@sdxc/idempotency

Idempotency-Key requests: replay the first response, refuse conflicting reuse.

Installation

npm add @sdxc/idempotency

The middleware runs on the remix router, the durable store over remix/data-table; refusals are @sdxc/problem documents, durations are @sdxc/duration inputs, and results are @sdxc/result values. All install alongside this package.

A POST that times out leaves its client unsure whether the resource was created. The IETF Idempotency-Key draft settles it: the client sends a unique Idempotency-Key with the request, the server runs it once and stores the outcome, and every retry with that key gets the stored outcome back. This package implements the server side as router middleware, plus the client helpers that mint keys and put them on requests. The server depends on one guarantee from its store, an atomic claim: of two concurrent requests with the same key, exactly one runs the handler.

Usage

Send A Key

import { generateIdempotencyKey, withIdempotencyKey } from "@sdxc/idempotency/client";
import { unwrap } from "@sdxc/result";

let key = generateIdempotencyKey(); // once per operation, reused by every retry
let response = await fetch(url, unwrap(withIdempotencyKey({ method: "POST", body }, key)));
// Idempotency-Key: "8e03978e-40d5-43e8-bc93-6894a57f9324"

The key is an RFC 9651 sf-string, so it travels quoted.

Protect A Route

import { idempotency } from "@sdxc/idempotency/middleware";
import { MemoryStore } from "@sdxc/idempotency/memory";

let idempotent = idempotency({
	store: new MemoryStore(),
	scope: (ctx) => ctx.get(Caller).id, // whoever your authentication middleware identified
	ttl: "24 hours",
});

router.post(routes.orders.create, {
	middleware: [idempotent],
	handler: async (ctx) => createOrder(ctx), // runs once per key
});

For each POST or PATCH, the middleware answers:

Store Records In SQL

import { DataTableStore, purgeExpired } from "@sdxc/idempotency/data-table";
import { idempotency } from "@sdxc/idempotency/middleware";

let idempotent = idempotency({
	store: (ctx) => new DataTableStore(ctx.get(Db)), // a remix/data-table Database
	scope: (ctx) => `api-key:${ctx.get(Caller).id}`,
	ttl: "24 hours",
});

let purged = await purgeExpired(db); // from a scheduled job: Result<number, IdempotencyStoreError>

Create the table by pasting IDEMPOTENCY_KEYS_SCHEMA_SQL into a migration.

API

@sdxc/idempotency

readIdempotencyKey(headers, options?)

Reads the Idempotency-Key field as an RFC 9651 Item whose value is a non-empty sf-string, into a Result<string | null, IdempotencyKeyError>. Parameters are ignored, and an absent header succeeds with null. options.maxLength (ReadIdempotencyKeyOptions) bounds the key, 255 characters by default.

formatIdempotencyKey(key)

Serializes a key as the field value, quoted and escaped, into a Result<string, IdempotencyKeyError>. Fails for an empty key or one with characters outside printable ASCII.

IDEMPOTENCY_KEY_HEADER

"Idempotency-Key".

fingerprint(request)

Resolves to the SHA-256 (hex) over the method, the path with its query, the content type's essence (no parameters, lowercased) and the body bytes. The origin is excluded. It consumes the body, so pass a clone.

IdempotencyStore

The contract the middleware claims through, three methods each resolving to a Result with an IdempotencyStoreError failure:

interface IdempotencyStore {
	claim(request: ClaimRequest): Promise<Result<ClaimOutcome, IdempotencyStoreError>>;
	complete(
		id: string,
		lease: string,
		response: StoredResponse,
		expiresAt: number,
	): Promise<Result<void, IdempotencyStoreError>>;
	release(id: string, lease: string): Promise<Result<void, IdempotencyStoreError>>;
}

claim is atomic: of concurrent claims on one id, exactly one answers { status: "claimed", lease }; the others see { status: "in-flight", fingerprint } or { status: "completed", fingerprint, response }. A record is claimable again once it expires, or while in flight past its lease (an invocation that died). complete and release act only while the lease still holds, so a resumed stale invocation never overwrites the outcome of the retry that took its claim over; release also leaves a completed record alone.

ClaimRequest, ClaimOutcome, StoredResponse

A claim carries the record id, the request fingerprint (or null), now, leaseMs and expiresAt, all in epoch milliseconds. A StoredResponse is plain JSON: status, lowercased header pairs without Set-Cookie, and the body as base64.

IdempotencyKeyError, IdempotencyStoreError

The key the client sent is unacceptable, with reason "invalid" or "too-long"; the store could not answer.

