Running doctest examples on a class results in OSError in REPL

[stenwessel]

Bug report

Bug description:

In the REPL, running doctests contained in the docstring of a class is broken since version 3.10. Example:

import doctest

class A:
    """Docstring of A, with tests.
    >>> 1 + 2
    3
    """
    pass

doctest.run_docstring_examples(A, globals(), name='A', verbose=True)

In version 3.9, this code works and executes the doctest in the docstring of the class A. In version 3.10 (and onward), the call to doctest.run_docstring_examples fails with OSError: source code not available.

This only fails in the REPL and not in regular Python files. In particular, this also fails in (Jupyter) notebook environments.

Cause

This change in behavior is caused by 48a6255 as a fix for #88814. Previously, inspect.getsourcefile() could only fail with TypeError, but this can now also be OSError. The new exception type is not being handled by doctest, only TypeError is caught:

cpython/Lib/doctest.py

Lines 924 to 927 in 713e428

	try:
	file = inspect.getsourcefile(obj)
	except TypeError:
	source_lines = None

The call to inspect.getsourcefile() is only used by doctest to determine line numbers, if available. Thus, failing to find the source file can safely fail silently in environments where line numbers are not available (such as the REPL).

Proposed changes

• Also catch OSError in the doctest finder (in the linked snippet).
• Update the documentation to reflect that inspect.getsourcefile() can also fail with OSError.

An alternative would be to change the exception type in 48a6255 to also be TypeError in the case the source file is not available. I do not think this a desired change.

I am happy to contribute a pull request with these changes.

CPython versions tested on:

3.9, 3.10, 3.11

Operating systems tested on:

Linux

Linked PRs

• gh-113329: Catch OSError in doctest finder and document that inspect.getsourcefile/getfile can raise OSError #113335

[gaogaotiantian]

This makes sense to me as we documented inspect.getsourcelines that it could raise OSError. The changes to doctest is reasonable to me as well.

I think the change should not be too difficult, so maybe you can make the PR if you want? I'm fairly confident that the doctest fix would be accepted, even if it's not, the documentation should definitely be updated.

[serhiy-storchaka]

Thank you for your report and PR @stenwessel. It LGTM in general if we keep OSError (we need also to add versionchanged directives and a What's New entry). But I wonder whether changing the exception type to TypeError is a better solution. I do not feel that OSError fits here.

[serhiy-storchaka]

@ambv, please take your attention.

[stenwessel]

Hi @serhiy-storchaka, I agree that the OSError might be an undesired exception type in this case, especially since this may break functionality in more places than just this doctest issue.

On the other hand, inspect.getsourcelines and inspect.getsource do fail with OSError as well when the source code cannot be found since version 3.3 (and this is documented as such). Additionally changing to TypeError will break code that catches TypeError assuming this means builtin class, while this is not necessarily the case.

Either way, I am of course happy to change the PR when necessary (also with the other updates for e.g. documentation and release notes, as I am not entirely sure what is required)