Nightly release failing with numpy 1.15.x

[effigies]

Bug report

Bug summary

The nightly release up on https://7933911d6844c6c53a7d-47bd50c35cd79bd838daf386af554a83.ssl.cf2.rackcdn.com/ is causing import errors with numpy 1.15.x. The problem resolves by upgrading to 1.16 or higher.

First detected on Travis CI in nipy/nipype#3152, but reproduced easily locally.

Code for reproduction

Using conda, install numpy 1.15.x and the latest nightly build:

conda create -n numpy115_mpl python=3.7 numpy=1.15
conda activate numpy115_mpl
pip install https://7933911d6844c6c53a7d-47bd50c35cd79bd838daf386af554a83.ssl.cf2.rackcdn.com/matplotlib-3.2.0rc2+1180.ga8ed66d0f-cp37-cp37m-macosx_10_9_x86_64.whl\#sha256\=fa2c4f132fbd257a1ecab26ce61a6789584b63581cffbeec5c1a4d0f8e24cabe

Attempt to import:

python -c 'import matplotlib'

Actual outcome

ModuleNotFoundError: No module named 'numpy.core._multiarray_umath'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/System/Volumes/Data/anaconda3/envs/numpy115_mpl/lib/python3.7/site-packages/matplotlib/__init__.py", line 187, in <module>
    _check_versions()
  File "/System/Volumes/Data/anaconda3/envs/numpy115_mpl/lib/python3.7/site-packages/matplotlib/__init__.py", line 172, in _check_versions
    from . import ft2font
ImportError: numpy.core.multiarray failed to import

Expected outcome

Matplotlib version

• Operating system: OSX
• Matplotlib version: 3.2.0rc2+1180.ga8ed66d0f
• Matplotlib backend (print(matplotlib.get_backend())): n/a
• Python version: 3.7.6
• Jupyter version (if applicable):
• Other libraries: numpy 1.15

[tacaswell]

I don't think that numpy guarantees that things complied with newer versions of numpy (but do guarantee that things compiled with old numpy work with new numpy). It looks like those wheels are built with numpy 1.18 (https://travis-ci.org/MacPython/matplotlib-wheels/jobs/637581926#L962).

Do the nightly matplotlib wheels work with newer versions of numpy? I am not convinced that we want to worry about pinning back the build dependencies on nightly wheel build.

[effigies]

Gotcha. If your release wheels are built with the oldest supported numpy, then I don't think there's a problem.

I've switched our --pre test to also use --upgrade, to ensure that we're testing all libraries with the newest version (what I was intending anyway...), so we're not going to hit this again unless the next matplotlib release also compiles with numpy 1.18.

[tacaswell]

Our release wheels should be built with the oldest supported version of numpy. I'm going to close this as no-action, thank you for reporting!