would prefer.

Common errors include omitting jitter, which preserves the synchronization the strategy exists to break, and leaving the growth uncapped, so a later attempt is scheduled well beyond any useful window. If the response carries an explicit hint about when to retry, that instruction should take precedence over the computed delay. Backing off on errors that can never succeed just delays a certain failure.

Backoff is one component of a retry policy, not a substitute for one, and it does nothing about safety, so idempotency is still required. It pairs with rate limits, where the appropriate reaction to a rejection is to slow down, and with circuit breakers, which stop the calls entirely once the failures look persistent rather than transient.

## Key points

- Delays grow multiplicatively and should be capped at a maximum.
- Jitter prevents synchronized retry waves from many clients.
- An explicit retry-after hint overrides the computed delay.
- Backoff reduces load during an incident instead of increasing it.

## In practice

Two hundred workers hit a rate limit at the same second. With a fixed one-second retry, all two hundred return together and fail again, repeatedly. With exponential backoff and full jitter, the first retries spread across zero to one second, the next across zero to two, and so on. The dependency drains its backlog and every task completes without a second failure wave.

## Related terms

- [Retry policy](/en/glossary/retry-policy)
- [Rate limit](/en/glossary/rate-limit)
- [Circuit Breaker](/en/glossary/circuit-breaker)
- [Queue](/en/glossary/queue)

[Back to the AI Glossary](/en/glossary)