If you are trying to use AsyncMock to mock a coroutine that returns awaitable objects, AsyncMock awaits on those objects instead of returning them as is.

Example:
mock = AsyncMock(return_value=asyncio.Future())
v = await mock() # blocks on trying to await the future

Expected:
mock = AsyncMock(return_value=asyncio.Future())
v = await mock()
assert isisnstance(v, asyncio.Future)

This problem affects side_effects and wraps.

New changeset 046442d by Lisa Roach (Jason Fried) in branch 'master':
bpo-38857: AsyncMock fix for awaitable values and StopIteration fix [3.8] (GH-17269)
046442d

New changeset b2744c1 by Andrew Svetlov (Lisa Roach) in branch '3.8':
[3.8] bpo-38857: AsyncMock fix for awaitable values and StopIteration fix [3.8] (GH-17269) (bpo-17304)
b2744c1

I think this deserves discussion :)

On one hand, it's a welcome change, on another it's kind of a regression.

Up until 3.8, our tests used to look like this:

---
# code under test

async def foo():
    return await bar()

# test

async def helper(value):
    return value


async def test_foo():
    with patch("bar", return_value=helper(42)):
        assert await foo() == 42

I feel that the default class patch() uses for new has crept in too quietly in 3.8.

At the same time, helper was only used because there was no AsyncMock.
(or at times, a 3rd party library, asynctest was used).

So, on one hand, it's a bit of a regression, but on the other, looking ahead, I would really like unittest.mock to do the right thing.

Can we have it both ways? If not, what way is a better way?

Its not possible to have it both ways. Also it stinks too much of trying to guess.

The root of your issue is you want a normal MagicMock not an AsyncMock. Its the automatic behavior of patch to pick AsyncMock vs MagicMock that is the heart of your issue. This bug fix doesn't involve that behavior at all, and AsyncMock was measurably broken without the fix, in an unavoidable way. Your breakage is avoidable by changes to how you patch.

with patch("bar", return_value=42)

To still do it the old way you would have to pass new_callable=MagicMock to patch.

Thank you for explanation, Jason!

I guess that the bug report and the patch were too technical for me to understand 😅

I'm happy with the behaviour in Python 3.8.1 and now I know it's going to stay, I'll just change the tests in our code base.