Fix pd.merge to preserve ExtensionArrays dtypes (pandas-dev#20745)

pandas/core/dtypes/common.py

@@ -1807,7 +1807,7 @@ def _get_dtype(arr_or_dtype):
         return arr_or_dtype
     elif isinstance(arr_or_dtype, type):
         return np.dtype(arr_or_dtype)
-    elif isinstance(arr_or_dtype, CategoricalDtype):
+    elif isinstance(arr_or_dtype, ExtensionDtype):
         return arr_or_dtype
     elif isinstance(arr_or_dtype, DatetimeTZDtype):
         return arr_or_dtype

pandas/core/internals.py

@@ -5541,8 +5541,14 @@ def concatenate_join_units(join_units, concat_axis, copy):
     if len(to_concat) == 1:
         # Only one block, nothing to concatenate.
         concat_values = to_concat[0]
-        if copy and concat_values.base is not None:
-            concat_values = concat_values.copy()
+        if copy:
+            if isinstance(concat_values, np.ndarray):
+                # non-reindexed (=not yet copied) arrays are made into a view
+                # in JoinUnit.get_reindexed_values
+                if concat_values.base is not None:
+                    concat_values = concat_values.copy()
+            else:
+                concat_values = concat_values.copy()
     else:
         concat_values = _concat._concat_compat(to_concat, axis=concat_axis)
 
@@ -5823,7 +5829,7 @@ def get_reindexed_values(self, empty_dtype, upcasted_na):
                 # External code requested filling/upcasting, bool values must
                 # be upcasted to object to avoid being upcasted to numeric.
                 values = self.block.astype(np.object_).values
-            elif self.block.is_categorical:
+            elif self.block.is_extension:
                 values = self.block.values
             else:
                 # No dtype upcasting is done here, it will be performed during

pandas/tests/extension/base/reshaping.py

@@ -95,3 +95,24 @@ def test_set_frame_overwrite_object(self, data):
         df = pd.DataFrame({"A": [1] * len(data)}, dtype=object)
         df['A'] = data
         assert df.dtypes['A'] == data.dtype
+
+    def test_merge(self, data, na_value):
+        # GH-20743
+        df1 = pd.DataFrame({'ext': data[:3], 'int1': [1, 2, 3],
+                            'key': [0, 1, 2]})
+        df2 = pd.DataFrame({'int2': [1, 2, 3, 4], 'key': [0, 0, 1, 3]})
+
+        res = pd.merge(df1, df2)
+        exp = pd.DataFrame(
+            {'int1': [1, 1, 2], 'int2': [1, 2, 3], 'key': [0, 0, 1],
+             'ext': data._constructor_from_sequence(
+                 [data[0], data[0], data[1]])})
+        self.assert_frame_equal(res, exp[['ext', 'int1', 'key', 'int2']])
+
+        res = pd.merge(df1, df2, how='outer')
+        exp = pd.DataFrame(
+            {'int1': [1, 1, 2, 3, np.nan], 'int2': [1, 2, 3, np.nan, 4],
+             'key': [0, 0, 1, 2, 3],
+             'ext': data._constructor_from_sequence(
+                 [data[0], data[0], data[1], data[2], na_value])})
+        self.assert_frame_equal(res, exp[['ext', 'int1', 'key', 'int2']])

pandas/tests/extension/category/test_categorical.py

@@ -75,6 +75,10 @@ def test_align(self, data, na_value):
     def test_align_frame(self, data, na_value):
         pass
 
+    @pytest.mark.skip(reason="Unobserved categories preseved in concat.")
+    def test_merge(self, data, na_value):
+        pass
+
 
 class TestGetitem(base.BaseGetitemTests):
     @pytest.mark.skip(reason="Backwards compatibility")

pandas/tests/extension/decimal/test_decimal.py

@@ -72,6 +72,14 @@ def assert_series_equal(self, left, right, *args, **kwargs):
 
     def assert_frame_equal(self, left, right, *args, **kwargs):
         # TODO(EA): select_dtypes
+        tm.assert_index_equal(
+            left.columns, right.columns,
+            exact=kwargs.get('check_column_type', 'equiv'),
+            check_names=kwargs.get('check_names', True),
+            check_exact=kwargs.get('check_exact', False),
+            check_categorical=kwargs.get('check_categorical', True),
+            obj='{obj}.columns'.format(obj=kwargs.get('obj', 'DataFrame')))
+
         decimals = (left.dtypes == 'decimal').index
 
         for col in decimals: