Selected writing · Software architecture
A shared language for systems.
Data librarians inherit systems they did not design. The 2025 deck gives them the vocabulary to ask structural questions — and deliberately stops before code.
Why it exists
A repository is easier to operate than to explain.
The Data Librarian programme at Graz trains people to run collections, not to draw diagrams. But the questions they meet are architectural: why this import is slow, why one change breaks three services, why no one can say what depends on what.
The deck starts from documentation rather than theory: Fedora’s feature list and architecture page, Islandora’s diagram, the OAIS reference model. Established systems explain themselves in a shared language; the workshop borrowed it.
The aim is a model for analysis aside from requirements. Requirements tell you what a system should do. Architecture tells you what it will cost to keep doing it.
The argument
Four questions, one diagram language.
The vocabulary comes first: software, software system, information system. A program is not a system, and a system is not its hardware. Most confusion in a meeting is a category error.
The C4 model supplies the zoom: context, container, component, class. One system, four levels, the same boxes at different distances. The workshop used draw.io and a generative model to draft a context diagram for DERLA — and treated the output as a draft, not an answer.
Coupling and cohesion supply the diagnosis. Weak cohesion and strong coupling produce the systems nobody wants to change: high entropy, high technical debt. The rule is old and still under-applied: strong cohesion inside a building block, weak coupling between them.
Monolith and distributed system are two extremes, never pure in practice. A monolith is one deployment unit: simple to test, to monitor, to troubleshoot; risky in coupling, hard to split across teams. A distributed system buys independent releases and error isolation and pays in complexity, which is routinely underestimated.
Conway’s Law closes the deck: a system takes the shape of the communication structure that builds it. Three teams produce three building blocks, sooner or later.
What the deck gives you
- a vocabulary
- C4
- coupling and cohesion
- monolith versus distributed
- Conway’s Law
What it leaves out
- programming
- code design
- vendor selection
- a reference architecture for your institution
What holds in 2026
The vocabulary aged; the tools did not.
C4 is still the cheapest way to make a system arguable, and coupling and cohesion still predict which system will hurt. The diagram tooling changed; generated diagrams are uneven, and a wrong diagram is worse than none.
Conway is the load-bearing idea. It connects this deck to the 2023 FORGE paper: if organisation determines technology, then architecture is an organisational document.
What I would change: less taxonomy, more one system end to end. A single repository traced from context to component would teach more than a catalogue of styles.
And I would say plainly what the deck only implies: you do not need permission to ask an architectural question. You need a diagram and three sentences.
The record
The deposit is the course deck: five hours of vocabulary, C4, and the cost of distribution.