[BUG] Groupby collect on struct columns losing field name

[ayushdg]

Describe the bug
Groupby collect aggregation on a struct columns causes loss of the struct field names in the output

Steps/Code to reproduce bug

df = cudf.DataFrame(
    {
        'a':['aa','aa','cc'],
        'd':[{"b": '1', "c": "one"}, {"b": '2', "c": "two"}, {"b": '3', "c": "one"}]
     }
)

df
	a	d
0	aa	{'b': '1', 'c': 'one'}
1	aa	{'b': '2', 'c': 'two'}
2	cc	{'b': '3', 'c': 'one'}

df.groupby('a').collect()

	d
a	
aa	[{'0': '1', '1': 'one'}, {'0': '2', '1': 'two'}]
cc	[{'0': '3', '1': 'one'}]

Expected behavior

	d
a	
aa	[{'b': '1', 'c': 'one'}, {'b': '2', 'c': 'two'}]
cc	[{'b': '3', 'c': 'one'}]

Environment overview (please complete the following information)

• Docker nightly (21.08)

[charlesbluca]

Are we expecting libcudf's groupby agg to return the underlying struct column with its metadata? If not, the fix here might be to reapply the source StructColumn's metadata to the resulting column's elements before returning.

That being said, I think the underlying issue here is that we shouldn't be allowing for a collect agg on StructColumns in the first place; looking in groupby.pyx, we can see that no aggs should be possible for StructColumns:

cudf/python/cudf/cudf/_lib/groupby.pyx

Lines 42 to 51 in 0a4e8a1

	# The sets below define the possible aggregations that can be performed on
	# different dtypes. These strings must be elements of the AggregationKind enum.
	_CATEGORICAL_AGGS = {"COUNT", "SIZE", "NUNIQUE", "UNIQUE"}
	_STRING_AGGS = {"COUNT", "SIZE", "MAX", "MIN", "NUNIQUE", "NTH", "COLLECT",
	"UNIQUE"}
	_LIST_AGGS = {"COLLECT"}
	_STRUCT_AGGS = set()
	_INTERVAL_AGGS = set()
	_DECIMAL_AGGS = {"COUNT", "SUM", "ARGMIN", "ARGMAX", "MIN", "MAX", "NUNIQUE",
	"NTH", "COLLECT"}

But we erroneously use _STRING_AGGS for struct dtypes:

cudf/python/cudf/cudf/_lib/groupby.pyx

Line 139 in 0a4e8a1

else _STRING_AGGS if is_struct_dtype(dtype)

[charlesbluca]

Opened up an issue to address the underlying bug #8499; though in theory, we should have been able to fix this bug by simply applying the input column's metadata to the resulting list's elements. This might be worth exploring if we want to add the collect agg later on.