Does uvicorn support pathsend?

[multimeric]

It seems that starlette now supports pathsend: Kludex/starlette#2435. However, using this feature requires an ASGI server that also supports pathsend. Does uvicorn?

[gi0baro]

@multimeric pathsend was proposed and added to ASGI spec by me to allow ASGI frameworks/apps to benefit from servers implementing such standard. The proposal was mainly driven by the existence of a similar behaviour in RSGI protocol.

Uvicorn doesn't support it, but it probably won't benefit at all from it, as pathsend is mainly designed for servers which are not written in Python. In order to support pathsend Uvicorn would implement the same Python code your application/framework already implement, behaving in the exact same way, and being limited by the same stack (GIL, Python objects, etc.).

[multimeric]

Thanks, so the benefits of pathsend will mostly be when using an ASGI server that can handle static files faster than Python?

[gi0baro]

I think the point is more about concurrency rather than speed. A pure Python pathsend implementation would still hold the GIL in the same way the application/framework does.

[Archmonger]

I'm not sure if I agree with this. The efficiency brought from pathsend is dependent on implementation. Pure-python ASGI webservers could implement pathsend as something akin to asyncio.to_thread( os.sendfile, ... ). When wrapping the operating system's syscall in a thread, it would absolutely would free up a ton of GIL time (especially for huge file transfers) compared to reading/sending each chunk in Python. In theory, this could open the doors to new ASGI use cases (e.g. video streaming).

If os.sendfile doesn't exist (such as on Windows) then it would need to fall-back to traditional python chunking.

@Kludex I was thinking of PRing this if you agree.

[gi0baro]

You can't use os.sendfile, the socket is handled by the event loop.

[Archmonger]

That was more of a pseudo-code example. I just drafted a functioning prototype with loop.sendfile (which uses os.sendfile under the hood). The performance implements ultimately relies on the user's event loop having built-in concurrency methodology for the syscall, which I believe is an acceptable expectation.

See PR #2840 for implementation.