pint support for Dataset

[keewis]

This is part of the effort to add support for pint (see #3594) to Dataset objects (although it will probably be a test-only PR, just like #3643).

• Tests added
• Passes isort -rc . && black . && mypy . && flake8
• Fully documented, including whats-new.rst for all changes and api.rst for new API

The list of failing tests from #3594:

• Dataset methods
  • __init__: Needs unit support in IndexVariable, and merge does not work yet (test bug is also possible)
  • aggregation: xarray does not implement __array_function__ (see running numpy functions on xarray objects #3917)
  • rank: depends on bottleneck and thus only works with numpy.array
  • ffill, bfill: uses bottleneck
  • interpolate_na: uses numpy.vectorize, which does not support NEP-18, yet
  • equals, identical: works (but no units / unit checking in IndexVariable)
  • broadcast_like: works (but no units / unit checking in IndexVariable)
  • to_stacked_array: no units in IndexVariable
  • sel, loc: no units in IndexVariable
  • interp, reindex: partially blocked by IndexVariable. reindex works with units in data, but interp uses scipy
  • interp_like, reindex_like: same as interp / reindex
  • quantile: works, but needs pint >= 0.12
  • groupby_bins: needs pint >= 0.12 (for isclose)
  • rolling: uses numpy.lib.stride_tricks.as_strided
  • rolling_exp: uses numbagg (supports NEP-18, but pint doesn't support its functions)

[keewis]

it seems assert_allclose was one of the sources of UnitStrippedWarnings, but since there's a bug in pint's isclose (fixed on master) the tests fail now.

Edit: pint will be released in the next few days, so most of the failing CI should pass after that.

Also, because it casted to numpy, there were a few bugs that were hidden.

[dcherian]

Thanks for working on this @keewis . Since there are no changes outside test_units.py, I think you should merge whenever you think this is ready.

[keewis]

I just found another issue: pint implements prod (but not yet nanprod) so the prod tests could be un-xfailed. However, we define a custom nanprod function that uses where to replace nan with 1. This won't work on quantities since, unlike nan and 0, a bare 1 cannot be put into quantities with a dimension (i.e. with a unit other than dimensionless).

I don't really understand the purpose of nanprod's min_count parameter (and _maybe_null_out) so I'm not sure how to fix that.

For now, I think we can merge this PR on green and I'll add that issue to the list in #3594.

[dcherian]

Looks like an incompatibility with latest pandas

____________________ TestDataset.test_resample[int-coords] _____________________

self = <xarray.tests.test_units.TestDataset object at 0x7fa84fb9e490>
variant = 'coords', dtype = <class 'int'>
    return func(*all_args, **all_kwargs)
xarray/core/common.py:1123: in resample
    grouper = pd.Grouper(
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

cls = <class 'pandas.core.groupby.grouper.Grouper'>, args = ()
kwargs = {'base': 0, 'closed': None, 'freq': '6m', 'label': None, ...}
TimeGrouper = <class 'pandas.core.resample.TimeGrouper'>, stacklevel = 2

    def __new__(cls, *args, **kwargs):
        if kwargs.get("freq") is not None:
            from pandas.core.resample import TimeGrouper
    
            # Deprecation warning of `base` and `loffset` since v1.1.0:
            # we are raising the warning here to be able to set the `stacklevel`
            # properly since we need to raise the `base` and `loffset` deprecation
            # warning from three different cases:
            #   core/generic.py::NDFrame.resample
            #   core/groupby/groupby.py::GroupBy.resample
            #   core/groupby/grouper.py::Grouper
            # raising these warnings from TimeGrouper directly would fail the test:
            #   tests/resample/test_deprecated.py::test_deprecating_on_loffset_and_base
    
            # hacky way to set the stacklevel: if cls is TimeGrouper it means
            # that the call comes from a pandas internal call of resample,
            # otherwise it comes from pd.Grouper
            stacklevel = 4 if cls is TimeGrouper else 2
            if kwargs.get("base", None) is not None:
>               warnings.warn(
                    "'base' in .resample() and in Grouper() is deprecated.\n"
                    "The new arguments that you should use are 'offset' or 'origin'.\n"
                    '\n>>> df.resample(freq="3s", base=2)\n'
                    "\nbecomes:\n"
                    '\n>>> df.resample(freq="3s", offset="2s")\n',
                    FutureWarning,
                    stacklevel=stacklevel,
                )
E               FutureWarning: 'base' in .resample() and in Grouper() is deprecated.
E               The new arguments that you should use are 'offset' or 'origin'.
E               
E               >>> df.resample(freq="3s", base=2)
E               
E               becomes:
E               
E               >>> df.resample(freq="3s", offset="2s")

[keewis]

yeah, I don't know how to only filter for pint warnings, I tried pytest.mark.filterwarnings("error:::pint[.*]") but that doesn't work

Edit: pytest.mark.filterwarnings("error::pint.UnitStrippedWarning") works so I'm merging.