Webhook Integration

Webhook Integration

Intro

Webhooks deliver at least once, in no particular order, over a network that fails. A correct webhook integration verifies HMAC signatures on the raw body, deduplicates by event ID, returns 2xx immediately and processes asynchronously, and retries with exponential backoff into a dead-letter queue.

Overview

Payload envelope

{
  "id": "evt_a1b2c3d4",
  "type": "order.shipped",
  "created_at": "2026-03-22T10:30:00Z",
  "data": { "order_id": "ord_xyz789", "carrier": "ups" }
}

• A unique event id for idempotency.
• A dotted type of the form resource.action (order.shipped, invoice.payment_failed).
• ISO 8601 created_at for ordering and replay-window enforcement.
• Domain payload nested under data so the envelope stays stable as the domain model evolves.

Signature verification (HMAC-SHA256)

Senders compute an HMAC over the raw request body using a shared secret and send it in a header (e.g. X-Webhook-Signature: sha256=<hex>). Consumers recompute and compare in constant time:

expected = hmac.new(secret.encode(), raw_body, hashlib.sha256).hexdigest()
if not hmac.compare_digest(expected, received_signature):
    return 401

Three things matter:

1. Sign the raw body, not parsed JSON. Re-serializing reorders keys and breaks the signature.
2. Constant-time comparison (hmac.compare_digest) — never == — to avoid timing attacks.
3. Replay protection. Include a timestamp in the signed payload or as a separate signed header, and reject anything older than ~5 minutes.

Support multiple active secrets so you can rotate without downtime.

Idempotency

Webhook delivery is at-least-once. Consumers must dedupe:

• Store processed event IDs in a database or cache with a unique constraint.
• Check before processing — INSERT ... ON CONFLICT DO NOTHING is the canonical pattern.
• TTL the idempotency store (e.g. 7 days) to bound storage. Senders should not retry beyond that window.

Retry handling

As a sender: retry on transient failure with exponential backoff (1 min, 5 min, 30 min, 2 hr, 8 hr, 24 hr is a typical schedule). Treat 2xx as success, 5xx and timeouts as transient, and 4xx (except 429) as permanent failure. Include X-Webhook-Retry-Count so consumers can log it.

As a consumer: verify the signature, enqueue the event for asynchronous processing, and return 200/202 immediately. Never do heavy work in the HTTP handler — the sender will time out and retry, amplifying load.

Dead-letter queues

Events that exhaust the retry schedule move to a DLQ with full context: payload, headers, last error, attempt count. Provide operator tooling to inspect and replay. Alert when the DLQ grows unexpectedly. Retain DLQ entries for at least 30 days so you can recover from outages discovered after the fact.

Security

• TLS only. Never accept webhooks over plain HTTP.
• Always verify HMAC signatures, even if the source IP is allowlisted.
• Restrict by sender IP ranges when the source publishes them (Stripe, GitHub, etc.).
• Enforce a payload size limit (1 MB is typical) and rate-limit the consumer endpoint.
• Store secrets in a vault (AWS Secrets Manager, Vault, etc.), not in code or environment variables checked into source control.

Gotchas

Agent-specific failure modes — provider-neutral pause-and-self-check items:

• Verifying the HMAC signature against parsed and re-serialized JSON instead of the raw body. Re-serializing parsed JSON reorders keys, collapses whitespace, and changes the byte sequence — the computed HMAC will not match the sender's signature. Always compute the HMAC against the raw request body bytes, before any parsing.
• Non-constant-time signature comparison. Using == to compare HMAC values allows a timing side-channel: an attacker can observe that the comparison returns faster when fewer bytes match, and iteratively guess the correct signature. Always use a constant-time comparison function such as hmac.compare_digest.
• No idempotency check — processing the same event twice. Webhook delivery is at-least-once. A network timeout or a 5xx response from your endpoint will cause the sender to retry, delivering the same event multiple times. A consumer that sends an email, charges a payment, or creates a database record without deduplicating by event ID will do so multiple times. Store processed event IDs and check before processing.
• Doing heavy processing in the HTTP handler instead of enqueuing. A handler that makes database queries, calls external APIs, or runs business logic before returning will frequently exceed the sender's request timeout (typically 5–30 seconds), causing the sender to retry — amplifying the work. Return 200 or 202 immediately after signature verification; enqueue the event payload for asynchronous processing.
• No replay-window check on the timestamp. A webhook signature protects the payload in transit but does not prevent replay attacks — a captured request can be replayed weeks later with a valid signature. Include a timestamp in the signed payload (or in a signed header), and reject events older than ~5 minutes to bound the replay window.
• Single signing secret with no rotation path. When a secret is leaked or an employee leaves, the secret must be rotated. If the consumer only accepts one active secret, rotation requires simultaneous updates to both the sender and consumer — a coordination window where webhooks either fail or are unverified. Support multiple active secrets so old events can still be verified during rotation.
• Returning 200 from a handler that errored internally. A handler that catches its own internal error and returns 200 tells the sender "delivery succeeded" — the sender will not retry, and the event is lost. Return 500 (after signature verification) so the sender retries. Reserve 4xx for cases where retrying will never succeed (malformed event, unsupported event type).

Full reference

Ordering

Webhooks do not guarantee delivery order. A consumer that assumes "shipped" must arrive after "created" will eventually be wrong. Two options:

1. Use created_at to detect out-of-order events and ignore the stale one.
2. Embed a monotonic version or sequence number on stateful resources, then discard any event whose version is older than what's already applied.

Both work; pick the one that fits the resource model.

Testing webhooks locally

• Tunnels: ngrok and cloudflared tunnel expose a local endpoint to a public URL so the sender can reach it.
• Inspection: webhook.site and requestbin capture raw requests for offline debugging when the sender doesn't expose a test mode.
• Automated tests must cover: invalid signature (reject), valid signature (accept), duplicate event ID (idempotent — process once), out-of-order events, malformed JSON, oversized payload, expired timestamp.

Failure modes worth designing for

The fix for the first two is the same on every platform: dedupe by event ID with a unique constraint, and move processing to a background queue so the HTTP handler returns in milliseconds.

Webhook registration API (when designing your own)

If you're publishing webhooks rather than consuming them, expose a simple registration API:

• POST /webhooks — register a url, an events[] filter, and receive a generated signing secret in the response (shown once).
• GET /webhooks/{id}/deliveries — recent delivery attempts with status, latency, and response body for debugging.
• POST /webhooks/{id}/deliveries/{delivery_id}/replay — manual replay for ops recovery.
• Admin dashboard with delivery success rate, latency percentiles, and DLQ depth per subscription.

Anti-patterns to avoid

• Synchronous processing in the HTTP handler. Always enqueue and return immediately.
• Verifying the signature against parsed JSON. Sign and verify the raw bytes the sender produced.
• == for signature comparison. Use a constant-time function.
• No idempotency store. "It probably won't happen twice" is wrong on the day it matters.
• Returning 200 from a handler that errored internally. The sender will not retry. Return 500 so retry happens, after the signature has been verified.
• No replay-window check. A leaked recording can be replayed weeks later.
• Single signing secret with no rotation path. Plan rotation before you need it.