Asyncio's loop.create_unix_server makes blocking calls

[cbornet]

Bug report

Bug description:

It seems create_unix_server makes some calls to blocking functions.

cpython/Lib/asyncio/unix_events.py

Line 273 in 4d56c40

async def create_unix_server(

In particular:

• os.fspath
• os.stat
• os.remove
• socket.bind (similar issue in trio: socket.bind() is a blocking call python-trio/trio#241)

CPython versions tested on:

3.11

Operating systems tested on:

No response

[ZeroIntensity]

cc @graingert

[gvanrossum]

The calls you mention are not worth making asynchronous -- we'd just be making a simple thing complicated for no reason.

• Modern filesystems use caching and filesystem operations are effectively synchronous. People have looked into making async versions of filesystem operations and IIRC the conclusion was that it wasn't worth the complexity -- no actual parallelism was gained.

• It's the same for bind() -- it does no actual I/O, it just copies some bytes from user space to kernel space.

[cbornet]

Thanks for the quick feedback.
Wouldn't delegating these operations to a thread allow to release the GIL and to not block the rest of the loop ? I think that's what other event loop systems such as Node.js do, even for simple syscalls such as stat.
As for the bind, the trio issue says that "it has to do filesystem traversal", doesn't it count as a blocking op ?

In general, it would be nice to have guidelines in the asyncio docs as to which framework functions are considered blocking and which are not.

[gvanrossum]

I don’t understand if you’re arguing from a conceptual purity concern, or if this has caused you problems in an actual project?

[cbornet]

It's more conceptual and looking for best practices, that make the system the most stable and performant.
Also, I wrote a tool lib to detect blocking calls called blockbuster (it's how I saw these os calls in asyncio) and would be interested in knowing which functions should trigger it. But that's probably something to discuss elsewhere.

[gvanrossum]

Okay, thanks. Then I will keep this closed.