Production

For the reader who has an application and has to run it. The lessons teach the framework; these pages teach the operational surfaces around it — what a deployment has to do in what order, what the runtime prints while it runs, what each failure is called, which boundaries hold on their own, and which ones are yours.

Everything here is measured on this repository's toolchain. Each page's code is a module under examples/learn_production/ that the site's own suite runs, and every output block is that module's real stdout, compared on every build. Nothing on these pages is a summary of documentation somewhere else.

Page	What it answers
Deploying it	the factory contract, the check step, migrations as a deploy step, and the choice between refusing to start and reporting not-ready
Observability	what the runtime logs and what it deliberately does not configure, the request pipeline's order, and the diagnostic context a log line survives
Troubleshooting	the error taxonomy, the one line that identifies each failure, and what to do about it
Security boundaries	what the runtime refuses to leak, what it refuses to import, and the large part of security that is a decision you make
Performance and concurrency	write contention under SQLite, the retry budget that bounds it, and why the numbers on this page are counts
Internals	the lifecycle as a state machine, close ordering, the request path, and the design rules behind them

How to read a page here

Each page names the shipped feature it relies on, or says which part is application code you write, in the three labels the rest of the site uses:

• Built-in — the runtime does it, and the page can show it.
• Built-in (extra) — a shipped implementation of a capability you opted into by installing an extra, such as xyberos[database].
• Application-side — the framework provides no such thing on purpose, and the page says what to build instead. A framework that imports nothing it does not need cannot also be your job queue, your rate limiter, or your secret store.

What is deliberately absent

Three absences are worth stating before the pages, because each one is a design decision rather than a gap:

• No server, and no request handling outside the optional http extra. uvicorn is your choice, and the adapter compiles the application into a Starlette application; the deploy page below and the traditional path's deploy lesson show the whole command.
• No background worker and no scheduler. The hybrid path's background lesson and the durable page show the two shapes the framework does support — a job row an application's own worker claims, and a journal that outlives the process with its limits stated — but neither is a scheduler.
• No sandbox. Plugins and providers are trusted, in-process Python code, which is why security is a page about your decisions rather than a page about guarantees.