Channel context manager eats exceptions when cancelled

[oremanj]

I ran into this today:

import trio

async def produce(send_channel):
    async with send_channel:
        print("in async with send_channel")
        raise RuntimeError("kaboom")

async def test():
    send_channel, receive_channel = trio.open_memory_channel(0)
    with trio.open_cancel_scope() as scope:
        scope.cancel()
        await produce(send_channel)

trio.run(test)
print("where's the kaboom?")

Prints:

in async with send_channel
where's the kaboom?

I'm guessing this is because channel.aclose() is a checkpoint? Seems like this exception-eating might be the same or an overlapping issue as #455, but it was especially surprising to me, and I think might create user confusion as people start using async with foo_channel: in more and more places.

[smurfix]

It is. There's not much that Trio can do internally to fix that IMHO, but what about a

def __aexit__(self, *tb):  # and/or __exit__
    with merge_exceptions(*tb):
        foo()

context manager that packages the exception(s) raised by foo, and the traceback if one exists, into a MultiError (or whatever Python's stdlib ends up naming the thing)? At least then we'd have a standard idiom that we can use, and the one additional line of boilerplate won't hurt (too much) (and can be checked for by a linter).

[njsmith]

Yeah, this is exactly the same as #455, so I'll close it to consolidate the discussion.