Systematize calculation testing

[dopplershift]

It would probably be good to find some way to make our calculation testing more systematic, in terms of testing certain features:

• Work with:
  • arrays (handled by Add a fixture that provides numpy, masked, and dask array functions #1935)
  • scalars
  • masked arrays (handled by Add a fixture that provides numpy, masked, and dask array functions #1935)
  • NaNs
• Don't mutate inputs in-place
• Return scalars as appropriate (?)
• For functions that deal with temperature (at least anywhere we manually "handle" units), verify that calculations work with degC/degF

This was inspired by #1209 and other issues. I'm not sure how to go about it. PyTest parameterization and/or fixtures could be useful, but I'm not sure how to handle encoding truth values with those. I suppose those could be smoke tests for those input cases, though checking behavior with NaNs and masked arrays could need some truth values for correctness.

[jthielen]

With regards to mutating inputs in-place, I've opened hgrecco/pint#925 as an upstream issue.

[rpmanser]

Fixtures could be made for each data type with meteorological variables within them and parameterizing could take care of edge cases.

For example, in the conftest.py file there would be:

@pytest.fixture
def arrays():
    return {
        'u': np.array([4., 2., 0., 0.]) * units('m/s'),
        'v': np.array([0., 2., 4., 0.]) * units('m/s'),
        # More variables here...
    }

Then a test would accept one of the data type fixtures and select the variables that are relevant to the function being tested:

def test_direction_array(arrays):
    """Test calculating wind direction."""
    direc = wind_direction(arrays['u'], arrays['v'])
    true_dir = np.array([270., 225., 180., 0.]) * units.deg

    assert_array_almost_equal(true_dir, direc, 4)

Something similar could be done for scalars, masked arrays, and NaNs.

Then for edge cases, parametrize could be used.

@pytest.mark.parametrize(
    "u, v, true_dir",
    [
        # Edge case for northerly and calm winds
        pytest.param(
            np.array([0., -0., 0.]) * units('m/s'),
            np.array([0., 0., -5.]) * units('m/s'),
            np.array([0., 0., 360.]) * units.deg
        )
        # Next edge case here
    ]
)
def test_direction(u, v, true_dir):
    """Test edge cases for wind direction"""
    direc = wind_direction(u, v)

    assert_array_almost_equal(true_dir, direc, 4)

I'm not sure how to handle truth values except for defining them in the test function, which is how everything is set up currently. There are also many tests where exceptions are raised, or specific units are being checked, which would probably remain untouched.

Anyway, would the fixtures be a good way to go about systematically testing data types in each function? And are there thoughts on parameterizing edge cases?

[dopplershift]

Adding another item to the list: For functions that deal with temperature (at least anywhere we manually "handle" units), verify that calculations work with degC/degF.