Idempotency keys

An idempotency key is a unique value that a client attaches to a request that would otherwise not be safe to repeat, such as creating a payment or launching a server. The server uses the key—often together with a fingerprint of relevant request data—to recognise retries and apply a documented duplicate-request policy. That policy is provider-specific: it can replay a stored response, report the earlier operation's result, reject a mismatched or concurrent reuse, or ensure that an already successful operation is not performed again. The mechanism exists because network failures are ambiguous: a client that never receives a response cannot tell whether its request was lost, failed midway, or succeeded. Within a provider's documented scope and retention window, an idempotency key can let the client retry without repeating the protected server-side operation. It does not by itself guarantee end-to-end exactly-once behavior or make every ambiguous failure safe to retry. (^[9]; provider comparison)

The pattern is widely used in payment and cloud-infrastructure APIs, and each provider publishes its own contract. Those contracts agree on the core mechanism but differ on details that matter in practice: where the key is sent, what the key is scoped to, how long the server remembers it, what happens when a key is reused with different parameters, and whether a failed first attempt is replayed or re-executed. This article describes the shared mechanism, then compares those properties across documented implementations. It is about the API-level contract; recovery strategies built on top of it, such as checkpoints and reconciliation after an uncertain outcome, are covered in Idempotency and recovery in agent tool calls.

Idempotence is a mathematical concept also used in computer science; HTTP applies it to request-method semantics (^[9]). RFC 9110 classifies methods by whether their intended effect on the server changes when an identical request is repeated. RFC 9110 states that "a request method is considered idempotent if the intended effect on the server of multiple identical requests with that method is the same as the effect for a single such request", and lists GET, HEAD, PUT, DELETE, OPTIONS, and TRACE as idempotent (^[1]). Two qualifications in the same section shape everything that follows. First, idempotence is about the server's state, not the response: "the fact that a method is idempotent does not imply that each request will receive the same response status code." Second, "the definition of idempotent only considers the intended effect on the server and does not consider side effects such as the logging of requests, charges incurred on third-party servers, or an increase in advertisement counter statistics" (^[1]).

POST is the notable omission: HTTP does not define the method itself as idempotent. A particular resource can still give a POST idempotent semantics. RFC 9110 therefore says a client should not automatically retry a non-idempotent method unless it knows the request semantics are actually idempotent or can detect that the original request was never applied. Idempotency keys are a common application-level contract for making retries of particular POST operations safer; they do not change the semantics of every POST. Stripe's documentation makes its own boundary explicit: "All POST requests accept idempotency keys. Don't send idempotency keys in GET and DELETE requests because it has no effect. These requests are idempotent by definition" (^[1]) (^[2]).

A retry is only dangerous when the client cannot tell what happened to the original request. Brandur Leach, writing for Stripe in 2017, enumerated the cases: "The initial connection could fail as the client tries to connect to a server. The call could fail midway while the server is fulfilling the operation, leaving the work in limbo. The call could succeed, but the connection break before the server can tell its client about it" (^[3]). From the client's side all three look identical: a timeout or a dropped connection. Retrying blindly is correct in the first case, ambiguous in the second, and creates a duplicate in the third.

Amazon's EC2 documentation describes the same problem from the infrastructure side. Mutating requests "typically return a result before the operation's asynchronous workflows have completed", operations "might also time out or encounter other server issues before they complete, even though the request has already returned a result", and if "the original request and the subsequent retries are successful, the operation is completed multiple times. This means that you might create more resources than you intended" (^[4]).

The mechanism has three parts.

The client generates a key. The key must be unique per logical operation, not per HTTP attempt: every retry of the same operation reuses the same key, and a genuinely new operation gets a new one. Stripe suggests "V4 UUIDs, or another random string with enough entropy to avoid collisions", allows keys "up to 255 characters long", and warns clients to "avoid using sensitive data (for example, email addresses or personal identifiers) as idempotency keys" (^[2]). Adyen recommends a UUID with "a maximum of 64 characters" (^[5]); EC2 accepts "a unique, case-sensitive string of up to 64 ASCII characters" (^[4]).

The server stores the key with the outcome. Stripe's description is representative: the layer works "by saving the resulting status code and body of the first request made for any given idempotency key, regardless of whether it succeeds or fails. Subsequent requests with the same key return the same result, including 500 errors" (^[2]). PayPal describes the same behaviour from the caller's perspective: if "the initial call fails with the HTTP 500 status code but the server has already refunded the payment, the caller does not need to worry that the server will refund the payment again" (^[6]).

The server checks that a replay really is a replay. A key reused by accident on a different operation must not silently return an unrelated stored response. Implementations therefore compare the retry's parameters against the original. Stripe's layer "compares incoming parameters to those of the original request and errors if they're not the same to prevent accidental misuse" (^[2]); EC2 fails the retry "with an IdempotentParameterMismatch error" (^[4]). The IETF draft generalises this comparison as an idempotency fingerprint, which a server may derive from a checksum of the payload, selected fields, or a request digest or signature (^[9]). A fingerprint is a digest used to identify content, the same idea that underlies content-addressed storage, applied here to a request rather than a stored artifact.

Providers agree on the mechanism above but not on its parameters. Six properties—transport plus the five discussed below—determine what a client can actually rely on, and they should be read from each provider's documentation rather than assumed.

