The Six Places Integrations Actually Fail

Credentials expire at the worst time

API keys rotate, OAuth tokens expire, certificates lapse. The integration stops, often quietly, and is noticed days later.

Monitor credential expiry as a first-class alert with thirty days' notice, and handle token refresh properly rather than assuming it works.

Rate limits bite at peak

You will never hit a rate limit during testing. You will hit it on the busiest morning of the month, which is also when nobody has time to investigate.

Queue, back off and retry rather than failing. And know the limit before designing the batch sizes.

Upstream formats change without warning

• A supplier redesigns their invoice template
• A vendor adds a field and reorders the export
• A date format changes from one convention to another
• An encoding changes and accented characters break

Validate structure on every input and alert on anything unexpected rather than processing it optimistically.

Duplicate writes

A retry after a timeout that actually succeeded creates two records. It is the most common data corruption in integrations and it is entirely preventable with an idempotency key.

Derive the key from the source document so the same input can never produce two outputs.

The last two

1. Index staleness — answers drawn from documents that changed weeks ago
2. Cost creep — prompts growing gradually until the bill has doubled

Both are invisible without monitoring and both are trivial to catch with it.

Frequently asked questions