traceback module can't format/print unhashable exceptions

[Trundle]

BPO	28603
Nosy	@terryjreedy, @Trundle, @berkerpeksag, @serhiy-storchaka, @JelleZijlstra, @zaneb
PRs	bpo-28603: Fix exception chaining for unhashable exceptions #4014 [3.6] bpo-28603: Fix formatting tracebacks for unhashable exceptions (GH-4014) #4024
Files	unhashable_exceptions.diff: Patch against default branch issue28603-list.patch issue28603-ignore.patch issue28603-list.patch hashex.py

^{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 2017-10-17.22:17:00.634>
created_at = <Date 2016-11-03.21:22:14.375>
labels = ['3.7', 'type-bug', 'library']
title = "traceback module can't format/print unhashable exceptions"
updated_at = <Date 2017-10-17.22:17:00.632>
user = 'https://github.com/Trundle'

bugs.python.org fields:

activity = <Date 2017-10-17.22:17:00.632>
actor = 'serhiy.storchaka'
assignee = 'none'
closed = True
closed_date = <Date 2017-10-17.22:17:00.634>
closer = 'serhiy.storchaka'
components = ['Library (Lib)']
creation = <Date 2016-11-03.21:22:14.375>
creator = 'Trundle'
dependencies = []
files = ['45342', '45870', '45871', '45872', '47223']
hgrepos = []
issue_num = 28603
keywords = ['patch']
message_count = 11.0
messages = ['280022', '280233', '280237', '283081', '283095', '296034', '304485', '304534', '304535', '304538', '304539']
nosy_count = 7.0
nosy_names = ['terry.reedy', 'Trundle', 'berker.peksag', 'serhiy.storchaka', 'JelleZijlstra', 'Chris Barth', 'zaneb']
pr_nums = ['4014', '4024']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue28603'
versions = ['Python 3.6', 'Python 3.7']

[Trundle]

The traceback module tries to handle loops caused by an exception's __cause__ or __context__ attributes when printing tracebacks. To do so, it adds already seen exceptions to a set. Unfortunately, it doesn't handle unhashable exceptions:

>>> class E(Exception): __hash__ = None
...
>>> traceback.print_exception(E, E(), None)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/lib/python3.5/traceback.py", line 100, in print_exception
    type(value), value, tb, limit=limit).format(chain=chain):
  File "/usr/lib/python3.5/traceback.py", line 439, in __init__
    _seen.add(exc_value)
TypeError: unhashable type: 'E'

CPython's internal exception printing pretty much does the same, except it ignores any exception while operating on the seen set (see https://hg.python.org/cpython/file/8ee4ed577c03/Python/pythonrun.c#l813 ff).

Attached is a patch that makes the traceback module ignore TypeErrors while operating on the seen set. It also adds a (minimal) test.

[berkerpeksag]

Thanks for the patch. Out of curiosity, how did you get an unhashable exception? Is there a way to reproduce this without creating a custom exception and set __hash__ to None? In other words, what's your use case?

[Trundle]

It was reported as bug to a project that uses the traceback module (bpython/bpython#651). Parsley (https://pypi.python.org/pypi/Parsley) is at least one library that uses unhashable exceptions, although I guess it's by accident: it defines __eq__, but not __hash__ and that makes the exception unhashable under Python 3.

[JelleZijlstra]

I ran into this bug through Thrift-generated exception classes (also reported there as https://issues.apache.org/jira/browse/THRIFT-4002).

I've added a few potential solutions:

• bpo-28603-listset.patch turns the seen set into a list if hashing fails. However, this adds a lot of complexity, especially in C, and because seen is changed halfway through the recursion, we may end up showing an exception twice.
• bpo-28603-list.patch uses a list instead of a set for seen. This is theoretically slower, but in practice it seems unlikely that exception __cause__ and __context__ would nest deep enough for this to be an issue.
• bpo-28603-ignore.patch takes a similar approach to Trundle's patch and just gives up when the value is not hashable. This means we lose cause/context information for these exceptions.

I prefer bpo-28603-list.patch.

[serhiy-storchaka]

There is no bpo-28603-listset.patch.

[ChrisBarth]

There are now several patches for this problem, which also affects me. What are the next steps to get this resolved?

[zaneb]

I also encountered this issue, and I'd like to further note that it is extraordinarily difficult to debug without access to stderr. If your logging facility uses the traceback module then it will be unable to record both the original (unhashable) exception, and the TypeError that it triggers (since the original exception is its __context__). The effect is that execution of a thread appears to simply stop without rhyme or reason (in fact it is still executing, but that fact is not apparent from the logs). I'm attaching a short reproducer that demonstrates this effect.

I believe the trade-offs inherent in the attached patches (either stopping the chain at an unhashable exception, or using a less-efficient data structure) can be avoided by comparing for identity rather than equality. The purpose of the check is to avoid an infinite loop; whether the objects compare as equal is something that is up to the author of the class, and has no relevance here except insofar as that objects ought to compare equal to themselves.

I submitted a pull request (bpo-4014) with an implementation of that.

[terryjreedy]

As I understand it, the patch amounts to ignoring any custom __eq__ and __hash__ on an Exception class when printing tracebacks and, in effect, using the default id-based versions inherited from object, as is being assumed. This seems right to me in this context.

[serhiy-storchaka]

New changeset de86073 by Serhiy Storchaka (Zane Bitter) in branch 'master':
bpo-28603: Fix formatting tracebacks for unhashable exceptions (bpo-4014)
de86073

[serhiy-storchaka]

New changeset 2712247 by Serhiy Storchaka (Miss Islington (bot)) in branch '3.6':
[3.6] bpo-28603: Fix formatting tracebacks for unhashable exceptions (GH-4014) (bpo-4024)
2712247

[serhiy-storchaka]

Thank you for your contribution Zane.