Designing integrations for the failure case
The happy path is the easy part. Retries, idempotency, dead-letter queues and ownership are what decide whether an integration survives its first bad night.
By IT NET Solution
An integration that moves data correctly on a quiet Tuesday is not finished work. It is a demonstration. Production is a different environment: partner APIs time out, ERP maintenance windows land on the same evening as a campaign, and one message arrives twice.
Decide what duplication means before it happens
Most retry strategies assume a delivery can be repeated safely. That assumption only holds if every message carries a business key and the receiving side can recognise work it has already done. Without that, a retry is an instruction to create a second order.
- Give every message a stable business key, not a fresh identifier per attempt.
- Make the receiving operation idempotent, or record what was applied and when.
- Write the outcome before acknowledging, so a crash cannot replay work silently.
- Prefer at-least-once delivery with a de-duplication step over fragile exactly-once claims.
Partial success is the normal state
In a multi-step flow — validate, reserve stock, create the order, notify the warehouse — any step can fail while the others have already committed. The design question is what the whole flow should mean at that moment. Either the steps are compensatable, or the flow is designed to resume from a recorded state.
Failure needs somewhere to go
Messages that cannot be processed should land in a queue with their payload and the reason, where an operator can correct and replay them. A failure that disappears into a log line is a failure nobody can act on at nine in the evening.
None of this is exotic. It is ordinary engineering applied to the boundary between systems — which is exactly where businesses tend to accept less rigour than they would inside a single application.
Topics
- Integration architecture
- Reliability
- Operations
Related reading
Decide master data ownership before you connect systems
Most integration conflicts are ownership conflicts. A short method for assigning the system of record for customers, products, prices and stock before the first connector is built.
Order-to-invoice flow across commerce, ERP and logistics
An illustrative integration programme: nightly exports and manual reconciliation replaced by a monitored, queue-based flow between a commerce platform, SAP Business One and a logistics provider.
Designing a CRM your sales team will actually use
Most CRM problems are design problems, not software problems. A practical view of module structure, pipeline logic and the fields that decide whether adoption holds.
Working on something similar?
Describe the systems, the constraints and what needs to change. We will outline a realistic path and the trade-offs behind it.