CancelledError suppressed where it should not about aiohttp HOT 4 OPEN

It assumes that, if the current task is cancelled, then it is due to the same exception we are just processing in the exceptbranch, as it re-raises it.

I suggest reading the linked article. The current task is cancelling because it is being asked to cancel (i.e. cancelled from outside the task). In this situation, the cancellation should be propagated. If the inner task has been cancelled and we are receiving the exception as it bubbles up, then the current task is not cancelling(). There are no assumptions in that logic.

To avoid maintaining rather verbose, complex code, I'd personally prefer to just fix this on 3.11+. I think you're the first user to actually report this, so I'm not too worried about the behaviour existing in the oldest versions of Python for a little while longer, there are similar issues in asyncio itself in those versions regardless.

from aiohttp.

This is something we just fixed in aiohttp-sse. Should probably have a look through all our code to find similar issues.

As in aiohttp-sse, I'd be fine with a fix that only works on Python 3.11, as it's awkward to fix on older versions and has been like that forever: https://github.com/aio-libs/aiohttp-sse/pull/459/files

Further information: https://superfastpython.com/asyncio-task-cancellation-best-practices/#Practice_1_Do_Not_Consume_CancelledError

from aiohttp.

I have read your patch ( https://github.com/aio-libs/aiohttp-sse/pull/459/files ). However I am not sure it is safe. It assumes that, if the current task is cancelled, then it is due to the same exception we are just processing in the exceptbranch, as it re-raises it. I am not sure this has no consequence. My experience with asyncio has been that some innocent assumptions can lead to surprising edge cases.

I wrote initially that I thought that my code was safer. But, reading your code that seems very logical and simple (even if requiring python >= 3.11), I am not sure. Maybe they are equivalent technically.

That would be two versions available for this patch:

• Your version which is much more simple (kudos for this) but requires python >= 3.11
• Mine which is universal (as long as asyncio is available) but takes more code to do the same, as it needs to go down into the asyncio Future machinery - yet despite the appearance it is not exotic (this code is like asyncio.gather, just simpified)

from aiohttp.

I agree. Your code is much (much) more maintainable and forward-looking.

from aiohttp.