issues

When doing a rolling mean on an array with no negative values, the result somehow contains negative values anyway. This shouldn't be possible, since the mean of non-negative values can never be zero. Further, it only happens when using the mean method directly, not when using reduce(np.mean) nor construct().mean().

MCVE Code Sample

Take the following xarray Dataarray

```Python import numpy as np import scipy as sp import xarray as xr

soundlen=10000 np.random.seed(1) noise = np.random.randn(soundlen) noise = sp.signal.hann(soundlen) noise2 = noise*2

xnoise = xr.DataArray(noise2, dims='temp', coords={'temp': np.arange(soundlen)}) print(xnoise.min()) ```

The result is 0. That is, it has no values less than 0.

Using reduce(np.mean) has no negative values, either, since the mean of non-negative values can never be negative:

```Python

print(xroll.reduce(np.mean).min())
<xarray.DataArray ()> array(2.90664355e-15) ```

Similarly, using mean through construct has no negative values:

Python print(xroll.construct('new').mean('new').min()) <xarray.DataArray ()> array(0.)

However, using the mean method directly does give negative values:

Python print(xroll.mean().min()) <xarray.DataArray ()> array(1.72090357e-15)

This mathematically shouldn't be possible.

Versions

xarray: 0.15.0 pandas: 0.25.3 numpy: 1.17.4 scipy: 1.4.1 netCDF4: 1.5.3 pydap: None h5netcdf: None h5py: 2.10.0 Nio: None zarr: None cftime: 1.0.4.2 nc_time_axis: None PseudoNetCDF: None rasterio: None cfgrib: None iris: None bottleneck: 1.3.1 dask: 2.11.0 distributed: 2.11.0 matplotlib: 3.1.3 cartopy: None seaborn: 0.10.0 numbagg: None setuptools: 44.0.0 pip: 20.0.2 conda: None pytest: 5.3.5 IPython: 7.12.0 sphinx: 2.4.3

Currently pandas supports a lot of str accessors that xarray doesn't. Many of them have useful functionality. I think it would be good if the xarray had more of these str accessors. I would be willing to begin working on this.

There seem to be three categories. One that has a one-to-one mapping between input and output (so one input element becomes one output element), and one that has a one-to-many mapping (one input element becomes multiple output elements), and one that has a many-to-one mapping (multiple input elements become one output element). Exactly how the one-to-many mapping should be handling, if at all, is an open issue. Some of the many-to-one mappings (specifically join) would need a dimension argument.

One-to-one

• [x] casefold(self)
• [x] normalize(self, form)

One-to-many

• [x] extract(self, pat[, flags, expand])
• [x] extractall(self, pat[, flags])
• [x] findall(self, pat[, flags])
• [x] get_dummies(self[, sep])
• [x] partition(self[, sep, expand])
• [x] rpartition(self[, sep, expand])
• [x] rsplit(self[, pat, n, expand])
• [x] split(self[, pat, n, expand])

Many-to-one

• [ ] cat(self[, others, sep, na_rep, join])
• [ ] join(self, sep)

In DataArray.copy, the _indexes attributes is not deep copied. After pull request #3840, this causes deleting a coordinate of a copy will also delete that coordinate from the original, even for deep copies.

MCVE Code Sample

```python a0 = xr.DataArray( np.array([[1, 2, 3], [4, 5, 6]]), dims=["y", "x"], coords={"x": ["a", "b", "c"], "y": [-1, 1]}, )

a1 = a0.copy() del a1.coords["y"]

xr.tests.assert_identical(a0, a0) ```

The result is:

``` xarray/testing.py:272: in _assert_internal_invariants _assert_dataarray_invariants(xarray_obj) xarray/testing.py:222: in _assert_dataarray_invariants _assert_indexes_invariants_checks(da._indexes, da._coords, da.dims)

indexes = {'x': Index(['a', 'b', 'c'], dtype='object', name='x')}, possible_coord_variables = {'x': <xarray.IndexVariable 'x' (x: 3)> array(['a', 'b', 'c'], dtype='<U1'), 'y': <xarray.IndexVariable 'y' (y: 2)> array([-1, 1])} dims = ('y', 'x')

def _assert_indexes_invariants_checks(indexes, possible_coord_variables, dims):
    assert isinstance(indexes, dict), indexes
    assert all(isinstance(v, pd.Index) for v in indexes.values()), {
        k: type(v) for k, v in indexes.items()
    }

    index_vars = {
        k for k, v in possible_coord_variables.items() if isinstance(v, IndexVariable)
    }
    assert indexes.keys() <= index_vars, (set(indexes), index_vars)

    # Note: when we support non-default indexes, these checks should be opt-in
    # only!
    defaults = default_indexes(possible_coord_variables, dims)

  assert indexes.keys() == defaults.keys(), (set(indexes), set(defaults))

E AssertionError: ({'x'}, {'y', 'x'})

xarray/testing.py:185: AssertionError ```

Expected Output

The test should pass.

Problem Description

Doing a deep copy should make a copy of everything. Changing a deep copy should not alter the original in any way.