BUG: preserve object dtype for Index set ops

doc/source/whatsnew/v1.1.0.rst

@@ -208,7 +208,7 @@ Other
 ^^^^^
 - Appending a dictionary to a :class:`DataFrame` without passing ``ignore_index=True`` will raise ``TypeError: Can only append a dict if ignore_index=True``
   instead of ``TypeError: Can only append a Series if ignore_index=True or if the Series has a name`` (:issue:`30871`)
--
+- Set operations on an object-dtype :class:`Index` now always return object-dtype results (:issue:`31401`)
 
 .. ---------------------------------------------------------------------------
 

pandas/core/indexes/base.py

@@ -524,6 +524,7 @@ def _shallow_copy(self, values=None, **kwargs):
             values = self.values
 
         attributes = self._get_attributes_dict()
+
         attributes.update(kwargs)
 
         return self._simple_new(values, **attributes)
@@ -2592,6 +2593,7 @@ def _union(self, other, sort):
                 # worth making this faster? a very unusual case
                 value_set = set(lvals)
                 result.extend([x for x in rvals if x not in value_set])
+                result = Index(result)._values  # do type inference here
         else:
             # find indexes of things in "other" that are not in "self"
             if self.is_unique:
@@ -2621,7 +2623,8 @@ def _union(self, other, sort):
         return self._wrap_setop_result(other, result)
 
     def _wrap_setop_result(self, other, result):
-        return self._constructor(result, name=get_op_result_name(self, other))
+        name = get_op_result_name(self, other)
+        return self._shallow_copy(result, name=name)
 
     _index_shared_docs[
         "intersection"
@@ -2682,9 +2685,10 @@ def intersection(self, other, sort=False):
         if self.is_monotonic and other.is_monotonic:
             try:
                 result = self._inner_indexer(lvals, rvals)[0]
-                return self._wrap_setop_result(other, result)
             except TypeError:
                 pass
+            else:
+                return self._wrap_setop_result(other, result)
 
         try:
             indexer = Index(rvals).get_indexer(lvals)

pandas/core/indexes/category.py

@@ -29,7 +29,6 @@
 from pandas.core.indexes.base import Index, _index_shared_docs, maybe_extract_name
 from pandas.core.indexes.extension import ExtensionIndex, inherit_names
 import pandas.core.missing as missing
-from pandas.core.ops import get_op_result_name
 
 _index_doc_kwargs = dict(ibase._index_doc_kwargs)
 _index_doc_kwargs.update(dict(target_klass="CategoricalIndex"))
@@ -385,12 +384,6 @@ def _has_complex_internals(self):
         # used to avoid libreduction code paths, which raise or require conversion
         return True
 
-    def _wrap_setop_result(self, other, result):
-        name = get_op_result_name(self, other)
-        # We use _shallow_copy rather than the Index implementation
-        #  (which uses _constructor) in order to preserve dtype.
-        return self._shallow_copy(result, name=name)
-
     @Appender(_index_shared_docs["contains"] % _index_doc_kwargs)
     def __contains__(self, key: Any) -> bool:
         # if key is a NaN, check if any NaN is in self.

pandas/core/indexes/datetimelike.py

@@ -791,11 +791,10 @@ def _union(self, other, sort):
         if this._can_fast_union(other):
             return this._fast_union(other, sort=sort)
         else:
-            result = Index._union(this, other, sort=sort)
-            if isinstance(result, type(self)):
-                assert result._data.dtype == this.dtype
-                if result.freq is None:
-                    result._set_freq("infer")
+            i8self = Int64Index._simple_new(self.asi8, name=self.name)
+            i8other = Int64Index._simple_new(other.asi8, name=other.name)
+            i8result = i8self._union(i8other, sort=sort)
+            result = type(self)(i8result, dtype=self.dtype, freq="infer")
             return result
 
     # --------------------------------------------------------------------

pandas/core/indexes/datetimes.py

@@ -33,7 +33,6 @@
     DatetimeTimedeltaMixin,
 )
 from pandas.core.indexes.extension import inherit_names
-from pandas.core.ops import get_op_result_name
 import pandas.core.tools.datetimes as tools
 
 from pandas.tseries.frequencies import Resolution, to_offset
@@ -367,18 +366,9 @@ def union_many(self, others):
             if this._can_fast_union(other):
                 this = this._fast_union(other)
             else:
-                dtype = this.dtype
                 this = Index.union(this, other)
-                if isinstance(this, DatetimeIndex):
-                    # TODO: we shouldn't be setting attributes like this;
-                    #  in all the tests this equality already holds
-                    this._data._dtype = dtype
         return this
 
-    def _wrap_setop_result(self, other, result):
-        name = get_op_result_name(self, other)
-        return self._shallow_copy(result, name=name, freq=None)
-
     # --------------------------------------------------------------------
 
     def _get_time_micros(self):

pandas/tests/indexes/test_base.py

@@ -1047,6 +1047,26 @@ def test_setops_disallow_true(self, method):
         with pytest.raises(ValueError, match="The 'sort' keyword only takes"):
             getattr(idx1, method)(idx2, sort=True)
 
+    def test_setops_preserve_object_dtype(self):
+        idx = pd.Index([1, 2, 3], dtype=object)
+        result = idx.intersection(idx[1:])
+        expected = idx[1:]
+        tm.assert_index_equal(result, expected)
+
+        # if other is not monotonic increasing, intersection goes through
+        #  a different route
+        result = idx.intersection(idx[1:][::-1])
+        tm.assert_index_equal(result, expected)
+
+        result = idx._union(idx[1:], sort=None)
+        expected = idx
+        tm.assert_index_equal(result, expected)
+
+        # if other is not monotonic increasing, _union goes through
+        #  a different route
+        result = idx._union(idx[1:][::-1], sort=None)
+        tm.assert_index_equal(result, expected)
+
     def test_map_identity_mapping(self, indices):
         # GH 12766
         tm.assert_index_equal(indices, indices.map(lambda x: x))