Decode_cf fails when scale_factor is a length-1 list

[rabernat]

Some datasets I work with have scale_factor and add_offset encoded as length-1 lists. The following code worked as of Xarray 0.16.1

import xarray as xr
ds = xr.DataArray([0, 1, 2], name='foo',
                  attrs={'scale_factor': [0.01],
                         'add_offset': [1.0]}).to_dataset()
xr.decode_cf(ds)

In 0.16.2 (just released) and current master, it fails with this error

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-2-a0b01d6a314b> in <module>
      2                   attrs={'scale_factor': [0.01],
      3                          'add_offset': [1.0]}).to_dataset()
----> 4 xr.decode_cf(ds)

~/Code/xarray/xarray/conventions.py in decode_cf(obj, concat_characters, mask_and_scale, decode_times, decode_coords, drop_variables, use_cftime, decode_timedelta)
    587         raise TypeError("can only decode Dataset or DataStore objects")
    588 
--> 589     vars, attrs, coord_names = decode_cf_variables(
    590         vars,
    591         attrs,

~/Code/xarray/xarray/conventions.py in decode_cf_variables(variables, attributes, concat_characters, mask_and_scale, decode_times, decode_coords, drop_variables, use_cftime, decode_timedelta)
    490             and stackable(v.dims[-1])
    491         )
--> 492         new_vars[k] = decode_cf_variable(
    493             k,
    494             v,

~/Code/xarray/xarray/conventions.py in decode_cf_variable(name, var, concat_characters, mask_and_scale, decode_times, decode_endianness, stack_char_dim, use_cftime, decode_timedelta)
    333             variables.CFScaleOffsetCoder(),
    334         ]:
--> 335             var = coder.decode(var, name=name)
    336 
    337     if decode_timedelta:

~/Code/xarray/xarray/coding/variables.py in decode(self, variable, name)
    271             dtype = _choose_float_dtype(data.dtype, "add_offset" in attrs)
    272             if np.ndim(scale_factor) > 0:
--> 273                 scale_factor = scale_factor.item()
    274             if np.ndim(add_offset) > 0:
    275                 add_offset = add_offset.item()

AttributeError: 'list' object has no attribute 'item'

I'm very confused, because this feels quite similar to #4471, and I thought it was resolved #4485.
However, the behavior is different with 'scale_factor': np.array([0.01]). That works fine--no error.

How might I end up with a dataset with scale_factor as a python list? It happens when I open a netcdf file using the h5netcdf engine (documented by @gerritholl in #4471 (comment)) and then write it to zarr. The numpy array gets encoded as a list in the zarr json metadata. 🙃

This problem would go away if we could resolve the discrepancies between the two engines' treatment of scalar attributes.

[dcherian]

I guess we need np.asarray(scale_factor).item()

[rabernat]

But what did we do before?

[mathause]

I think it just did np.array([1, 2, 3.]) * [5] numpy coverts the [5] to an array with one element, which is why it worked:

xr.coding.variables._scale_offset_decoding(np.array([1, 2, 3.]), [5], None, np.float)

xarray/xarray/coding/variables.py

Line 217 in 49d03d2

def _scale_offset_decoding(data, scale_factor, add_offset, dtype):

[rabernat]

Ok then I am 👍 on @dcherian's solution.