issue_comments

I closed the PR by accident but I think I'll keep it closed, unless there is interest to get this in. Testing dtypes could be an interesting option, however,

• there are O(1000) tests that do not preserve dtype - so this would be an enormous cleanup required if we ever seriously wanted to do this
• some elements cannot preserve dtypes (str types in pandas index (not sure what the index refactor does here) and empty str arrays)

True, but for the problem that the diff repr is not shown I think it makes sense to go for this, e.g. the following works fine:

```python def test_x():

data_array1 = xr.DataArray(data=1 * unit_registry.m)
data_array2 = xr.DataArray(data=2 * unit_registry.m)

with warnings.catch_warnings():
    warnings.simplefilter("always")

    xr.testing.assert_allclose(data_array1, data_array2)

```

to really fix this we need some way to say "I really want to convert to a numpy.ndarray" (unfortunately, np.asarray seems to be commonly used as "ensure that this is a numpy.ndarray"). That's the same issue as with cupy.Array, which will raise in __array__.

yes, I agree that it makes sense to set warnings to "always" in the assert_* functions.

Thanks! That drove me nuts but I finally figured out the problem:

https://github.com/pydata/xarray/blob/5735e163bea43ec9bc3c2e640fbf25a1d4a9d0c0/xarray/tests/test_units.py#L29

promotes the warning to an error so diff_array_repr fails (because the units are stripped...) and that is what we see. That also fails on master - run the following in with pytest:

```python import pint import pytest import xarray as xr

pytestmark = [ pytest.mark.filterwarnings("error::pint.UnitStrippedWarning"), ]

unit_registry = pint.UnitRegistry(force_ndarray=True)

def test_y():

data_array1 = xr.DataArray(data=1 * unit_registry.m)
data_array2 = xr.DataArray(data=2 * unit_registry.m)

xr.testing.assert_allclose(data_array1, data_array2)

```

@keewis do you have an idea how to fix this? One idea is to supress warnings in diff_array_repr and diff_dataset_repr. Do we need to strip the units before sending them to these functions?

the difference is that the unit registry has force_ndarray_like enabled. This is important because a pint.Quantity wrapping a python scalar is not a duck array, so the unit gets stripped.

@keewis The following test is giving me trouble: pytest xarray/tests/test_units.py::test_align_dataarray[int64-10-data-compatible_unit]

https://github.com/pydata/xarray/blob/5735e163bea43ec9bc3c2e640fbf25a1d4a9d0c0/xarray/tests/test_units.py#L461

I now try to reproduce what it does but I cannot even manage that... I think the following is what the test does but it fails with DimensionalityError: Cannot convert from 'dimensionless' to 'meter' - can you see what I do wrong/ different from the test?

```python import numpy as np import pint import xarray as xr

unit_registry = pint.UnitRegistry()

dtype = np.int64

data_unit1 = unit_registry.m data_unit2 = unit_registry.mm

array1 = np.linspace(0, 10, 2 * 5).reshape(2, 5).astype(dtype) array2 = np.linspace(0, 8, 2 * 5).reshape(2, 5).astype(dtype)

x = np.arange(2) # * dim_unit1 y1 = np.arange(5) # * dim_unit1 y2 = np.arange(2, 7) # * dim_unit2

u1 = np.array([3, 5, 7, 8, 9]) # * coord_unit1 u2 = np.array([7, 8, 9, 11, 13]) # * coord_unit2

coords1 = {"x": x, "y": y1} coords2 = {"x": x, "y": y2}

data_array1 = xr.DataArray(data=array1 * data_unit1, coords=coords1, dims=("x", "y")) data_array2 = xr.DataArray(data=array2 * data_unit2, coords=coords2, dims=("x", "y"))

fill_value = 10

actual_a, actual_b = xr.align(data_array1, data_array2, join="outer", fill_value=fill_value * data_unit1)

```

The underlying problem is that when I run this test with check_dtype=True it raises an error (which is fine) but instead of a nice diff_repr it elevates the UnitStrippedWarning to an error and I have not the slightest clue why...

sometimes dask can get the dtype wrong before the compute (see below)

I suspect this is only for object arrays since you need not get back an object array? Maybe check in lazy_array_equiv only if all dtypes are not object