Trigpoint
Collections › Software & Systems Architecture
SHELF

Software & Systems Architecture

Choosing a structure for a knowledge system
Discipline: Computer Science
“Software architecture is only for people who write code”: that is a misreading.
17 volumes · 9 lab cases Every source: bibliography checked ✓
↓ Jump to the node list

You choose a structure for a knowledge system every day: where truth sits, how derivations arise, where the boundary falls and who decides it. This field has asked each of those questions for decades. The answers are not in the code; they are in the trade-offs.

Node sequence

Suggested reading order · where truth lives, how derivations arise, where the boundary falls, what the boundary follows, what you give up once distributed, and when both ends of reading and writing are models
№ 01§ Where truth livesOrder within the group: state the principle → see how it lands in the data layer → then an ancestor that is often misread as it
01

Don't repeat yourself (DRY)

Hunt & Thomas (1999)

What the original says is that every piece of knowledge in a system must have a single, unambiguous, authoritative representation.

4 min read
02

Single source of truth (concept history)

Devlin & Murphy (1988) → Loshin (2008)

A single source of truth says that every element is maintained in one place and everywhere else refers to it.

Lab cases ×35 min read
03

Model-view-controller (MVC)

Reenskaug (1979)

Separate the data from the display, so that one body of data can have several presentations that stay in step.

4 min read
№ 02§ How derived things ariseOrder within the group: how the source of fact is stored → the reading side can be a different model. The two are often sold together, though both original sources say they can be separated
04

Event sourcing

Fowler (2005)

Store every change as one immutable event, and obtain the current state by replaying that sequence.

4 min read
05
№ 03§ Where the boundary fallsOrder within the group: criteria for siting it (cost structure, not feature grouping) → then the language boundary → then the direction of dependency
06

Modularity and thin crossing points

Baldwin & Clark (2000) / Baldwin (2008)

Where a boundary should open is settled by the cost of the traffic across it, not by which category the work falls into.

5 min read
07

Domain-driven design (DDD)

Evans (2003)

The main battleground of complexity is the domain itself, not the technical layer.

5 min read
08
№ 04§ What the boundary followsOrder within the group: the constraint first (structure will mirror communication paths) → then what can be designed (make interfaces explicit, and expect the relationships to evolve)
09

Conway's law

Conway (1968)

Organisations that design systems produce designs that mirror their own communication structures.

Lab cases ×25 min read
10

Team topologies

Skelton & Pais (2019)

Relationships between teams want declaring explicitly, the way an interface is declared, rather than resting on tacit understanding.

4 min read
№ 05§ What you gave up once you distributed itOrder within the group: the decision (split) → the cost (the theorem) → then a position that reverses it explicitly. This closing group rewrites the first one, which is the evidence that this shelf is not doctrine
11

Microservices

Lewis & Fowler (2014)

Split one application into a set of services that are released independently, and the unit of deployment and scaling gets smaller.

5 min read
12

CAP theorem

Brewer (2000) / Gilbert & Lynch (2002)

When a network partition happens, you can have consistency or availability, and not both.

Lab cases ×25 min read
13

Data mesh

Dehghani (2019, 2022)

Once scale arrives, holding one globally uniform canonical model makes that model the bottleneck.

5 min read
№ 06§ When both ends of reading and writing are modelsOrder within the group: the external evidence → why it still breaks → which two axes it breaks on → then the highest-risk class. This group is also the honesty anchor for the shelf: it lays out this site's own gatekeeping
14

Retrieval-augmented generation (RAG)

Lewis et al. (2020)

Split the model's memory in two: what is pressed into the weights, and what is fetched at the moment of generating.

4 min read
15

Hallucination in NLG

Ji et al. (2023)

Hallucination is generation that reads fluently while being unfaithful to its source, or making no sense at all.

4 min read
16

Faithfulness vs factuality

Maynez et al. (2020)

Faithfulness asks whether a statement is supported by the source. Factuality asks whether it is true of the world.

