Implementation for using hmr with uvicorn by uvicorn-hmr

[CNSeniorious000]

• Related to FastAPI with on-demand HOT reload! 🙌 fastapi/fastapi#13192

Currently uvicorn --reload restart the python process every reload, which makes heavy 3rd-party modules being imported everytime.

I implemented a uvicorn-hmr package to solve this problem. The differences between uvicorn main:app --reload and uvicorn-hmr main:app can be concluded as:

1. uvicorn-hmr stop and restart uvicorn servers in the same process, so 3rd-party librarys are cached by default.
2. uvicorn-hmr depends on hmr, which performs fine-grained on-demand dependency tracking on every imported name, it your main module depends on a and b and only a.py is changed, b will not be invalidate at all.

Everybody can have a try by simply replacing uvicorn --reload main:app with uvicorn-hmr main:app.

I've tested this on several of middle-sized ASGI Python servers. Looking forward to any responses!

Caution

hmr are sometimes refer to a feature that updates the page in the browser on the client side when the server code changes. This is not that. This package is a server-side HMR, that reloads the server code when it changes.

[Kludex]

Hi @CNSeniorious000 ,

What's the behavior of the lifespan events using hmr?

Can we include this on uvicorn itself? Would it be too hard with the current implementation?

[CNSeniorious000]

It's actually quite straightforward.

Here's how I start a uvicorn server:

    def start_server(app: "ASGIApplication"):
        nonlocal stop_server

        server = Server(Config(app, host, port, env_file=env_file, log_level=log_level))
        finish = Event()

        def run_server():
            server.run()
            finish.set()

        Thread(target=run_server, daemon=True).start()

        def stop_server():
            server.should_exit = True
            finish.wait()

Whenever hmr detects that the app instance has changed, uvicorn-hmr simply calls stop_server() and then start_server(app). This means all the usual ASGI lifespan events are handled gracefully by uvicorn, since the server thread is shut down with server.should_exit = True.

Can we include this on uvicorn itself?

I think it's possible and worthwhile! The hmr package takes care of dependency graph tracking, and uvicorn-hmr is just ~40 lines of code that re-launch the server thread when necessary. Including this hot-reloading behavior into uvicorn itself may push the DX of Python backend development a step forward.