Improved round tripping for attrs

[hmaarrfk]

Is your feature request related to a problem?

Consider the following dataset:

import xarray as xr
dataset = xr.Dataset()
dataset.attrs = {
    'empty_attribute': [],
    'attribute_of_length_1': ['one_item_only'],
    'attribute_of_length_2': ['one_item', 'two_items'],
}
dataset.to_netcdf('foo.nc')

loaded = xr.open_dataset('foo.nc')

from pprint import pprint
pprint(loaded.attrs)

pprint(loaded.attrs)
{'attribute_of_length_1': 'one_item_only',
 'attribute_of_length_2': ['one_item', 'two_items'],
 'empty_attribute': array([], dtype=float64)}

1. The attrs that contains an empty list has become an array of type float
2. The attrs that contains a list of one element, has been downcast to a scalar
3. The attrs that contains two elements has been kept as a list of strings.

Describe the solution you'd like

A away to ensure that lists stay lists.

Does this option already exists?

Describe alternatives you've considered

Writing nice error prone if statements

Additional context

I tried to look through https://github.com/pydata/xarray/blob/main/xarray/tests/test_dataset.py and didn't find anything that was testing this special behavior.

[hmaarrfk]

With the default engine

 h5dump foo.nc
HDF5 "foo.nc" {
GROUP "/" {
   ATTRIBUTE "_NCProperties" {
      DATATYPE  H5T_STRING {
         STRSIZE 34;
         STRPAD H5T_STR_NULLTERM;
         CSET H5T_CSET_ASCII;
         CTYPE H5T_C_S1;
      }
      DATASPACE  SCALAR
      DATA {
      (0): "version=2,netcdf=4.9.2,hdf5=1.14.3"
      }
   }
   ATTRIBUTE "attribute_of_length_1" {
      DATATYPE  H5T_STRING {
         STRSIZE 13;
         STRPAD H5T_STR_NULLTERM;
         CSET H5T_CSET_ASCII;
         CTYPE H5T_C_S1;
      }
      DATASPACE  SCALAR
      DATA {
      (0): "one_item_only"
      }
   }
   ATTRIBUTE "attribute_of_length_2" {
      DATATYPE  H5T_STRING {
         STRSIZE H5T_VARIABLE;
         STRPAD H5T_STR_NULLTERM;
         CSET H5T_CSET_UTF8;
         CTYPE H5T_C_S1;
      }
      DATASPACE  SIMPLE { ( 2 ) / ( 2 ) }
      DATA {
      (0): "one_item", "two_items"
      }
   }
   ATTRIBUTE "empty_attribute" {
      DATATYPE  H5T_IEEE_F64LE
      DATASPACE  NULL
      DATA {
      }
   }
}
}

With specifying the engine as h5netcdf:

$ h5dump foo_h5netcdf.nc
HDF5 "foo_h5netcdf.nc" {
GROUP "/" {
   ATTRIBUTE "_NCProperties" {
      DATATYPE  H5T_STRING {
         STRSIZE 48;
         STRPAD H5T_STR_NULLPAD;
         CSET H5T_CSET_ASCII;
         CTYPE H5T_C_S1;
      }
      DATASPACE  SCALAR
      DATA {
      (0): "version=2,h5netcdf=1.6.1,hdf5=1.14.3,h5py=3.13.0"
      }
   }
   ATTRIBUTE "attribute_of_length_1" {
      DATATYPE  H5T_STRING {
         STRSIZE H5T_VARIABLE;
         STRPAD H5T_STR_NULLTERM;
         CSET H5T_CSET_UTF8;
         CTYPE H5T_C_S1;
      }
      DATASPACE  SIMPLE { ( 1 ) / ( 1 ) }
      DATA {
      (0): "one_item_only"
      }
   }
   ATTRIBUTE "attribute_of_length_2" {
      DATATYPE  H5T_STRING {
         STRSIZE H5T_VARIABLE;
         STRPAD H5T_STR_NULLTERM;
         CSET H5T_CSET_UTF8;
         CTYPE H5T_C_S1;
      }
      DATASPACE  SIMPLE { ( 2 ) / ( 2 ) }
      DATA {
      (0): "one_item", "two_items"
      }
   }
   ATTRIBUTE "empty_attribute" {
      DATATYPE  H5T_IEEE_F64LE
      DATASPACE  SIMPLE { ( 0 ) / ( 0 ) }
      DATA {
      }
   }
}
}

however, even though the difference DATASPACE SCALAR vs DATASPACE SIMPLE exists, it doesn't affect how the data is read in.