01The network is not reliable, and neither is the vendor
Every integration eventually meets a timeout, a rate limit or a schema change. Code written for the happy path will lose data the first time one of those happens at 2am.
Durable integrations assume failure: retry with backoff, keep a dead-letter queue, and make every operation safe to repeat.
02Idempotency is the cheapest insurance you can buy
Give every operation a key derived from the business event, and make repeating it a no-op. This one decision removes an entire category of duplicate orders, double charges and repeated emails.
03Reconcile on a schedule
Event streams drift. A nightly job that compares both sides and reports differences will catch the problems your monitoring misses — usually before finance or a customer does.
04Monitor per integration, not per server
Volume, latency and failure rate for each connection, with alerts on anomaly rather than on absolute thresholds. When a partner silently changes behaviour, the shape of the traffic tells you first.