ARCH Case studies — postmortems & runbooks
Written the way incidents get written up: what was built, what broke, what the design does about it. Anonymized, load-bearing, and free of hero narratives — the system is the protagonist.
Presence, conflict resolution, and sync that degrades politely when the network does not. CRDT sync over Durable Objects — and what actually degrades first at ten thousand concurrent editors.
Tenant isolation with Workers, D1 and R2. Routing at the edge, billing boundaries in the schema, and no shared state by accident.
Retrieval that can prove where an answer came from. Evaluation loops before feature loops; a retrieval pipeline that regressed quietly, and the eval harness that now catches it.
Agents as narrow tools with contracts. Tool boundaries, checkpoints, human gates — autonomy budgeted as a governed resource.
Module boundaries a new hire can learn in a week and a team can hold for five years. Ownership maps, dependency rules, and the small set of conventions worth enforcing by machine.
Five case studies in system design. Each answers the same question: where do the boundaries go, and what happens when a piece fails? The interesting parts are load-bearing. Everything else is deliberately dull.