hooks: matplotlib: assert that mpl_data_dir is not empty

[rokm]

If the statement in exec_statement() fails for whatever reason, mpl_data_dir ends up empty, which in turn cases the current working directory being collected as matplotlib/mpl-data.

This leads to difficult-to-debug issues (see confusion around .git directory under #5504), so it's better to just signal the error.

[bwoodsend]

Hmm, looks like there's something washed up by this. I notice that this failing job is the only one that installs matplotlib from source.

[rokm]

Hmm, looks like there's something washed up by this. I notice that this failing job is the only one that installs matplotlib from source.

Ah, so it's building it from source.

The test itself ends up with empty mpt_data_dir because matplotlib fails to import due incompatibility in numpy:

 ------- Starting build. -------
RuntimeError: module compiled against API version 0xe but this version of numpy is 0xd
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/Users/runner/hostedtoolcache/Python/3.9.2/x64/lib/python3.9/site-packages/matplotlib/__init__.py", line 207, in <module>
    _check_versions()
  File "/Users/runner/hostedtoolcache/Python/3.9.2/x64/lib/python3.9/site-packages/matplotlib/__init__.py", line 192, in _check_versions
    from . import ft2font
ImportError: numpy.core.multiarray failed to import

Search results for module compiled against API version 0xe but this version of numpy is 0xd seem to suggest that the problem lies in pip not honoring the pinned version of build dependencies: pypa/pip#9542.

So this should probably go away if we pin numpy to the latest available version.

[rokm]

Or maybe if we get the CI to install pinned version of numpy before matplotlib.

[bwoodsend]

Ughh pypa/pip#9542 looks hellish. I'd rather just re-pin matplotlib?

[rokm]

Ughh pypa/pip#9542 looks hellish. I'd rather just re-pin matplotlib?

You're right. Re-pinning matplotlib to 3.3.4 should probably also do the trick as it has wheels for all python versions against which we test. I'll push a commit once AppVeyor finishes its current round.

[AppVeyorBot]

✅ Build pyinstaller 1.0.965 completed (commit c137ed5a81 by @rokm)

[bwoodsend]

white_check_mark Build pyinstaller 1.0.965 completed (commit c137ed5 by @rokm)

Yipee it worked!