iapi.buzz
Rate limits · reviewed November 2026

API Rate Limits, 429 Errors and Retry Backoff: How to Handle Them Without Making Things Worse

Sooner or later an API tells your program to slow down. This page explains the 429 status code and the Retry-After header, shows which failures are worth retrying, compares the standard backoff strategies, and gives you a calculator that turns your settings into a concrete retry schedule.

Why rate limits exist

An API serves many clients from the same machines. A limit on requests per window stops one client from using up the shared capacity, and it keeps a buggy loop from taking a service down. The limit is a promise about fair use, and the status code that enforces it is 429 Too Many Requests.

RFC 6585 defines 429 as a signal that the user sent too many requests in a given amount of time. Providers choose their own windows and numbers, so the only reliable figures are the ones in the documentation of the API you call. This page therefore gives you the method, not a list of provider limits.

The headers that tell you how long to wait

Two headers matter most. The standard one is Retry-After, and many providers add their own counters such as the requests remaining in the window. Because the counter names differ by provider, read your API documentation for them.

Retry-After as seconds

A whole number of seconds to wait before the follow-up request.

Retry-After: 3600

Retry-After as a date

A fixed moment in HTTP-date format. Convert it to a delay from your own clock.

Retry-After: Fri, 31 Dec 1999 23:59:59 GMT

Where it can appear

RFC 9110 describes it with 503 and 3xx responses, and RFC 6585 says a 429 response may include it.

Source: IETF RFC 9110 section 10.2.3 (Retry-After); RFC 6585 section 4 (429 Too Many Requests).

Which failures are worth retrying

A retry helps only when the cause is temporary. A request that is wrong stays wrong, and a retry of it just adds load. The table gives this site's rule of thumb, based on the meanings in RFC 9110 and RFC 6585.

ResponseRetry?Reason
400, 401, 403, 404, 422NoThe request or the credentials are at fault. Fix them first.
429 Too Many RequestsYes, after waitingThe limit resets with time. Use Retry-After when present.
500 Internal Server ErrorCautiouslyThe cause is unknown. Retry only methods that are safe to repeat.
502 Bad Gateway, 504 Gateway TimeoutCautiouslyAn upstream server failed or was slow. It often clears.
503 Service UnavailableYes, after waitingThe server is overloaded or in maintenance, and may send Retry-After.
501 Not ImplementedNoThe server lacks the needed function. Waiting will not add it.

“Cautiously” refers to idempotent methods. GET, PUT and DELETE can be repeated with the same effect, as RFC 9110 defines them. A POST that creates something may already have worked, so check the state or use an idempotency key before you repeat it.

Backoff strategies compared

If every client that failed retries after exactly one second, they all return at the same moment and the server fails again. Backoff spreads the retries out, and jitter adds randomness so that clients do not move in step. The AWS Architecture Blog post “Exponential Backoff And Jitter” compares the variants. It concludes that plain exponential backoff without jitter is the clear loser, because it takes more work and more time than the jittered versions.

Exponential, no jitter

Wait base times 2 to the power of the attempt number, up to a cap.

sleep = min(cap, base * 2^attempt)

Full jitter

Pick a random wait between zero and the exponential value.

sleep = random(0, min(cap, base * 2^attempt))

Equal jitter

Keep half of the exponential value and randomize the other half.

temp = min(cap, base * 2^attempt)
sleep = temp/2 + random(0, temp/2)

Source: AWS Architecture Blog, “Exponential Backoff And Jitter” (Marc Brooker). Attempt numbers in this page start at 0 for the first retry. Always set a cap and a maximum number of retries so that a long outage does not make your program wait forever.

Retry schedule calculator

RetryShortest wait (s)Longest wait (s)Longest total so far (s)

Method: wait for retry n is min(cap, base times factor to the power n), counting n from 0. Full jitter shows a range from 0 to that value. Equal jitter shows half of it up to the full value. The longest total adds the longest waits, so a real run will usually finish sooner with jitter. Pacing is the window length divided by the allowed requests, which is an average spacing, not a guarantee about how a provider counts a window.

Frequently asked questions

What should my client do when it gets a 429?

Stop sending the request, wait, and then try again. If the response has a Retry-After header, use that value. If it has none, use a growing delay with jitter and a limit on the number of attempts.

What is the difference between 429 and 503?

A 429 says you sent too many requests in a given time. A 503 says the server cannot handle the request now, for example from overload or maintenance. Both can carry Retry-After, and both are usually worth retrying after a wait.

Why add jitter instead of waiting a fixed time?

With a fixed wait, every client that failed together retries together and creates the same spike again. Random waits spread the retries over time, which is the point of the AWS Architecture Blog comparison.

How many retries are enough?

There is no universal number. Choose a small maximum, add a cap on the single wait, and report the failure to the user or a log when you reach it. The right values depend on how long your provider says outages and limits last.

About this site

iapi.buzz is an independent informational site about the everyday mechanics of web APIs. It sells nothing and does not review or rank any provider. The standards named here are IETF RFC 9110 and RFC 6585, and the backoff comparison comes from the AWS Architecture Blog. Last reviewed: November 2026. Educational content, not professional advice.