The key is unique within some boundary, and the boundary is provider-defined. Adyen states that "keys are stored at a company account level. The system checks that the idempotency keys are unique to the company account" (^[5]). Square requires that keys "be unique for every CreatePayment request", which scopes them to an operation type (^[7]). EC2 distinguishes regional idempotency, where "you can use the same request, including the same client token, in a different Region", from zonal idempotency, where two AllocateHosts calls with the same token "succeed if they specify different values for the AvailabilityZone parameter" (^[4]). The IETF draft leaves the choice open: "uniqueness of the key MUST be defined by the resource owner and MUST be implemented by the clients of the resource" (^[9]).

A stored key eventually expires, and after expiry a "retry" is a new request. Stripe removes keys "automatically after they're at least 24 hours old" and will "generate a new request if a key is reused after the original is pruned" (^[2]). Adyen keys "are valid for a period of 7 to 14 days after first submission" (^[5]). PayPal says the server stores the ID "for a period of time" that depends on the API, and gives the refund endpoint as an example where "the user can make the call again with the same ID … for up to 45 days" (^[6]). The consequence for clients is that a retry window longer than the retention period is not protected, and an application that queues failed operations for later retry needs to know the number.

Several documented implementations reject a key reused with different content, but the signal differs; the cited Adyen and PayPal pages do not state a mismatch rule. EC2 returns IdempotentParameterMismatch (^[4]); Stripe "errors" (^[2]); Square returns "an error indicating that you used the idempotency key previously", adding that "this behavior might vary depending on the API" (^[8]). The IETF draft specifies 422 Unprocessable Content with a problem-details title of "Idempotency-Key is already used" (^[9]).

A retry can arrive while the original is still executing. The IETF draft separates this from mismatch and specifies 409 Conflict with the title "A request is outstanding for this Idempotency-Key" (^[9]). Adyen documents that a duplicate submitted "before the first request has completed" returns "HTTP 422 – Unprocessable Entity or HTTP 409 - Conflict status with the error code 704: 'request already processed or in progress'" (^[5]). Stripe treats this case as one where no result is saved at all: if "the request conflicts with another request that's executing concurrently, we don't save the idempotent result because no API endpoint initiates the execution. You can retry these requests" (^[2]).

This is the property most often assumed and least often checked. Stripe saves and replays everything once execution starts, "including 500 errors", but "only after the execution of an endpoint begins", so validation failures are not stored and can be retried after correction (^[2]). EC2's guidance describes idempotency in terms of successful completion, "if the original request completes successfully, any subsequent retries complete successfully without performing any further actions", and separately advises retrying 500-series responses with backoff and not retrying most 400-series responses (^[4]). Adyen and PayPal do not state whether failed first attempts are replayed. A client that retries a cached 500 with the same key at Stripe will receive the same 500 again; whether that is the desired behaviour depends on whether the failure was transient, which the client cannot always know.

"Not stated" means the cited page does not say; the provider may document it elsewhere or behave in an undocumented way. The table reflects the pages as retrieved in September 2026.

The IETF HTTP APIs working group adopted a draft, The Idempotency-Key HTTP Header Field, by Jayadeba Jena and Sanjay Dalal, whose abstract states that the header "can be used to make non-idempotent HTTP methods such as POST or PATCH fault-tolerant". Revision 07 is dated 15 October 2025, defines the field as an Item Structured Header whose "value MUST be a String", and had expired by April 2026 without progressing to an RFC (^[9]). Its main contributions are the vocabulary used above, in particular the fingerprint concept and the separation of the mismatch case (422) from the in-progress case (409). Whether the working group revives the draft is not known at the time of writing; the provider contracts in the table predate it and do not depend on it.

Keys are not end-to-end exactly-once delivery. An idempotency key protects one hop: the client's request to one server. If that server calls a third party, RFC 9110's exclusion of "charges incurred on third-party servers" applies, and the downstream call needs its own key or its own reconciliation (^[1]). Keys also do nothing for operations that were never sent, so a client that crashes between generating a key and sending the request must persist the key first if it wants to resume.

Retention bounds the retry window. Because keys expire, a client should retry within the documented window and treat anything later as a new operation requiring a fresh check of the server's state, for example by querying for the resource before creating it.

A replayed error is still an error. Where failures are cached, the retry returns the original failure. The client has to decide whether to issue a new key, and it should do so only after establishing that the original did not take effect; Stripe's rule that results are stored only once execution begins means a stored 500 may correspond to a partially completed operation.

Keys should carry no meaning. Random keys with enough entropy avoid collisions and, as Stripe notes, avoid leaking personal identifiers into server-side logs (^[2]). Deriving a key deterministically from the operation's content is sometimes done so that independent processes converge on the same key, but it turns the key into a fingerprint and reintroduces the collision and privacy questions the random approach avoids.

Retries still need pacing. Where the provider contract prevents duplicate execution, the key can make a retry safer; it does not make the retry cheap. Leach's post pairs idempotency with exponential backoff, waiting a time proportional to "2^n, where n is the number of failures", and adds "random 'jitter' to each client's wait time" so that aligned retry schedules do not overwhelm a recovering server (^[3]). See Exponential backoff with jitter.

• Idempotency and recovery in agent tool calls: checkpoints, reconciliation, and completion verification built on top of the contract described here.
• Exponential backoff with jitter: the pacing half of a safe retry policy.
• Content-addressed storage: digests as identifiers, the idea behind idempotency fingerprints.
• Exactly-once delivery: the broader distributed-systems problem that idempotency keys partially address.
• HTTP request methods: the safe and idempotent method classes that idempotency keys extend to POST.