fix miscellaneous numpy=2.0 errors

[keewis]

• towards ⚠️ Nightly upstream-dev CI failed ⚠️ #8091
• closes AttributeError: np.unicode_ was removed in the NumPy 2.0 release. Use np.str_ instead #8133

Edit: looking at the relevant numpy issues, it appears numpy will stay a bit unstable for the next few weeks / months. Not sure how quickly we should try to adapt to those changes.

[keewis]

not sure what to do with the float_ removal, but otherwise this should be ready for review/merging. That way, at least the import check in the upstream-dev CI succeeds.

[mathause]

LGTM - I think this can be merged?

Isn't np.float_ and float equivalent (https://numpy.org/doc/stable/user/basics.types.html)?

[kmuehlbauer]

Isn't np.float_ and float equivalent

Not for numpy 2.0, where it has been removed: https://numpy.org/devdocs/release/2.0.0-notes.html#numpy-2-0-python-api-removals.

Maybe just replacing np.float_ with np.float64 as suggested in the error message will fix it?

[mathause]

Right, thanks. I found no deprecation warning or error in our upstream-dev CI, though. Maybe that's still to come? Anyway, given the description that seems to be the right approach.

[headtr1ck]

Thanks for this PR.
Maybe we should add some type ignores to silence Mypy. Still not sure what the best approach is in such cases.

[dcherian]

Yeah we'll need to do something about the. mypy errors:

xarray/core/dataset.py:30: error: Cannot find implementation or library stub for module named "numpy.exceptions"  [import]
xarray/core/dataset.py:30: note: See https://mypy.readthedocs.io/en/stable/running_mypy.html#missing-imports
xarray/tests/test_dataset.py:20: error: Cannot find implementation or library stub for module named "numpy.exceptions"  [import]
xarray/tests/test_dataarray.py:18: error: Cannot find implementation or library stub for module named "numpy.exceptions"  [import]

[keewis]

Maybe just replacing np.float_ with np.float64 as suggested in the error message will fix it?

We can't, as we cannot know the default floating-point size in advance: as far as I remember, 32bit OSs (windows in particular) still have float32 as the default float. Instead, just using the python float as dtype might work?

Yeah we'll need to do something about the mypy errors:

I'd be tempted to just add numpy.exceptions.* to the list of ignored modules in pyproject.toml, but numpy does have a .pyi file for that, so not sure how to deal with the transition.

[keewis]

this should be ready for merging now. Note that there are lots of failing upstream-dev mypy errors and dask seems to have some internal inconsistency in its rolling implementation, but these can be fixed / discussed in separate issues / PRs.

[mathause]

numpy suggests replacing np.float_ with np.float64: https://numpy.org/devdocs/release/2.0.0-notes.html

[max-sixty]

FYI looking through the mypy upstream errors, they're almost all matplotlib (with a couple of numpy ones)

[headtr1ck]

FYI looking through the mypy upstream errors, they're almost all matplotlib (with a couple of numpy ones)

See #8030