← Notes
Reading list, summer 2026
Twice a year we stop building and read for a fortnight. It started as a reaction to shipping something we did not understand, and it has become the most reliable source of our better decisions. This is what stayed with us from the summer.
Papers
- Simple Testing Can Prevent Most Critical Failures (Yuan et al.) — the one that keeps paying rent. Most catastrophic failures are triggered by a handful of specific inputs, and most of those inputs are never tested. We now write the boring test that nobody wants to write.
- The Tail at Scale (Dean & Barroso) — still the clearest explanation of why your p99 is nobody's p50.
- Time, Clocks, and the Ordering of Events (Lamport) — read properly for the first time. The part that landed was not the algorithm but the framing: order is a property of what you observed, not of the world.
- On the Dangers of Stochastic Parrots — useful less as a position than as a checklist of questions to ask before deploying anything that generates text.
- Crash-Only Software (Candea & Fox) — eleven pages that made us delete a shutdown path.
- An Incremental Approach to Compiler Construction — because we had never built one and it turns out the discipline transfers.
Books
- Normal Accidents (Perrow) — the argument that some systems are coupled too tightly to be safe, no matter how careful the operators. Uncomfortable reading for anyone who designs platforms.
- The Design of Everyday Things (Norman) — reread for the fourth time. Every year the examples feel less like product design and more like operations.
What changed in our work
Two things, both small. We added a test for the failure path that we had been avoiding because it was awkward to construct. And we deleted a graceful-shutdown handler in Atlas that had been making restarts slower and less predictable for eighteen months. Both changes came from a paper, which is a better hit rate than most of our planning.