FIX: _safe_first_finite on all non-finite array

[rcomer]

PR Summary

This function used to be named _safe_first_non_none (#23751), and at v3.6.0 we have

import numpy as np
import matplotlib.cbook

arr = np.full(5, np.nan)
print(matplotlib.cbook._safe_first_non_none(arr))

nan

So this PR reinstates previous behaviour. Currently on main, _safe_first_finite raises StopIteration when passed an all-nan array.

Fixes #18294 and fixes #24818. The examples from both those issues now successfully produce empty plots. The example from #18294 no longer throws the originally reported warning, but the example from #24818 does throw

/home/ruth/git_stuff/matplotlib/lib/matplotlib/axes/_axes.py:1185: RuntimeWarning: All-NaN axis encountered
  miny = np.nanmin(masked_verts[..., 1])
/home/ruth/git_stuff/matplotlib/lib/matplotlib/axes/_axes.py:1186: RuntimeWarning: All-NaN axis encountered
  maxy = np.nanmax(masked_verts[..., 1])

PR Checklist

Documentation and Tests

• Has pytest style unit tests (and pytest passes)
• [N/A] Documentation is sphinx and numpydoc compliant (the docs should build without error).
• [N/A] New plotting related features are documented with examples.

Release Notes

• [N/A] New features are marked with a .. versionadded:: directive in the docstring and documented in doc/users/next_whats_new/
• [N/A] API changes are marked with a .. versionchanged:: directive in the docstring and documented in doc/api/next_api_changes/
• [N/A] Release notes conform with instructions in next_whats_new/README.rst or next_api_changes/README.rst

[rcomer]

I see that a different approach was taken to handle the change in behaviour for bar #24149.

[oscargus]

I think that this approach is quite a bit simpler than the one used in #24149. On the other hand it is a bit more explicit what happens there.

My suggestions:

• Pick the first element of the array instead of np.nan (may be all np.inf and then returning np.nan is not so obvious...)
• Update the doc-string and state what happens if no parameter is finite
• Update the doc-string to replace skip_none with skip_nonfinite (old issue)
• (Change bar to use the new method?)

[rcomer]

Thanks @oscargus your suggestions make sense to me. I’ll put this in draft until I have time to implement them.

[jklymak]

I think there a few other places where StopIteration is try/excepted for _safe_first_element? Did you check all those?

[rcomer]

I hadn't, but I have now. I only found these two:

matplotlib/lib/matplotlib/dates.py

Lines 1828 to 1831 in 7103779

	try:
	x = cbook._safe_first_finite(x)
	except (TypeError, StopIteration):
	pass

matplotlib/lib/matplotlib/units.py

Lines 182 to 185 in 7103779

	try: # If cache lookup fails, look up based on first element...
	first = cbook._safe_first_finite(x)
	except (TypeError, StopIteration):
	pass

These both predate the change that led to StopIteration being thrown for all-nan arrays. The commit that introduced the first says it was for zero length objects, and these will still trigger a StopIteration when the logic arrives here. I'm not sure about the second - should units.Registry.get_converter also be able to handle zero length objects?

[jklymak]

I don't really know - just noted that it happens a few other places, and the more logic we can move into a consistent helper function, the better in my opinion.

[rcomer]

the more logic we can move into a consistent helper function, the better in my opinion.

I certainly agree with that. I wonder if further consolidation should wait for a separate PR though, given recent clarifications about what should and should not be backported

• my current change fixes a regression from v3.6, so it would be good to get in a patch release
• a change that tidies the code but (hopefully) makes no difference to the user should maybe wait till v3.8