adapt to NEP 51

[keewis]

With NEP 51 (and the changes to numpy main), scalar types no longer pretend to be standard python types in their string representation. This fixes most of the errors in the tests but there are still a few remaining in the doctests (in particular, doctests for private plotting utils).

• towards ⚠️ Nightly upstream-dev CI failed ⚠️ #8091

[dcherian]

Should we keep the numpy representation since that's more accurate?

We could use pytest-accept to update the tests

[keewis]

Should we keep the numpy representation since that's more accurate?

I wouldn't think so, but I might be wrong: we already have the dtype information right beside it so it wouldn't add much, and they are usually taken out of arrays and are only formatted that way because we iterate over the array's values.

That said, there are definitely a few instances where we actually do return numpy scalars in the doctests that should be updated (although that means we'd have to wait until we adopt numpy>1.25 (numpy=2.0?), and accept that you have to use at least that version to run the doctests until we can discontinue support for numpy<2.0)

[headtr1ck]

That said, there are definitely a few instances where we actually do return numpy scalars in the doctests that should be updated (although that means we'd have to wait until we adopt numpy>1.25 (numpy=2.0?), and accept that you have to use at least that version to run the doctests until we can discontinue support for numpy<2.0)

You could also parse them to python floats/ints before displaying them. This would work with all numpy versions.

[keewis]

You could also parse them to python floats/ints before displaying them.

Cast, you mean?

[headtr1ck]

Simply value = float(da.item()). This will display the same for all np versions.

[keewis]

Simply value = float(da.item())

The float wouldn't even be necessary, arr.item() (and arr.tolist()) always return python types.

However, for some we're not the ones who issue those (an example: fg.axs[0, 0].get_xlim() in xarray.plot.facetgrid.FacetGrid._set_lims returns a tuple of two numpy.floats, but get_xlim is from matplotlib).

According to numpy/numpy#24470, the recommendation is to test a single numpy version, so I'd worry about those once numpy=2.0 has been released (we could also switch to a more advanced doctest runner like pytest-doctestplus, but that should be a new PR).

[keewis]

this should now be ready for merging, and the rolling failures is an issue upstream in dask (see #8091)