You already know: distributed-systems reliability, replication and failure modes (D14); serving platforms and gateways (D17); ML pipelines and feature stores (D11); the costs of every component (D12, D17).
Why this matters for Atlas. Individual components are rarely what makes large systems hard. The difficulty lives in their shape: where the boundaries are, what crosses them, which failures spread, and how easily the whole can change. Staff engineers are expected to see that shape, reason about it quantitatively, and steer it. This lesson covers the core lenses: boundaries and coupling, interfaces and contracts, failure domains and blast radius, back-of-the-envelope capacity, and designing for evolution, each with a small calculation that makes the reasoning concrete.
1. Boundaries and coupling
The question. Where should one service, team or module end and another begin?
Start from nothing. Draw boundaries around things that change together and fail together, and keep what crosses them small. Measures of trouble: fan-out (how many things a change touches), cycles (A depends on B depends on A, so neither can change alone), and shared mutable state (two services writing the same table are one service in disguise). Conway's law adds an organizational lens: systems mirror the communication structure of the teams that build them, so boundaries that cut across team lines will be expensive to maintain; sometimes the right move is to change the team structure (the "inverse Conway manoeuvre").