Webhook Delivery Service · View 17 of 20 · Operations
The split that settles arguments
- Platform time and consumer time are measured and reported separately at every stage. Conflating them makes the platform the suspect in every slow integration (ADR-13).
- The bottom row is not internal telemetry — it is the product. Backlog depth, delivery health and the request inspector are what let a developer diagnose their own integration without a support ticket.
- Retry share is a first-class metric because a small population of failing endpoints can consume most of the delivery capacity while producing no successful deliveries.
Alert thresholds, all assumptions
- Fan-out lag above 60 s, backlog older than 1 hour, retries above 25% of attempts, and any endpoint auto-disabled.
- Attempt records retained 90 days at full fidelity for the console; the response body is stored as a bounded 1 KB prefix only.
Gaps
- The empty cell is honest: the platform has no trace inside a consumer's system and never will. Everything after the POST leaves the estate is inferred from a status code and a duration.
- A 2xx returned by a consumer that then failed internally is invisible to every signal on this page. Only the consumer's own idempotency and the availability of replay address it.