serve_forever blocks upon cancellation when clients are connected (upstream change in CPython >= 3.12)

[pskopnik]

An upstream change in CPython >= 3.12 changes the behaviour of asyncio.Server.serve_forever() to wait for all clients to disconnect voluntarily before returning upon cancellation. There is discussion to change this behaviour, but no patch has been finalized yet python/cpython#123720.

websockets is affected, because it relies directly on serve_forever(). Common patterns for running a server will start blocking in Python >= 3.12 upon cancellation until all clients disconnect.

• serve() context manager + serve_forever(), described in the Quick examples.

  async with serve(hello, "localhost", 8765) as server:
      await server.serve_forever()

  This is the easiest to reproduce the problem: Create the script from the example, start the script, connect a client (any TCP client works, or a websockets client from same example page), then stop the server via Ctrl-C.

• Plain serve() + serve_forever(), described in websockets.asyncio.server.serve API docs.

  server = await serve(handler, host, port)
  await server.serve_forever()

Waiting for an arbitrary future continues to work, also described in websockets.asyncio.server.serve docs. Upon cancellation, the serve() context manager is exited and will shut down the server including all connected clients.

async with serve(hello, "localhost", 8765) as server:
    await asyncio.get_running_loop().create_future()

Not sure how to continue here. This issue describes the problem and will hopefully save others the time to research. We could update all examples in the docs to use the only working pattern. Or add a note at least to server.serve mentioning the issue. To me it seems likely that the issue will be fixed in CPython and also backported to 3.12, so implementing a custom serve_forever() may not be worthwhile.

[aaugustin]

Thank you for investigating the situation and reporting your findings. websockets sticks to Python's behavior in such cases because having a slightly different behavior from the stdlib is likely to cause even more confusion.

We could update all examples in the docs to use the only working pattern.

I just made the opposite change because the await asyncio.get_running_loop().create_future() pattern didn't feel intuitive and didn't look popular 🤦

[pskopnik]

I just made the opposite change because the await asyncio.get_running_loop().create_future() pattern didn't feel intuitive and didn't look popular 🤦

I feel the same way. serve_forever() was a great addition, helping to bring the behaviour of running a server in line with any other cancellable async task.

Looking at pypistats indicates there are ~ 24 % of package downloads >= 3.12 and hence affected by this issue.

What do you think of adding an admonition to the relevant docs pages? Assuming (hoping) that the CPython change is only temporary, it makes little sense to change all examples or introduce new API.