What does maintaining custom software actually involve after it launches?
Three different kinds of work. Keep-alive: security patches, dependency updates, certificate and credential rotation. Adaptive: reacting when a vendor changes an API, a login flow, or a data format on their schedule, not yours. Growth: changes your own business asks for. Software that nobody touches still decays, because everything it depends on keeps moving. Budget for adaptive work annually even if you plan no new features.
Software rots even when the code is frozen
A finished application is not a finished product, because it sits on a stack that keeps changing underneath it. Language runtimes reach end of life. Libraries publish security advisories. Cloud providers deprecate services. Browsers change how they handle cookies and redirects. An OAuth token expires on a schedule someone forgot to write down.
None of this is caused by you and none of it is optional. This is the single most common surprise for owners who commissioned a build and assumed the relationship ended at launch.
The three buckets, and why they should be separated
- Keep-alive. Patching, upgrades, backups, monitoring, credential rotation. Predictable, boring, non-negotiable.
- Adaptive. Work forced by an outside party — an API version sunset, a new required scope, a changed field name, a platform policy update. Unpredictable in timing, very predictable in existence.
- Growth. New capability you asked for. This is the only bucket that should feel like a project.
- Incident. Something broke. Response time here is a business decision, not a technical one: decide in advance how long the tool can be down before it costs you money.
Integrations are the part that ages fastest
Anything talking to a third-party system carries that system's release schedule. Field service platforms, ad platforms and phone systems all evolve, and their changes arrive as deprecation notices in an inbox nobody checks. Assume the integration surface will need attention at least once a year per connected system.
The practical defense is an inventory: every connection, what it reads and writes, which account owns the credential, when it expires, and where the deprecation notices go. Teams that keep this list handle vendor changes as maintenance. Teams that do not handle them as outages. It is a standing part of how connected systems are run, not an afterthought.
Monitoring is maintenance
The dangerous failure in operational software is not a crash. It is silence — a sync that stops importing and produces a report that looks plausible and is quietly missing a day. Alarms should fire on the absence of expected activity, not just on errors.
That means a check that says "we expected roughly this many records in the last hour and got none." It is cheap to add and it is what keeps operational systems trustworthy over years rather than months.
Topics: maintenance · support · dependencies · operations
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.