IDEMPOTENCY_PROBLEM_ENTRIES

The four @sdxc/problem catalog entries, idempotencyKeyMissing (400), idempotencyKeyInvalid (400), idempotencyKeyInUse (409) and idempotencyKeyReused (422), to spread into a defineProblems catalog. Their slugs are the wire contract clients match on.

@sdxc/idempotency/middleware

idempotency(options)

The record id is the SHA-256 of prefix, scope, method, pathname and key. A handler that throws releases the key and the error propagates. Events go to the invocation's @sdxc/logger log (idempotency.replayed, idempotency.in_flight, idempotency.reused, idempotency.store_unavailable) without the key or the scope.

IdempotencyOptions, IdempotencyProblems

The options above, and the four builders problems takes, which a catalog containing IDEMPOTENCY_PROBLEM_ENTRIES satisfies.

@sdxc/idempotency/data-table

DataTableStore

new DataTableStore(db) stores records in a remix/data-table Database whose schema includes idempotency_keys. claim is one conditional upsert followed by a read, so it stays atomic on databases without interactive transactions, such as Cloudflare D1, as well as on Durable Object SQLite. complete and release are single statements guarded by id and lease.

idempotencyKeys, IdempotencyKeyRow

The table definition and its row type.

IDEMPOTENCY_KEYS_SCHEMA_SQL

create table plus an index on expires_at, for a migration. Apply it through a migration runner: D1's exec() treats each line as its own statement.

purgeExpired(db, now?)

Deletes every record whose expires_at has passed and resolves to how many went, as a Result<number, IdempotencyStoreError>.

@sdxc/idempotency/memory

MemoryStore

new MemoryStore() is an in-process store for tests and single-isolate development, atomic within one isolate.

@sdxc/idempotency/client

generateIdempotencyKey()

A random UUID.

deriveIdempotencyKey(...parts)

Resolves to the SHA-256 (hex) of the parts, keeping part boundaries: the same parts always give the same key.

withIdempotencyKey(init, key)

A copy of init with the header set, other headers kept, as a Result<RequestInit, IdempotencyKeyError>.

applyIdempotencyKey(request, key?)

Sets the header on a POST or PATCH that has none, generating a UUID when key is omitted, as a Result<Request, IdempotencyKeyError>. Other methods and requests that already carry a key come back unchanged.

Pattern: Refusals From Your Problem Catalog

import { IDEMPOTENCY_PROBLEM_ENTRIES } from "@sdxc/idempotency";
import { DataTableStore } from "@sdxc/idempotency/data-table";
import { idempotency } from "@sdxc/idempotency/middleware";
import { defineProblems } from "@sdxc/problem";

const PROBLEMS = defineProblems("https://docs.example.com/errors/", {
	...IDEMPOTENCY_PROBLEM_ENTRIES,
	notFound: { slug: "not-found", status: 404, title: "The resource does not exist" },
});

let idempotent = idempotency({
	store: (ctx) => new DataTableStore(ctx.get(Db)),
	scope: (ctx) => `api-key:${ctx.get(Caller).id}`,
	ttl: "24 hours",
	required: true,
	problems: PROBLEMS,
});

Mount it after authentication, since the scope needs the caller, and after rate limiting, so a replay still spends budget. Publish the ttl in the API reference; the draft asks servers to state when keys expire.

Pattern: A Durable Object As The Store

When records belong next to one tenant's data, the object runs a DataTableStore over its own SQLite and exposes the three methods over RPC; the Worker hands the middleware a thin adapter:

idempotency({
	store: (ctx) => {
		let stub = ctx.get(TenantStub);
		return {
			claim: (request) => stub.idempotencyClaim(request),
			complete: (id, lease, response, expiresAt) =>
				stub.idempotencyComplete(id, lease, response, expiresAt),
			release: (id, lease) => stub.idempotencyRelease(id, lease),
		};
	},
	scope: (ctx) => ctx.get(Caller).id,
	ttl: "24 hours",
});

Pattern: Keys For Retried Jobs

A queue message id is the same on every delivery, so a key derived from it survives the job's own retries with nothing stored:

import { deriveIdempotencyKey, withIdempotencyKey } from "@sdxc/idempotency/client";
import { unwrap } from "@sdxc/result";

let key = await deriveIdempotencyKey(message.id, "create-order");
await fetch(url, unwrap(withIdempotencyKey({ method: "POST", body }, key)));

Clients must resend the same bytes: the same JSON with keys reordered fingerprints differently and gets 422. A handler that commits a write and then answers 5xx needs its own shouldStore, or a retry repeats the write.