I found some issue that I suppose is a bug.

Let us have long running coroutines. We use them in gather, and one of them raises an error. Since then we cannot cancel the gather anymore, thus remaining children are not cancelable and executed until complete or raise an exception themselves.

===

import asyncio


async def coro_with_error():
    # Coro raises en error with 1 sec delay
    await asyncio.sleep(1)
    raise Exception('Error in coro')


async def cancellator(coro):
    # We use this to cancel gather with delay 1 sec
    await asyncio.sleep(1)
    coro.cancel()


async def success_long_coro():
    # Long running coro, 2 sec
    try:
        await asyncio.sleep(2)
        print("I'm ok!")
        return 42
    except asyncio.CancelledError:
        # Track that this coro is really cancelled
        print('I was cancelled')
        raise


async def collector_with_error():
    gather = asyncio.gather(coro_with_error(), success_long_coro())
    try:
        await gather
    except Exception:
        print(f"WHOAGH ERROR, gather done={gather.done()}")
        print(f'EXC={type(gather.exception()).__name__}')
        # We want to cancel still running success_long_coro()
        gather.cancel()


async def collector_with_cancel():
    # Gather result from success_long_coro()
    gather = asyncio.gather(success_long_coro())
    # schedule cancel in 1 sec
    asyncio.create_task(cancellator(gather))
    try:
        await gather
    except Exception:
        print(f"WHOAGH ERROR, gather done={gather.done()}")
        print(f'EXC={type(gather.exception()).__name__}')
        # We want to cancel still running success_long_coro()
        gather.cancel()
        return

# First case, cancel gather when children are running
print('First case')
loop = asyncio.get_event_loop()
loop.create_task(collector_with_cancel())
# Ensure test coros we fully run
loop.run_until_complete(asyncio.sleep(3))
print('Done')

# Second case, cancel gather when child raise error
print('Second case')
loop = asyncio.get_event_loop()
loop.create_task(collector_with_error())
# Ensure test coros we fully run
loop.run_until_complete(asyncio.sleep(3))
print('Done')

===

Actual output:

First case
I was cancelled
WHOAGH ERROR, gather done=True
EXC=CancelledError
Done
Second case
WHOAGH ERROR, gather done=True
EXC=Exception
I'm ok!
Done

Expected output:

First case
I was cancelled
WHOAGH ERROR, gather done=True
EXC=CancelledError
Done
Second case
I was cancelled
WHOAGH ERROR, gather done=True
EXC=Exception
Done

Documentations says:

If gather() is cancelled, all submitted awaitables (that have not completed yet) are also cancelled.
But it mentions no cases on child coros' exceptions.

From doc:

If return_exceptions is False (default), the first raised exception is immediately propagated to the task that awaits on gather(). Other awaitables in the aws sequence won’t be cancelled and will continue to run.
Which is true, exception is propagated, the gather has an exception set, marked as done() so its children are not cancelled.

I believe asyncio should allow cancellation in that case.

Hi Dimitri,
You are right, gather.cancel() doesn't work once it has propagated an exception. This happens because after propagating the exception to the caller, gather is marked as done. And cancel doesn't work after a Future object has been marked done.
You can test the same by printing the return value of gather.cancel(). It will be False

I also believe that the documentation of gather should explicitly mention this. But, depending on the fact, whether this is an expected behaviour, current code base might also need changes.

Therefore I have created two patches, one updating the current documentation according to the current functionality, and other changing the codebase which supports cancelling even after raising exceptions.

I will try to contact one of the core developers on deciding which one is the way to go.

The behavior change is not backward compatible, we cannot apply it.

Also, please use github pull requests for proposing a patch: https://devguide.python.org/pullrequest/

Hi Andrew,
Thanks for replying!

I understand that the behavior change is not backward compatible. But the current behavior is a bit ambiguous ad mentioned by Dimitri, therefore I have updated the documentation and opened a Pull request.

I would be very glad if you could review it.
Thanks!

Also, would adding an optional argument like force_cancel=False to gather.cancel, which won't return even if gather is done be backward compatible. Should I open a PR with this change also, so that you can better review it and see if it looks good.

New changeset d42528a by Vinay Sharma in branch 'master':
bpo-37703: improve asyncio.gather documentation regarding cancellation (GH-15312)
d42528a

New changeset 58f59a9 by Miss Islington (bot) in branch '3.8':
bpo-37703: improve asyncio.gather documentation regarding cancellation (GH-15312)
58f59a9

New changeset 46634b7 by Miss Islington (bot) in branch '3.9':
bpo-37703: improve asyncio.gather documentation regarding cancellation (GH-15312)
46634b7