Integrations that don't break: engineering the glue, not just the call

Integrations fail in boring, predictable ways: a webhook that never arrived, a payload field that changed shape, a rate limit hit during your busiest hour. The systems that survive treat every external call as if it will fail — because eventually it will.

Four patterns carry the load. Idempotency: attach a request ID so a retried payment or webhook can't double-charge or double-count. Retries with backoff and a dead-letter queue: failures get stored, not lost, and a human inspects them. Contract tests: when a vendor changes a payload, your test suite — not your customers — finds out. And reconciliation jobs that compare your records against the vendor's daily, catching silent drift.

Webhooks specifically deserve distrust: they're dropped, duplicated and delivered out of order. The robust pattern is treating webhooks as hints and pulling the authoritative state on receipt — a small API cost that eliminates a whole class of missing-event bugs.

Finally, observability: every integration gets structured logs, a latency metric and an alert when its error rate moves. When Razorpay changes something at 11 a.m. on your biggest sale day, the difference between a two-minute fix and a two-day apology is whether your monitoring spoke first.