Knowledge · Decision guide

Integration becomes dependable in everyday operation.

“The API is available” does not tell you whether your orders were processed. Reliable integration covers normal operation, interruptions and recovery to a consistent state. These questions help you plan a connection that the team can understand and act on when something goes wrong.

Practical guidance

Define success in business terms.

A sent message, a technical acknowledgement and an order created in the ERP are different states. Agree which event completes the business transaction. For an order, that might be an unambiguously linked ERP order number; for a stock update, confirmation of processing in the destination. Check what evidence the systems actually provide.

Each transaction needs a reference that can be followed across systems. It should show when processing started, which step last succeeded and whether action is required. Store the information needed for this purpose and agree retention and access. Logging every payload indefinitely does not automatically make the process easier to understand.

From need to solution

Different failures need different responses.

01

Temporary interruption

A destination is briefly unavailable. A limited retry after a delay may help, provided it cannot create a second business transaction. Once the agreed limit is reached, the issue should be visibly escalated.

02

Invalid business data

A product is missing, a unit is unknown or required information is absent. Retrying blindly will not fix it. The responsible owner needs to correct the record before it is processed again.

03

Uncertain processing state

The connection fails after sending. The order may still have been created. Before retrying, check its state through the reference or use a dependable mechanism that prevents duplicate effects.

04

Permanent system change

Access, field formats or business rules have changed. Correcting one record may not be enough. The integration needs adjustment and representative tests, with affected records visibly held for later processing.

Practical guidance

Retries must not create duplicate orders.

The principle is called idempotency: a business transaction can be triggered again without applying its effect twice. It requires a stable key and explicit treatment of transactions already processed. An external order number may help, but it must be unique within the agreed scope. Two stores must not accidentally use matching numbers for different orders.

Some interfaces support suitable mechanisms; others need additional state tracking and reconciliation. The implementation depends on the systems. Pay particular attention to the moment when creation succeeds at the destination but the response is lost. A “sent” flag at the source proves neither successful receipt nor that a retry would be safe.

Practical guidance

Monitor unfinished work as well as servers.

Technical availability matters. Operations also need the age and number of pending transactions, rejected records and unusual differences. Missing expected activity can itself be a signal. Acceptable delay depends on the data: product descriptions and stock used to make sales commitments do not necessarily have the same urgency.

Each meaningful alert needs an owner and an expected response. “Error 422” alone is not actionable. A useful message identifies the transaction, explains the reason, states the last known status and gives the next step. Group related failures so one outage does not generate hundreds of identical messages. Agree escalation and cover before launch.

Practical guidance

Plan what happens after an outage.

When a system returns, the backlog cannot always be sent in any order. Creating a customer may be a prerequisite for creating an order; an old stock update must not overwrite a newer state. Rules for ordering, freshness and obsolete transactions are part of the integration, not an afterthought.

Also define how manual interventions are recorded. If an order was entered manually during the outage, automated recovery must be able to recognise it. Finish recovery with reconciliation: which transactions completed, which remain open and which need a business decision? This prevents residual errors being forgotten once the technical incident ends.

Practical guidance

Include these scenarios in acceptance checks.

Alongside a normal order, test at least a duplicate message, temporary unavailability, an invalid record and a lost response. Add a business-specific exception from your daily work. Record the expected state, visible feedback and available corrective action for each. A test is useful only if its outcome can be evaluated clearly.

This also makes operational support concrete: which flows are monitored, who responds when and which corrections are included? Availability is not a blanket promise. A traceable integration with clear responsibilities provides the foundation for diagnosing problems and resuming work reliably.

The right next step

Apply this to your business.

Tell us where you are today and what you want to improve. We will identify the next useful step together.