Improve traceback of cancelled tasks / add cancel() msg argument

@carljm

BPO	31033
Nosy	@carljm, @cjerdonek, @ambv, @socketpair, @1st1, @matrixise, @thehesiod, @miss-islington
PRs	bpo-31033: Make traceback for cancelled asyncio tasks more useful #19951 bpo-31033: Add a msg argument to Future.cancel() and Task.cancel() #19979 bpo-31033: Use _PyErr_ChainStackItem() inside gen_send_ex(). #20173 [3.9] Use _PyErr_ChainStackItem() inside gen_send_ex(). (GH-20173) #20202

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2020-05-18.05:56:13.267>
created_at = <Date 2017-07-25.13:21:40.773>
labels = ['type-feature', '3.9', 'expert-asyncio']
title = 'Improve traceback of cancelled tasks / add cancel() msg argument'
updated_at = <Date 2020-05-19.02:14:19.517>
user = 'https://github.com/socketpair'

bugs.python.org fields:

activity = <Date 2020-05-19.02:14:19.517>
actor = 'lukasz.langa'
assignee = 'none'
closed = True
closed_date = <Date 2020-05-18.05:56:13.267>
closer = 'chris.jerdonek'
components = ['asyncio']
creation = <Date 2017-07-25.13:21:40.773>
creator = 'socketpair'
dependencies = []
files = []
hgrepos = []
issue_num = 31033
keywords = ['patch']
message_count = 23.0
messages = ['299079', '299083', '299084', '299101', '299103', '299107', '299112', '299891', '300152', '314953', '314954', '341967', '368236', '368243', '368405', '368987', '368988', '368993', '369173', '369179', '369180', '369305', '369315']
nosy_count = 8.0
nosy_names = ['carljm', 'chris.jerdonek', 'lukasz.langa', 'socketpair', 'yselivanov', 'matrixise', 'thehesiod', 'miss-islington']
pr_nums = ['19951', '19979', '20173', '20202']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue31033'
versions = ['Python 3.9']

History:

First, I tried to debug code around asyncio.Task() cancelling. So I wrote:
=====
try:
...
except Exception as e:
print(e)
=====

When task was cancelled, an empty string printed. I wondered why. So I change the code to

====

print(repr(e))

====

and it printed 'CancelledError' as expected.

Next, I tried:

====

print(Exception())

====

It prints empty string too!

So I came up to propose API change. I propose to add argument to the .cancel() methods (for Task and for Future). This argument should be passed to the CancelledError constructor. This will greatly improves debugging -- it allows to easily know why Future/Task was cancelled.

Also, this change does not break current code. Argument must be optional.

Well, don't use "print(e)" when you are printing errors in in Python. This is pretty standard behaviour for all Python code, not just asyncio. A lot of code in stdlib raises exceptions, almost none allows to customize them.

you could use gdb or pdb for the debugging.

Yes, I agree with you about my weird way of debugging. But anyway, changing API with adding ability to pass actual cause would be welcome.

Yes, I agree with you about my weird way of debugging. But anyway, changing API with adding ability to pass actual cause would be welcome.

I'm not opposed to the idea, btw. If we do decide to add an argument to 'cancel', we probably should do the same for concurrent.futures.

Another possibility would be to allow cancellation via Future.set_exception:

  task.set_exception(asyncio.CancelledError('message'))

Hmmm....

task.set_exception(Exception('xxx'))

Will it cancel whole chain of depending futures in a RIGHT way ?
Or we must require exception passed here to be subclassed from CancelledError ?

Will it cancel whole chain of depending futures in a RIGHT way ?

I was thinking about this:

def Future.set_exception(exc):
if isinstance(exc, asyncio.CancelledError):
return self._cancel(exc)
# here goes old code

def Future.cancel():
return self._cancel(asyncio.CancelledError())

Although now, that I'm looking at it, I don't like this idea, because setting as exception is a different operation from cancelling a task/future, and we shouldn't mix them. It's OK to set a CancelledError without cancelling. Also this would be a backwards incompatible change.

So back to square one -- we can consider passing extra args to .cancel() methods.

A couple thoughts on this issue:

First, I think the OP's original issue could perhaps largely be addressed without having to change cancel()'s signature. Namely, simply passing a hard-coded string to CancelledError in the couple spots that CancelledError is raised would cause the exception to display:

cpython/Lib/asyncio/futures.py

Line 153 in 733d0f6

raise CancelledError

cpython/Lib/asyncio/futures.py

Line 173 in 733d0f6

raise CancelledError

The "raise CancelledError" could be changed to "raise CancelledError('future is cancelled')", a bit like how InvalidStateError is handled a couple lines later:

    if self._state == _CANCELLED:
        raise CancelledError
    if self._state != _FINISHED:
        raise InvalidStateError('Result is not ready.')

Second, in terms of making cancellations easier to debug, is it a deliberate design decision that the CancelledError traceback "swallows" / doesn't show the point at which the coroutine was cancelled?

For example, running the following code--

    async def run():
        await asyncio.sleep(1000000)

    loop = asyncio.new_event_loop()
    task = asyncio.ensure_future(run(), loop=loop)
    loop.call_later(2, task.cancel)
    loop.run_until_complete(task)

Results in the following output:

    Traceback (most recent call last):
      File "test-cancel.py", line 46, in <module>
        loop.run_until_complete(task)
      File "/Users/.../python3.6/asyncio/base_events.py", line 466,
          in run_until_complete
        return future.result()
    concurrent.futures._base.CancelledError

In particular, it doesn't show that the task was waiting on asyncio.sleep(1000000) when the task was cancelled. It would be very useful to see full tracebacks in these cases. (Sorry in advance if this second point is off-topic for this issue.)

