The Journal

Notes on building software
that stays good.

Hard-won lessons from the studio. Most production failures come down to one thing: the state a system reports is not the state it is actually in. Here is what we have learned catching that gap.

EngineeringJul 20263 min read

A merge is not a deploy

Closing a pull request tells you the code got accepted. It tells you nothing about whether it is live. Why we gate "shipped" on the artifact users can reach, not the merge event.
EngineeringJul 20263 min read

"Success" is not proof of correctness

A status of "complete" and an HTTP 200 mean a process finished, not that it did the right thing. As more of what we build is machine-generated, that is where the costly bugs hide.
EngineeringJul 20263 min read

Never trust a self-declared risk level

The riskiest changes are often the ones that look safe. Why we stopped letting anything, person or AI, grade its own homework, and judge a change on what it actually does.