Integration projects are usually sold on connector count. They succeed or fail on what happens when a connection drops at 2am, when two systems disagree about a customer, or when a vendor changes an API without warning. The architecture you buy should make those events ordinary and recoverable, not invisible until a month-end report fails.
Write requirements before comparing platforms
- Every source and target system, with a named business owner for each.
- Direction of flow and the system of record for customers, prices, stock and payments.
- Average and peak volume, including seasonal or month-end spikes.
- Latency tolerance: real time, near real time, hourly or overnight.
- Whether ordering matters, and whether duplicates are tolerable.
- How far back a failed message must be replayable.
- Data classification and any fields that must not leave a system.
- Who receives an error, who fixes it and within what service window.
- How often schemas change and who approves a breaking change.
Five patterns you will be offered
Point-to-point connections
Each pair of systems gets its own integration. This is fast for two systems and becomes a maintenance maze at six. It suits a narrow, stable scope with a single owner.
Hub-and-spoke through an integration platform
A central platform maps, routes and monitors messages. It reduces duplicate connectors and creates one place to see failures. The trade-off is another platform to license, secure and operate.
Event-driven messaging
Systems publish events and consumers react, which decouples teams and handles bursts well. It demands discipline around event contracts, ordering, idempotency and replay, or the business ends up with silent data loss.
API gateway with managed services
A gateway handles authentication, rate limits, logging and routing while each service owns its data. This is a strong fit for custom applications, but it requires engineering maturity and a clear API ownership model.
Change data capture and pipelines
Changes are read from a database log and delivered to a warehouse or another system. It is efficient for analytics and bulk synchronisation, but it can bypass business rules and needs careful handling of deletes and schema changes.
Non-negotiables in the design
- Idempotent writes, so a retry cannot create a duplicate order or payment.
- Retries with backoff and a maximum attempt count, never an infinite loop.
- A dead-letter queue with a visible owner and a replay action.
- Schema versioning and a documented compatibility rule.
- Contract tests that run before either side deploys.
- Correlation IDs across systems, so one transaction can be traced end to end.
- Rate-limit handling that queues or slows down instead of failing silently.
- Secrets in a managed store, rotated on a schedule.
- A written field map with transformations, defaults and null rules.
A scorecard for the shortlist
- Connector depth (20%): not the number of connectors, but whether the ones you need support the objects, custom fields and incremental sync.
- Failure handling (20%): dead-letter view, replay, alerting and a documented recovery procedure.
- Transformation and mapping (15%): can your team change a map without a vendor change request?
- Observability (10%): per-message status, latency, error trends and exportable logs.
- Security and residency (10%): authentication options, encryption, subprocessors and data location.
- Environments (10%): a real sandbox with test credentials and a safe promotion path.
- Support and SLA (10%): response times, escalation and who owns an outage.
- Exit (5%): export of mappings, logs and configuration, plus transition assistance.
Pricing questions that change the total
- Is the fee per connection, per environment, per message or per data volume?
- What counts as a message, and how are retries and replays billed?
- Are sandbox and staging environments included?
- What happens when volume exceeds the plan: overage, throttling or a forced upgrade?
- Are monitoring, alerting and log retention included or separate?
- What does implementation, mapping and training cost, and who does the work?
- What is the termination notice and the exit assistance fee?
Pilot with failure injection
Choose two integrations that matter, including one write path. Run them for 30 days with normal traffic and deliberately break things: revoke a credential, return a 500, send a duplicate message, delay a response beyond the timeout and change a field name in the sandbox. Acceptance should be measured, not demonstrated: no duplicate records, every failure visible within the agreed window, a successful replay and a reconciliation report that matches both sides. A vendor that will not support failure injection has not operated integrations at scale.
Contract and exit terms
- Named data owner and a written prohibition on secondary use.
- Data processing terms, subprocessor list and breach notification window.
- Service credits tied to measurable availability and support response.
- Advance notice of breaking API or platform changes.
- Export of mappings, transformation logic, logs and configuration.
- A parallel-run period and transition assistance at contract end.
Proposal warning signs
- A connector count presented as an architecture.
- No sandbox, or a sandbox that only works with clean demo data.
- Pricing per record with no cap, forecast or alerting.
- Errors visible only in a support ticket.
- Custom mapping locked behind professional services.
- No answer for duplicate delivery or replay after an outage.
How we build and monitor integrations
We design and build integrations through Custom Software & Integrations, including contracts, mapping, retries and monitoring. How an integration engagement is sized and priced appears on the pricing page. For the design conventions that pair with this guide, read the REST API design checklist.
Sources
- OpenAPI Specification, accessed 17 November 2026.
- RFC 9457, Problem Details for HTTP APIs, accessed 17 November 2026.
- OWASP, API Security Top 10, accessed 17 November 2026.
- AsyncAPI, documentation, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.