Yes, specifying argument to cancel in raise CancelledError would be perfect
specifying argument to .cancel() is still actual.
Yes, important part of exceptions should not be swallowed (please open separate issue). Also, it will be nice to show which code calls .cancel(), not only sleep() you menitioned.

I was about to open a new bug, but I think my idea overlaps with this one. From what I understand there are two ways to cancel tasks:

calling task.cancel()
explicitly raising a CancelledError

with #2 you can get a traceback by catching the exception which yields why the task was cancelled (per callstack, and you create the CancelledError with a custom message).

#1 I think is what this is talking about.

The most useful information IMHO is the callstack, and via #1 this callstack is completely lost.

I like the idea of having an argument to construct the CancelledError with, but I like even more the ability to tell the exception that will be raised to have the traceback of the point where the task was originally cancelled.

Right now I have to use forbidden fruit to override the Task.cancel method and store the callstack in a custom attribute.

I was talking with asvetlov and and he thought perhaps there could be an environment variable to enable this by default and then when the cancellation error is raised it can have the caller's callstack...I think ideally this would be the default if it could be done cheaply (given if a regular exception throw contains the callstack, then a cancellation exception should always as well.

Ideally I think when a task is cancelled it instantiates the exception with the callstack at that point, and then does something like a "with e" to raise the exception from the future so you get it when it's raised.

I like the idea of having an argument to construct the CancelledError with, but I like even more the ability to tell the exception that will be raised to have the traceback of the point where the task was originally cancelled.

Why don't we make CancelledErrors having proper traceback a default behaviour?

@yselivanov even better :)

New changeset 1ce5841 by Chris Jerdonek in branch 'master':
bpo-31033: Add a msg argument to Future.cancel() and Task.cancel() (GH-19979)
1ce5841

The msg argument has now been added (second PR). I'm going to keep this issue open until the traceback issue has also been addressed (the other PR), as that was one part of the discussions here.

I just want to flag one issue after rebasing my traceback PR onto what was merged. If task.cancel() is called like this--

    task.cancel("POSSIBLY LONG CANCEL MESSAGE")

There is the question of whether the passed message should be repeated each time the CancelledError is raised, or only show it in the innermost, originating exception. My preference is to do the latter because it is simpler, less verbose, and seems more correct from a Python perspective. But I wanted to flag this because the message won't be visible in the leading, outermost exception.

There is a third alternative which is to mutate the exception each time (delete the message from the earlier exception and add it to the new exception). But that seems more fraught and what I'd consider surprising behavior.

Lastly, to illustrate, here is the more verbose option (the one I think it **shouldn't** look like):

Traceback (most recent call last):
  File "/.../cpython/test-cancel.py", line 4, in nested
    await asyncio.sleep(1)
  File "/.../cpython/Lib/asyncio/tasks.py", line 670, in sleep
    return await future
asyncio.exceptions.CancelledError: POSSIBLY LONG CANCEL MESSAGE

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/.../cpython/test-cancel.py", line 11, in run
    await task
asyncio.exceptions.CancelledError: POSSIBLY LONG CANCEL MESSAGE

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/.../cpython/test-cancel.py", line 15, in <module>
    loop.run_until_complete(run())
  File "/.../cpython/Lib/asyncio/base_events.py", line 642, in run_until_complete
    return future.result()
asyncio.exceptions.CancelledError: POSSIBLY LONG CANCEL MESSAGE

Elevating to release blocker to make sure it's included. The PR is good.

New changeset da742ba by Chris Jerdonek in branch 'master':
bpo-31033: Improve the traceback for cancelled asyncio tasks (GH-19951)
da742ba

Thanks so much, Yury.

(Removing the "release blocker" flag now that it has been merged.)

New changeset ff7a8b0 by Chris Jerdonek in branch 'master':
Use _PyErr_ChainStackItem() inside gen_send_ex(). (GH-20173)
ff7a8b0

New changeset f02c304 by Miss Islington (bot) in branch '3.9':
Use _PyErr_ChainStackItem() inside gen_send_ex(). (GH-20173) (bpo-20202)
f02c304

Navigation Menu

Repository navigation

Improve traceback of cancelled tasks / add cancel() msg argument #75216

3 remaining items

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Improve traceback of cancelled tasks / add cancel() msg argument #75216

Description

Activity

socketpair commented on Jul 25, 2017

1st1 commented on Jul 25, 2017

matrixise commented on Jul 25, 2017

socketpair commented on Jul 25, 2017

1st1 commented on Jul 25, 2017

socketpair commented on Jul 25, 2017

1st1 commented on Jul 25, 2017

cjerdonek commented on Aug 8, 2017

socketpair commented on Aug 11, 2017

thehesiod commented on Apr 4, 2018

1st1 commented on Apr 4, 2018

thehesiod commented on May 9, 2019

3 remaining items

cjerdonek commented on May 15, 2020

cjerdonek commented on May 16, 2020

cjerdonek commented on May 16, 2020

1st1 commented on May 18, 2020

cjerdonek commented on May 18, 2020

cjerdonek commented on May 18, 2020

cjerdonek commented on May 18, 2020

ambv commented on May 19, 2020

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions