Learn system design by reading the code that proved it

Books teach system design in the abstract: "use a write-ahead log", "prefer event loops for I/O-bound work", "content-address your data". Those are true and useless until you have watched a real codebase pay for them — where the loop lives, what the log record looks like, what trade was made to get there.

The atlas points at 34 (and growing) canonical codebases and walks through the decisions frozen into their code: the two or three defining trade-offs, the data structures that carry the design, the concurrency model, the seams where the system extends. You read the atlas entry, then you read the code alongside it. That is the whole method.

1. Pick a domain you want to understand — cache, storage engine, VCS, web server, database, runtime... Browse catalog.yaml (entries are tagged by domain and difficulty).
2. Read the catalog entry's "why" paragraph — it tells you what design lesson this codebase is unusually good at teaching.
3. Shallow-clone the project so you can follow along:

   mkdir -p clones && git clone --depth 1 <repo-url> clones/<id>
   

   Deep-dives pin the commit they were written against; expect small drift, the file/function landmarks are stable for years in mature projects.
4. Read the deep-dive top to bottom once (architecture → data structures → concurrency → trade-offs), then work the reading path in the actual clone, opening every cited file. The reading path is the product; the prose is the guide.
5. Steal the lessons — each deep-dive ends with "what to steal": transferable design moves you can apply to your own systems.

Time budget: a deep-dive plus its reading path is a focused 3–8 hours (depending on the entry's difficulty rating). Do not rush the reading path.