Integration isn't just moving data between systems

Most integrations that fail in production were never wrong about the endpoint. They were wrong about ownership, ordering, or what should happen the third time a request times out.

Four things to decide before writing the connection

1. 01Ownership — which system holds the authoritative version of each field.
2. 02Direction — one-way sync unless bidirectional is genuinely required.
3. 03Trigger — event-driven where possible, scheduled reconciliation as a safety net.
4. 04Failure — retries, idempotency, and a visible place where problems surface.

Events over polling

Events describe what changed and when. Polling only tells you what a record looks like now, which makes ordering and cause impossible to reconstruct later.

{
  "event": "subscription.activated",
  "occurredAt": "2026-08-06T09:14:22Z",
  "customerRef": "cus_1042",
  "source": "commerce",
  "idempotencyKey": "sub_1042_activated_1"
}

A payload like this can be replayed safely, audited, and consumed by more than one downstream layer without changing the producer.

Make it observable

• Log the event, not just the result.
• Expose a queue or table of failures a human can actually read.
• Reconcile on a schedule so silent drift can't accumulate.