Production Engineering

A script that works on your laptop and a service someone else operates are separated by a list of unglamorous things: it has to be installable, its dependencies have to resolve the same way twice, a gate has to refuse the change that breaks it, an operator has to be able to see what it did, and its internal seams have to survive being pulled apart when one part needs to scale. This domain is that list, worked in order.

It is taught local-first, which is the honest constraint of a browser runtime and turns out to be a better teacher anyway. You do not push to a registry; you read the wheel and the sdist a registry expects, and prove by hand that a lockfile makes a build reproducible. You do not run a cloud pipeline; you author a real GitHub Actions workflow and drive it through an in-process CI simulator where a job's needs is visibly its gate, a failed gate skips what depends on it, and fail-fast cancels the rest. You do not deploy a container; you build a linter for the release-hygiene mistakes in a Dockerfile — unpinned base, cache-busting layer order, running as root, no multi-stage build. Real serving, sockets and queues stay prose rather than being faked. Observability then gets built from scratch — structured logs bound to a correlation id, a metrics registry, trace spans across a request — and architecture closes the domain by inverting dependencies so the core depends on ports rather than concretes, with idempotency and retry-with-backoff making a remote call survivable. Throughout, this repository is the worked example.