Every mature codebase I have ever been handed has a tech tree drawn on some wiki page: the architecture diagram, boxes and arrows, each service in its place with its dependencies listed like prerequisites in a strategy game. And every time, the diagram has been lying in exactly the way I have accused the game of lying — not in its components, but in its shape. The boxes are real. The claim that the system is *made of* the boxes is the fiction. [author assertion]
What the system is actually made of is complements that nobody drew. The payments service "works" because one engineer remembers why the retry logic has a forty-second pause in it. The search index stays honest because a cron job a departed contractor wrote in 2019 is still running on a machine nobody inventories. The deployment pipeline unlocks new features because two teams have an unwritten agreement about Friday afternoons. Remove any two of these and the famous architecture fizzles into a curiosity — boxes, correctly labelled, through which nothing flows. [author assertion]
Software people rediscover forgetting faster than civilisations do. A library that nobody on the team has touched in two years is not *owned*; it is a rumour with a version number. Knowledge that is not continuously re-enacted decays, and codebases re-enact their knowledge only in the act of changing the code — which is why the components everyone is afraid to touch are precisely the ones the organisation has quietly forgotten how to think about. We call this "legacy" and treat it as an age problem. It is a practice problem. The code is the same age as everything else; it has simply stopped being practised. [author assertion]
The institutional layer binds here too. An elegant internal tool sits idle for years in one firm and transforms another in a quarter, and the difference is rarely the tool: it is who may mandate it, whose budget absorbs the migration, which team's status is threatened by its adoption, which guild is allowed to object. I have watched a superior build system lose to an inferior one because the inferior one's owner ran the onboarding. The scarce resource is seldom the technology. It is the surrounding tissue that lets the technology be used, financed, repaired and taught — and in software, the tissue is other engineers' habits, which no diagram has ever once depicted. [author assertion]
To be fair to the diagram, it gets two deep things right, as the game does. The first is opportunity cost: a team's build capacity is scarce, and choosing a line means not choosing another, with consequences that surface quarters later. The second is path dependence: early choices genuinely foreclose later ones, and a codebase that optimised for the next unlock and one that invested in complements are different bets whose grades arrive slowly. [author assertion]
If I were redesigning the architecture review, I would draw the complements web instead: for each service, the practices, people, agreements and rituals without which the box is decorative — with decay marked where a practice has lapsed. It would be a worse diagram and a better model, and it would make a specific kind of engineer uncomfortable, because some of the most load-bearing boxes on it would have people's names where the technology should be. [author assertion]
So my question, for anyone who maintains a system older than its diagram: which box on your architecture page is one resignation — or one lapsed habit — away from being a museum piece? Name the complement. I will go first in the replies, and mine involves a cron job. [author assertion]
