Hire Integration Engineers

Integration is where enterprise programmes fail quietly. Nothing errors. The same order is simply counted twice, or a status update lands out of order and a shipment shows as delivered before it was picked. Both are discovered in a reconciliation weeks later rather than in a log the same afternoon.

That shapes the work: contracts agreed before code, idempotency treated as a requirement rather than a defensive measure, ordering guarantees stated rather than assumed, and a dead-letter path a human can actually work through. The tooling — MuleSoft, Kafka, an API gateway, a middleware somebody chose in 2016 — matters less than whether the person has had to explain a reconciliation break to finance.

Idempotency in this context means that processing the same message twice leaves the system in the same state as processing it once, and tells the caller the same thing both times. It is a design property agreed before the flow is built. It cannot be added convincingly afterwards, because by then there is production data that may or may not already be doubled and no reliable way to tell which.

Ordering is the guarantee most often assumed and least often stated. Kafka orders events within a partition and promises nothing across partitions, which makes the partition key a business decision rather than an infrastructure one. If events for a single order can land on three partitions, a cancellation can be applied before the booking it cancels, and the bug will be intermittent enough to survive several rounds of testing.

The layering question is the equivalent argument on the API side. An API-led approach only pays for itself if the system API genuinely hides the system behind it, and deadline pressure reliably produces experience APIs that reach past the layer to a database. The test is uncomfortable and quick: could the system behind that API be replaced without any consumer changing? Where the answer is no, the organisation is paying for three deployments and getting one integration.

Search and indexing sit in this group for the same reason. Elasticsearch is usually fed by the same pipelines and fails in the same way — quietly, with an analyser splitting part numbers on hyphens, until a customer searches for one and gets nothing. The common thread across all of it is that these systems do not error when they are wrong, so somebody has to design the check that notices.