python-trio/async_generator

asynccontextmanager does not support callable classes

mrosales opened this issue · 4 comments

I was expecting something like the following example to work, but it does not. This originally came up in: fastapi/fastapi#1204

from async_generator import asynccontextmanager

class GenClass:
    async def __call__(self):
        yield "hello"


cm = asynccontextmanager(GenClass())

with cm as value:
    print(value)

The result is an exception:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/.../lib/python3.6/site-packages/async_generator/_util.py", line 100, in asynccontextmanager
    "must be an async generator (native or from async_generator; "
TypeError: must be an async generator (native or from async_generator; if using @async_generator then @acontextmanager must be on top.

Put @asynccontextmanager on the class's __call__ method, not on the class itself. That will produce a class whose instances yield async generators when they are called. If that doesn't meet your needs, can you clarify where it falls short?

@asynccontextmanager is intended to be used as a function decorator. That is, it accepts as argument an async generator function (i.e., a function that returns an async generator iterator) and returns a function that returns an async context manager. So even if it did support being applied to a class instance, you would use async with cm() as value, not with cm as value. But @asynccontextmanager does check specifically for an async generator function, not just a thing-that-when-called-returns-an-async-generator, so this won't work. You can fool it (look at the definition of isasyncgenfunction in _impl.py to see what you need to spoof) but decorating the __call__ method will likely work much better.

I just looked at the linked issue. If you are given the class definition and can't modify it, you could do something like cm = asynccontextmanager(MyClass.__call__); async with cm(MyClass()) as value: ...

Sure, you're recommendation of a workaround makes sense and that's more or less what I did to unblock the python3.6 failure on the MR that I opened for the linked project issue, but I think the functionality that I was trying to call out is slightly different.

First, the example was a typo, you are correct that it should be async with, but the point that I was trying to show is that the TypeErrror was thrown on the line before it. I was filing this as a bug because mostly because it differs from the behavior of the native python3.7+ function in contextlib

# python 3.6
from async_generator import asynccontextmanager

# this line throws a TypeError
cm = asynccontextmanager(GenClass())
# python 3.7
from contextlib import asynccontextmanager

# this works just fine
cm = asynccontextmanager(GenClass())

@mrosales fyi this works on the contextlib2 backport:

from contextlib2 import asynccontextmanager


class GenClass:
    async def __call__(self):
        yield "hello"


cm = asynccontextmanager(GenClass())

async def amain():
    async with cm() as value:
        print(value)

import anyio

anyio.run(amain)