BUG/EA: groupby on an EA should return the EA type

[jreback]

In [1]: df = pd.DataFrame({'Int': pd.Series([1, 2, 3], dtype='Int64'), 'A': [1, 2, 1]})
   ...: df
   ...: 
Out[1]: 
  Int  A
0   1  1
1   2  2
2   3  1

In [2]: df.groupby('A').Int.sum()
Out[2]: 
A
1    4
2    2
Name: Int, dtype: int64

[2] should be of type Int64; IOW this aggregation needs to be passed thru to the EA _from_sequence

[5hirish]

@jreback Hi working on this issue; did some code walkthrough, learned that here https://github.com/pandas-dev/pandas/blob/master/pandas/core/groupby/groupby.py#L834 in the method _python_agg_general at result, counts = self.grouper.agg_series(obj, f) the obj type is Int64 but the result type is int64. So, I checked the agg_series method, seems like a Cython converter converts it from object to int64 after calculating the aggregation result. So, I am kind of lost as to where I should put _from_squence after checking if it's of type ExtensionArray

[jreback]

so look at _wrap_aggregated_output

there are a number of paths here though

[5hirish]

@jreback

1. Tests under pytest pandas/tests/arrays/ like test_integer.py/test_preserve_dtypes, test_integer.py/test_reduce_to_float are failing because:

AssertionError: Attributes are different
Attribute "dtype" are different
     [left]:  Int64      <-- result
     [right]: float64  <-- expected

there is custom datatype Int64 in the test input data of the above tests and test expects it to be converted to int64 or float64. Does pandas consider Int64 as int64, cause Int64 satisfies both the following conditions: if is_extension_array_dtype(dtype): and if numeric_only and is_numeric_dtype(dtype) or not numeric_only ?

2. Tests under pandas/tests/sparse/ like test_groupby.py/test_groupby_includes_fill_value are failing because:

AssertionError: Attributes are different           
Attribute "dtype" are different
    [left]:  Sparse[float64, nan]   <-- result
    [right]: float64                       <-- expected

[TomAugspurger]

@5hirish probably best to keep the conversation in the PR now that it's open.