How to Maintain Integrations Over Time
The first sign that an integration has stopped may be a customer asking why a payment was recorded twice. The monitoring page is green, the original developer left months ago, and nobody remembers which system owns the retry. An integration can keep running for years without attention, until a small change reveals that “it works” was never an operating plan.
An integration moves or synchronizes information between systems, often through an application programming interface, file exchange, webhook, or managed connector. Maintaining it means treating the connection as a service with an owner, expected behavior, observable health, recovery steps, and a budget. You do not need an enterprise monitoring platform to begin. You do need to know what should happen when a record does not arrive.
Build an inventory that answers operational questions
Start with a list of every connection you can find. Include the systems at both ends, the business purpose, data exchanged, direction, trigger, frequency, volume, technical method, authentication, environment, support contact, and last known review. Add the source repository or configuration location, deployment process, current version, and any third party service in the path.
Record a business owner and a technical owner for each integration. The business owner decides whether the data and timing are still useful. The technical owner can inspect, change, test, and restore the connection. In a small company, one person may hold both roles, but name a backup who can find the documentation and contact the vendor. “IT” or “the contractor” is not a useful owner when an incident needs a decision.
Do not inventory only what is approved. Ask finance, sales, operations, and customer support which reports or manual exports they rely on. Look for scheduled spreadsheets, personal access tokens, vendor created recipes, email rules, and scripts that run from a laptop. An undocumented file transfer can be more operationally important than a formally managed interface.
Capture the consequence of failure, not just a technical severity. Does a missed update delay a shipment, create a duplicate invoice, or merely postpone a report by an hour? Note how much data can be safely replayed, who notices a gap, and whether a late delivery is acceptable. This context determines the right alert and recovery objective.
Define what healthy means
A successful HTTP response does not prove that the right business result happened. A request may have been accepted but later rejected. A job may finish while sending zero records because its cursor or filter is wrong. Define health using both technical signals and business outcomes.
For each connection, document expected volume or a reasonable range, maximum age of the latest successful record, acceptable error rate, and the outcome to verify. If a nightly process usually moves around 400 orders, an execution with a success status and no orders deserves attention. Avoid treating every natural volume change as an outage. A low volume alert should consider business hours, seasonality, holidays, and known pauses.
Track a small set of useful signals. These may include run start and finish, duration, records read and written, records rejected, retry count, oldest pending item, and last successful business transaction. Keep correlation identifiers that let an operator follow one transaction through both systems. Do not put passwords, access tokens, payment data, or unnecessary personal information in logs.
An integration can be healthy at several layers. The source system may respond, the connection may authenticate, the payload may pass schema checks, the destination may accept it, and the business record may appear correctly. Decide which layers matter for each flow and where evidence can be observed. A daily reconciliation can catch missing outcomes that an immediate request alert will not.
Design alerts for action, not noise
An alert should explain what failed, what was affected, when the issue began, and what the recipient can do next. Route it to a named person or monitored team channel. Include a link to the runbook and relevant log search, not a full copy of sensitive payloads. Escalate unresolved critical failures through an agreed path.
Separate an isolated transient error from a sustained problem. A brief network timeout may succeed on a later retry. A revoked credential, repeated validation failure, or growing queue needs attention. Use thresholds that reflect business impact, such as a critical queue older than 20 minutes or a daily financial reconciliation missing a batch, rather than paging staff for every first attempt.
Group related failures. If one vendor outage causes a thousand retries, send one actionable incident with the scale and latest status instead of a thousand messages. Have an owner mute or adjust an alert only with a documented reason and an expiry or review date. Permanent silence is not resolution.
Review alerts after incidents. Did the signal arrive early enough? Was it routed to someone able to act? Did it lead to a clear recovery step? Remove alerts that repeatedly produce no action and create better ones for the actual failure. A monthly check of alert volume can reveal both blind spots and notification fatigue.
Rotate credentials without interrupting the flow
Integrations commonly use API keys, OAuth access tokens, certificates, or service accounts. A credential rotation replaces an expiring or potentially exposed secret with a new one. The safest procedure is usually to create and test a replacement before disabling the old credential, when the vendor permits overlap.
First identify where the credential is stored and which deployments use it. Store secrets in a managed secret store or the platform's protected environment configuration, not source code, a ticket, a wiki, or an employee's personal notes. Limit access to the people and services that need it. Keep separate credentials for development, test, and production when possible.
Write the rotation steps for the exact vendor and integration. Confirm who can issue a replacement, how long it takes to take effect, whether a test environment exists, and how to roll back. Deploy the new value to a test environment, verify a read and a representative write, then update production. Check logs and business outcomes after the change. Revoke the old credential only after the new one is confirmed.
Illustrative scenario. A billing connector uses a key that expires in 30 days. The runbook says the business owner requests a replacement from the vendor, the technical owner stores it in the production secret manager, and a test invoice is synchronized in the sandbox. During a scheduled window, the team switches production, watches the next three batches, and then revokes the old key. If the vendor allows no overlap, the team arranges a short change window and a manual billing fallback. This is a hypothetical procedure and should be adapted to the actual service.
