I think I've changed my opinion since this morning, and now having this worked around in astropy seems a more right solution, after all collection without optional dependencies should work there, thus this "feature" recovered a few bugs already.

As discussed offline, I have no strong feeling either way.

I think the correct fix is to collect doctests without importing files, instead e.g. parsing files with ast or something similar to find docstrings.

I think the correct fix is to collect doctests without importing files, instead e.g. parsing files with ast or something similar to find docstrings.

How does the doctest that ships with pytest do it? Will we risk deviating too much from pytest?

pytest's doctest does the exact same we had here for 0.6.0 in #103:

                module = self.fspath.pyimport()
            except ImportError:
                if self.config.getvalue("doctest_ignore_import_errors"):
                    pytest.skip("unable to import module %r" % self.fspath)
                else:
                    raise

Sorry, my question in #108 (comment) is a reply to #108 (comment) . (I clarified my comment above.)

oh, I assumed the first one was a question.

Let's go ahead with this one for now, and maybe #110 for the future.

I'm merging this now, but they may be some more outstanding issues, as using this with pytest 5.4, there are still asdf related ModuleNotFoundErrors popping up.

Thank you @pllim!

I haven't caught up with all the notifications yet but I am hoping that maybe astropy/astropy#10260 will handle it. If not, perhaps the last resort is to have setup.cfg ignore it on the astropy side. 🤞