5 min read
17

Fabricated citations

Walters & Wilder (2023)

Walters and Wilder (2023) measured fabricated citations in ChatGPT output: complete and convincing in form, and the thing cited does not exist.

5 min read

After reading this shelf

Several pages on this shelf describe some failure that does not report itself. See "Don't repeat yourself (DRY)", "Single source of truth (concept history)", "Event sourcing", "Command query responsibility segregation (CQRS)", "Domain-driven design (DDD)", "Retrieval-augmented generation (RAG)", "Hallucination in NLG" and "Fabricated citations".

For three of these practices, the pages state where the failure moves to once the practice is adopted. Naming one place as authoritative: the problem moves from "which copy is right" to "are the references pointing at it still right", and those references raise no error when the target is renamed or moved. See "Single source of truth (concept history)". Keeping a separate projection for reading: the projection lags behind the original, and the lag itself gives no signal. See "Command query responsibility segregation (CQRS)". Having a model retrieve before it answers: the ceiling on quality moves from how much the model remembers to how good the retrieval store is; when the wrong passage is retrieved the output is just as fluent, and the error often cannot be seen. See "Retrieval-augmented generation (RAG)".

Two further pages state that aligning everything to one place does not make that place correct. "Don't repeat yourself (DRY)" states that the principle makes no promise of correctness; if the one place is wrong, the error only becomes more consistent. "Faithfulness vs factuality" states that when the source itself is wrong, reproducing it accurately still counts as faithful.

On boundaries, three pages write time into the criterion. "Microservices" relays advice from one of its authors: find stable boundaries inside a monolith first, then split. "Team topologies" states that close, frequent collaboration belongs to the discovery period and should be stepped down once the boundary is understood. "Modularity and thin crossing points" states that the cost of crossing a boundary changes, so boundaries need periodic reassessment.

From these the shelf draws two conclusions. First, after adopting any of the three practices above, schedule a fixed check alongside it: verify the references, stamp the projection with the time it was generated, and check sentence by sentence that the output says what its source says. Checking the source itself is a separate task, and none of these structures does it for you. Second, do not cut a boundary before it is understood, and reassess it some time after cutting. Neither conclusion is a statement made by any single source.

At what point one structure should give way to another, the pages on this shelf give directions and questions, not numbers. "Single source of truth (concept history)" and "Data mesh" go only as far as saying that the number of domains and the rate of change "pass a certain point". "Modularity and thin crossing points" states that its source gives no operational measure. "Team topologies" states that when to step down is a matter of judgement. "Conway's law" states that how accurate the law is cannot be answered on that page.

Most pages on this shelf rest on trade books, blog posts and documents published by their own authors, none of them peer reviewed; each page says so for itself. "Conway's law" carries one quantitative study from 2012, which this site checked only as far as the abstract and whose strength it does not report.

The main sources for the four pages in the group "When both ends of reading and writing are models" were published between 2020 and 2023. "Fabricated citations" states that the two rates on that page are tied to that study, those two model versions and a 2023 measurement, and cannot be taken as figures for today's models. "Hallucination in NLG" states that hallucination can at present only be narrowed and must not be assumed to have reached zero.

Applying the structures of software systems to individual knowledge work and to working with models is this site's analogy; these sources did not test it. "Team topologies" and "Data mesh" also state that their practices address multiple teams or large organisations, and that an individual can borrow only part of them.

Read it yourself · Theory layer

The whole theory layer is here, with no account needed. Read it and go.

Trigpoint Pass · Application layer

The case layer (our own tests of source of truth, boundaries and distribution trade-offs built into governance rules, including a metric that corroded and one commit that swept another session's changes in) and the full verification layer of most nodes (how far each source was checked, counter-evidence and failure boundaries).

Trigpoint Pass ▸

Tell me when a new shelf opens

Tell me when a new shelf opens

Shelf notices are sent only when a new shelf opens or when the reading rules change.