I moved the check to the correct position in the method and add ConftestImportFailure to the Union. However, I could not remove the second #type: ignore as ExceptionInfo does not accept a tuple of the form Tuple[Type[Exception], Exception, TracebackType], which excinfo.value.excinfo is.

Thanks for the feedback @bluetech! I applied the suggested changes.

The code LGTM, just two things:

• The tests fail on Python 3.5 -- seems like the exception type is different there.

• According to the "patch" coverage report, the if isinstance(excinfo.value, ConftestImportFailure): branch isn't covered, which means the tests don't exercise it. Is it wrong or are the tests checking something else?

Also, did you see my comment about code sharing above? (It's OK if you don't want to dig into that)

Thanks for the feedback!

• I don't know why it's failing, but I saw a bunch of PRs failing on Python 3.5 but got merged.
• The test covers exactly than bit of code, but I do not know why coverage flags it as not covered.
• Yes, I had a look at it, but wasn't able to understand it in a way to be able to extend it and share the code (sorry for that)

I don't know why it's failing, but I saw a bunch of PRs failing on Python 3.5 but got merged.

The CI is gated on Python 3.5, i.e. the merge button is disabled if it is not passing.

Looking at the CI logs, the issue is that the test asserts for

E   ModuleNotFoundError: No module named 'unknown'

but on Python 3.5 it is

E   ImportError: No module named 'unknown'

so one way to take care of it is to assert just for

E   *Error: No module named 'unknown'

The test covers exactly than bit of code, but I do not know why coverage flags it as not covered.

Yes that's weird. I verified now that the if is indeed entered by the test.

Fixed the py35 build.

@asottile What do you think about the patch? Can it be merged?