Software Development and Architecture: Systems That Survive the Next Change

Before a single line exists, the business side is sorted: which terms mean what in which domain, who owns which data, which rules are law and which are habit that grew over time. Out of that sorting come the borders of the later system. The result is an agreed model of terms and a list of the places where two departments use the same word differently — the most common source of special cases later on.

The target design settles which building blocks the system consists of, which block carries which responsibility and how they talk to one another. That includes an honest justification of the granularity: for many houses a monolith with clear internal borders is the more workable answer than distributed services, which presuppose an operations crew that does not exist. What is written down is not diagrams but decisions, together with alternatives and reasons.

This is the layer where building happens: line-of-business applications, backend services, web interfaces, reporting logic, automation. External capacity works inside the house’s way of working and not beside it — same repositories, same review steps, same release paths. Where a technology is new to the house, bringing your own people up to speed is part of the brief, not of a follow-up phase.

Hardly any application stands alone. Here it is settled which system leads for which record, which interfaces have to answer synchronously and which work better event-driven, and how version changes stay possible without touching every connection at once. Contracts over interfaces are made explicit and are checked — not out of formalism, but because they are the place where changes otherwise become most expensive.

A test suite is only one if a red run means something. That includes tests on the right layer rather than as many as possible, checks on dependencies and their licences, catching known vulnerabilities during the build, and the question of how faults become visible in operation before users report them. Security in the application code belongs here; the organisational level of protection is covered in cyber security consulting.

From its first day in operation an application is a standing cost. This layer is about reproducible releases, about observability in the running system, and about who actually develops the application further once the external mandate ends. A handover that consists of documentation is not one — handed over is what your own team has changed without asking back.

Starting point: a process runs on spreadsheets and shouted instructions, and an off-the-shelf solution does not fit the particularity that makes the business. What gets built first is the core that carries that particularity; everything else is bought in or connected. Touchstone: the cycle time of the supported process and the share of manual rework — not the number of requirements delivered.

Starting point: a load-bearing application is technically at its end but indispensable to the business and nowhere fully described. Replacement happens in sections with parallel operation, starting where the way back is shortest. Touchstone: the share of the legacy system that is demonstrably switched off — a system replaced to 90 percent costs as much as a whole one.

Starting point: a design exists — from your own house or from a vendor — and an independent assessment is missing before budget is committed. Reviewed are the business boundaries, the dependencies, the operating assumptions and the question of what the design costs in the event of a reversal. Touchstone: named risks with the condition under which they occur, and a statement on how reversible each load-bearing decision is.

Starting point: the target date is fixed, the team carries day-to-day operations alongside, and the profiles wanted cannot be hired permanently at short notice. Reinforcement happens inside the existing way of working, not as a second team beside it. Touchstone: the release frequency after people join — if it does not rise within a few weeks, capacity was not the bottleneck.