# Retries and backoff in plain words

> A retry is trying a failed operation again. Backoff is waiting longer before each new try. When retrying helps, when it makes an outage worse, and the four settings every retry needs.

Source: https://hintvo.com/blog/retries-and-backoff-in-plain-words | Last updated: November 20, 2026 | Reviewed: 2026-11-20

A retry is trying a failed operation again, in the hope that the failure was temporary. Backoff is waiting a little longer before each new try, so that a struggling service gets room to recover instead of a flood of repeats.

## When a retry helps

Some failures go away on their own: a dropped connection, a timeout, a service that was restarting. Trying again a moment later often works. Other failures never go away, and retrying them only wastes time.

| Failure | Retry? |
| --- | --- |
| A timeout or a dropped connection | Yes |
| The service says it is overloaded or temporarily unavailable | Yes, after a wait |
| An invalid request or a missing permission | No. It will fail the same way |
| Not found | Usually not |

## Why you wait longer each time

If a thousand clients all retry immediately, a service that was already struggling gets a thousand extra requests at its worst moment. Waiting one second, then two, then four, spreads the load out. Doubling the wait like this is called exponential backoff.

## Why you add randomness

If every client waits exactly one second, they all come back together one second later. Adding a small random amount to each wait, called jitter, spreads them apart.

```text
attempt 1  fails
           wait about 1 second
attempt 2  fails
           wait about 2 seconds
attempt 3  fails
           wait about 4 seconds
attempt 4  succeeds, or give up and report the failure
```

## The four settings every retry needs

- **How many attempts.** There must be a limit. Retrying forever turns one failure into a permanent one.
- **How long to wait,** and how the wait grows.
- **Which failures to retry.** Only the temporary ones.
- **What happens when it gives up.** Someone or something has to find out: a list of failed jobs, an alert, a message to the user.

## The trap: repeating something that is not safe to repeat

A retry repeats the operation. If the operation charges a card or sends an email, a retry after a lost reply does it twice. That is why retries and [idempotency](https://hintvo.com/blog/idempotency-in-plain-words) are designed together, and why the idempotency comes first.

## Questions you will hear in a meeting

- Where does the retry live: in the caller, in the queue worker, or in both?
- If two layers each retry three times, do we really make nine attempts?
- What does the user see while we are retrying?
- Where do the jobs go that failed every attempt?

> **Where Hintvo fits:** when a term like this comes up in a call, the In plain words card explains it in one sentence, in the context of your project. When someone asks how your own code handles it, the answer names the file and line. In the sample meeting on the [home page](https://hintvo.com/#replay), the question is where the retry logic should live, and the answer comes from the project's own queue consumer.

## Related

- [Idempotency in plain words: why a retry must not charge twice](https://hintvo.com/blog/idempotency-in-plain-words.md): What does idempotency mean, in plain words?
- [Message queues in plain words](https://hintvo.com/blog/message-queues-in-plain-words.md): What is a message queue and when do you need one, in plain words?
- [Incident calls: find the code path while people are talking](https://hintvo.com/blog/incident-call-find-the-code-path-fast.md): How do I find the failing code path quickly during an incident call?