Set a rotation calendar from the credential's real expiration and the vendor's process. Avoid arbitrary frequent changes that nobody can execute reliably, but do not wait until the last day. Treat a suspected exposure differently from routine expiry. Revoke compromised credentials promptly, assess access logs, and follow your incident process.
Prepare for API deprecations and behavior changes
An API deprecation means a vendor has announced that a version or feature will eventually be removed or changed. Notices may arrive through developer email, a dashboard, release notes, or a partner. Assign someone to monitor those channels and record affected integrations, dates, and required work.
Maintain a dependency record for each interface. Include API version, endpoint or event type, software library, authentication method, and vendor documentation link. When a notice arrives, determine whether your use actually depends on the changed behavior. A version number alone does not tell you whether the integration is affected.
Test changes against representative data and failure cases in a nonproduction environment. Compare counts, key fields, timestamps, pagination, rate limits, and error responses. Watch for an API that changes a field from optional to required, alters how empty values are represented, or changes webhook retry behavior. Schema validation can detect some changes before they create bad records.
Agree on a deadline with enough room for testing, release, and rollback. Vendors sometimes extend removal dates, but do not make an extension your plan. Keep a compatibility layer or small adapter where it simplifies future changes, but avoid abstracting every vendor detail into a framework the team cannot maintain.
Write a runbook a tired person can use
A runbook is a short operating guide for routine checks and incident recovery. It should begin with the service purpose and the impact of a failure. Include owners and backups, dashboards, logs, dependencies, expected timing, alert meaning, safe replay instructions, escalation contacts, and steps to disable the connection without causing duplicate or lost transactions.
Write recovery as a sequence of checks. Confirm the source event exists. Check whether it was received. Search for the correlation identifier at the destination. Determine whether the record was rejected, delayed, or applied twice. Only then choose retry, repair, or manual entry. State what must be reconciled after recovery. A “restart the job” instruction can multiply a duplicate if the operation is not idempotent.
Idempotency means that processing the same request more than once has the same intended effect as processing it once. For example, an order creation request can carry a stable identifier so retries update or return the existing order instead of creating another. If the integration does not have this protection, document the duplicate check and approval needed before replaying data.
Keep operational instructions separate from secrets. A runbook can say where to request or retrieve an approved credential, but it should not contain the credential itself. Review access to the documentation and avoid including customer details in screenshots. Add the date and reviewer so an operator can see whether the instructions have gone stale.
Assign ownership and a maintenance rhythm
Set a light review schedule based on risk. A payment or order connection may need weekly health review and a quarterly recovery exercise. A noncritical reporting export may need only a monthly check. At minimum, revisit ownership, credentials, vendor notices, alert quality, runbook accuracy, and the cost of the connector during a regular operations review.
Use a change process that is proportional to impact. Record why a change is needed, what data and workflows it touches, how it will be tested, who approves release, and how it will be reversed. Coordinate changes to either endpoint. A destination team that renames a field can break a stable integration even if the integration code has not changed.
Plan for departures. Ensure at least two people can access the repositories, deployment process, monitoring, and vendor support path. Remove departing staff access promptly, but transfer ownership of service accounts and vendor contacts before accounts are closed. A system that depends on one person's login is a continuity risk, not merely a documentation gap.
Choose maintenance by business impact
Not every connection deserves a custom monitoring stack. A managed platform may be appropriate when the workflow is standard, the connector is maintained by a reputable provider, and the retry, history, and alert controls are sufficient. Custom code may be sensible when the process has important business rules, specialized validation, or a need for clear audit and replay behavior.
Compare the complete operating burden. Ask who sees failures, how long run history is retained, whether errors can be corrected and replayed safely, what usage limits apply, and how a vendor support request is handled. Ask what happens if the connector provider closes or changes pricing. For custom code, include developer availability, hosting, tests, and version updates. A lower initial fee can produce a higher cost of ownership if the team has no capacity to maintain it.
Start the audit on Monday morning
Pick the integration most likely to cause a financial, customer, or operational problem. Name its business and technical owners. Follow one recent record from the source to the destination and note where you can prove each step. Check when the last successful business result occurred, not only when a job last ran.
Then document one recovery path and test it safely with a nonproduction record. Review the credential's owner and expiry, find any vendor change notices, and send a test alert to the person expected to respond. Repeat for the next highest impact connection. Small, verified improvements are more valuable than a long inventory no one maintains.
Integrations are part of daily operations, not plumbing that can be forgotten after launch. If ownership, alerts, and recovery are clear, routine change becomes manageable. When a set of inherited connections has no obvious owner or safe replay path, Ozcorp InfraOps can help assess the operating risks and organize maintenance around the work your team depends on.
Discuss your infrastructure requirements
Review the current service offer or schedule a conversation about the scope you need.
Stay in touch
Follow Ozcorp
Leave your details for occasional notes about new articles and Ozcorp updates.