Upstream CI is failing

[Illviljan]

What happened?

The upstream CI has been failing for a while. Here's the latest:
https://github.com/pydata/xarray/actions/runs/5501368493/jobs/10024902009#step:7:16

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/home/runner/work/xarray/xarray/xarray/__init__.py", line 1, in <module>
    from xarray import testing, tutorial
  File "/home/runner/work/xarray/xarray/xarray/testing.py", line 7, in <module>
    import numpy as np
ModuleNotFoundError: No module named 'numpy'

Digging a little in the logs

  Installing build dependencies: started
  Installing build dependencies: finished with status 'error'
  error: subprocess-exited-with-error
  
  × pip subprocess to install build dependencies did not run successfully.
  │ exit code: 1
  ╰─> [3 lines of output]
      Looking in indexes: https://pypi.anaconda.org/scipy-wheels-nightly/simple
      ERROR: Could not find a version that satisfies the requirement meson-python==0.13.1 (from versions: none)
      ERROR: No matching distribution found for meson-python==0.13.1
      [end of output]

Might be some numpy problem?

Should the CI be robust enough to handle these kinds of errors? Because I suppose we would like to get the automatic issue created anyway?

[keewis]

actually, no, it's not numpy but pandas that caused that issue... we're installing pandas, numpy, matplotlib, and scipy from the scipy-wheels-nightly repository, and apparently pip selects the source distribution of pandas and fails because the build dependencies cannot be satisfied.

I don't know why it selects the source dist, but the fix might be easy: switch to scientific-python-nightly-wheels, because apparently pandas doesn't seem to publish to scipy-wheels-nightly anymore?

[keewis]

as for making CI emit an error in that case: yes, that would be ideal, but then we'd need to rethink our test workflow... we currently have a step that fails early if the import doesn't succeed (because running pytest on the entire testsuite is expensive, even if we're only collecting tests), but that obviously does not create pytest log files.