Always return ndarray

pandas/core/arrays/base.py

@@ -739,14 +739,22 @@ def _create_method(cls, op, coerce_to_dtype=True):
         ----------
         op : function
             An operator that takes arguments op(a, b)
-        coerce_to_dtype :  bool
+        coerce_to_dtype :  bool, default True
             boolean indicating whether to attempt to convert
-            the result to the underlying ExtensionArray dtype
-            (default True)
+            the result to the underlying ExtensionArray dtype.
+            If it's not possible to create a new ExtensionArray with the
+            values, an ndarray is returned instead.
 
         Returns
         -------
-        A method that can be bound to a method of a class
+        Callable[[Any, Any], Union[ndarray, ExtensionArray]]
+            A method that can be bound to a class. When used, the method
+            receives the two arguments, one of which is the instance of
+            this class, and should return an ExtensionArray or an ndarray.
+
+            Returning an ndarray may be necessary when the result of the
+            `op` cannot be stored in the ExtensionArray. The dtype of the
+            ndarray uses NumPy's normal inference rules.
 
         Example
         -------
@@ -757,7 +765,6 @@ def _create_method(cls, op, coerce_to_dtype=True):
         in the class definition of MyExtensionArray to create the operator
         for addition, that will be based on the operator implementation
         of the underlying elements of the ExtensionArray
-
         """
 
         def _binop(self, other):
@@ -778,7 +785,12 @@ def convert_values(param):
                 try:
                     res = self._from_sequence(res)
                 except Exception:
-                    res = np.asarray(res, dtype=object)
+                    # https://github.com/pandas-dev/pandas/issues/22850
+                    # We catch all regular exceptions here, and fall back
+                    # to an ndarray.
+                    res = np.asarray(res)
+            else:
+                res = np.asarray(res)
 
             return res
 

pandas/core/series.py

@@ -2323,12 +2323,14 @@ def combine(self, other, func, fill_value=None):
             pass
         elif is_extension_array_dtype(self.values):
             # The function can return something of any type, so check
-            # if the type is compatible with the calling EA
-            # ExtensionArray._from_sequence can raise anything, so we
-            # have to catch everything.
+            # if the type is compatible with the calling EA.
             try:
                 new_values = self._values._from_sequence(new_values)
             except Exception:
+                # https://github.com/pandas-dev/pandas/issues/22850
+                # pandas has no control over what 3rd-party ExtensionArrays
+                # do in _values_from_sequence. We still want ops to work
+                # though, so we catch any regular Exception.
                 pass
 
         return self._constructor(new_values, index=new_index, name=new_name)

pandas/tests/extension/decimal/test_decimal.py

@@ -279,6 +279,15 @@ def _from_sequence(cls, scalars, dtype=None, copy=False):
         raise KeyError("For the test")
 
 
+class DecimalArrayWithoutCoercion(DecimalArrayWithoutFromSequence):
+    @classmethod
+    def _create_arithmetic_method(cls, op):
+        return cls._create_method(op, coerce_to_dtype=False)
+
+
+DecimalArrayWithoutCoercion._add_arithmetic_ops()
+
+
 def test_combine_from_sequence_raises():
     # https://github.com/pandas-dev/pandas/issues/22850
     ser = pd.Series(DecimalArrayWithoutFromSequence([
@@ -293,8 +302,12 @@ def test_combine_from_sequence_raises():
     tm.assert_series_equal(result, expected)
 
 
-def test_scalar_ops_from_sequence_raises():
-    arr = DecimalArrayWithoutFromSequence([
+@pytest.mark.parametrize("class_", [DecimalArrayWithoutFromSequence,
+                                    DecimalArrayWithoutCoercion])
+def test_scalar_ops_from_sequence_raises(class_):
+    # op(EA, EA) should return an EA, or an ndarray if it's not possible
+    # to return an EA with the return values.
+    arr = class_([
         decimal.Decimal("1.0"),
         decimal.Decimal("2.0")
     ])