Fix: coerce unknown types to O dtype

xarray/core/dtypes.py

@@ -278,17 +278,11 @@ def should_promote_to_object(
     """
     np_result_types = set()
     for arr_or_dtype in arrays_and_dtypes:
-        try:
-            result_type = array_api_compat.result_type(
-                maybe_promote_to_variable_width(arr_or_dtype), xp=xp
-            )
-            if isinstance(result_type, np.dtype):
-                np_result_types.add(result_type)
-        except TypeError:
-            # passing individual objects to xp.result_type (i.e., what `array_api_compat.result_type` calls) means NEP-18 implementations won't have
-            # a chance to intercept special values (such as NA) that numpy core cannot handle.
-            # Thus they are considered as types that don't need promotion i.e., the `arr_or_dtype` that rose the `TypeError` will not contribute to `np_result_types`.
-            pass
+        result_type = array_api_compat.result_type(
+            maybe_promote_to_variable_width(arr_or_dtype), xp=xp
+        )
+        if isinstance(result_type, np.dtype):
+            np_result_types.add(result_type)
 
     if np_result_types:
         for left, right in PROMOTE_TO_OBJECT:
@@ -326,13 +320,27 @@ def result_type(
     if xp is None:
         xp = get_array_namespace(arrays_and_dtypes)
 
-    if should_promote_to_object(arrays_and_dtypes, xp):
-        return np.dtype(object)
+    try:
+        if should_promote_to_object(arrays_and_dtypes, xp):
+            return np.dtype(object)
+    except TypeError:
+        # Unknown python objects will raise a TypeError in `xp.result_type`;
+        # We pass the decision on its impact on array type to after we've attempted to promote.
+        pass
+
     maybe_promote = functools.partial(
         maybe_promote_to_variable_width,
         # let extension arrays handle their own str/bytes
         should_return_str_or_bytes=any(
             map(utils.is_allowed_extension_array_dtype, arrays_and_dtypes)
         ),
     )
-    return array_api_compat.result_type(*map(maybe_promote, arrays_and_dtypes), xp=xp)
+    try:
+        result = array_api_compat.result_type(
+            *map(maybe_promote, arrays_and_dtypes), xp=xp
+        )
+    except TypeError:
+        # Unknown python objects will raise a TypeError in `xp.result_type`;
+        # We assume the user wants them to be there and therefore promote to object dtype instead of raising.
+        return np.dtype(object)
+    return result

xarray/core/utils.py

@@ -213,7 +213,7 @@ def maybe_coerce_to_str(index, original_coords):
 
     try:
         result_type = dtypes.result_type(*original_coords)
-    except (TypeError, ValueError):
+    except ValueError:
         pass
     else:
         if result_type.kind in "SU":

xarray/tests/test_dtypes.py

@@ -32,6 +32,10 @@ class DummyArrayAPINamespace:
         ([np.dtype("<U2"), str], np.dtype("U")),
         ([np.dtype("S3"), np.bytes_], np.dtype("S")),
         ([np.dtype("S10"), bytes], np.dtype("S")),
+        ([type("Foo", (object,), {"foo": "bar"})()], np.object_),
+        ([np.float32, type("Foo", (object,), {"foo": "bar"})()], np.object_),
+        ([np.str_, type("Foo", (object,), {"foo": "bar"})()], np.object_),
+        ([np.bytes_, type("Foo", (object,), {"foo": "bar"})()], np.object_),
     ],
 )
 def test_result_type(args, expected) -> None:

xarray/tests/test_utils.py

@@ -29,31 +29,45 @@ def new_method():
             old_method()
 
 
-@pytest.mark.parametrize(
-    ["a", "b", "expected"],
-    [
-        [np.array(["a"]), np.array(["b"]), np.array(["a", "b"])],
-        [np.array([1], dtype="int64"), np.array([2], dtype="int64"), pd.Index([1, 2])],
-    ],
-)
-def test_maybe_coerce_to_str(a, b, expected):
-    index = pd.Index(a).append(pd.Index(b))
-
-    actual = utils.maybe_coerce_to_str(index, [a, b])
-
-    assert_array_equal(expected, actual)
-    assert expected.dtype == actual.dtype
-
-
-def test_maybe_coerce_to_str_minimal_str_dtype():
-    a = np.array(["a", "a_long_string"])
-    index = pd.Index(["a"])
-
-    actual = utils.maybe_coerce_to_str(index, [a])
-    expected = np.array("a")
-
-    assert_array_equal(expected, actual)
-    assert expected.dtype == actual.dtype
+class TestMaybeCoerceToStr:
+    @pytest.mark.parametrize(
+        ["a", "b", "expected"],
+        [
+            [np.array(["a"]), np.array(["b"]), np.array(["a", "b"])],
+            [
+                np.array([1], dtype="int64"),
+                np.array([2], dtype="int64"),
+                pd.Index([1, 2]),
+            ],
+        ],
+    )
+    def test_maybe_coerce_to_str(self, a, b, expected):
+        index = pd.Index(a).append(pd.Index(b))
+
+        actual = utils.maybe_coerce_to_str(index, [a, b])
+
+        assert_array_equal(expected, actual)
+        assert expected.dtype == actual.dtype
+
+    def test_maybe_coerce_to_str_minimal_str_dtype(self):
+        a = np.array(["a", "a_long_string"])
+        index = pd.Index(["a"])
+
+        actual = utils.maybe_coerce_to_str(index, [a])
+        expected = np.array("a")
+
+        assert_array_equal(expected, actual)
+        assert expected.dtype == actual.dtype
+
+    def test_maybe_coerce_to_str_python_obj_dtype(self):
+        """No change to dtype if the array contains a custom python object."""
+        a = np.array([type("Foo", (object,), {"foo": "bar"}), "a_long_string"])
+        index = pd.Index(["a"], dtype=object)
+
+        actual = utils.maybe_coerce_to_str(index, [a])
+
+        assert_array_equal(index, actual)
+        assert index.dtype == actual.dtype
 
 
 class TestArrayEquiv: