Fix type casting in Series.__setitem__

python/cudf/cudf/core/column/column.py

@@ -854,6 +854,8 @@ def can_cast_safely(self, to_dtype: Dtype) -> bool:
         raise NotImplementedError()
 
     def astype(self, dtype: Dtype, **kwargs) -> ColumnBase:
+        if dtype == self.dtype:
+            return self
         if is_categorical_dtype(dtype):
             return self.as_categorical_column(dtype, **kwargs)
 

python/cudf/cudf/core/column/lists.py

@@ -742,6 +742,8 @@ def astype(self, dtype):
         >>> s2.dtype
         ListDtype(float64)
         """
+        if self.dtype == dtype:
+            return self
         return self._return_or_inplace(
             self._column._transform_leaves(
                 lambda col, dtype: col.astype(dtype), dtype

python/cudf/cudf/core/scalar.py

@@ -392,4 +392,6 @@ def _dispatch_scalar_unaop(self, op):
         return getattr(self.value, op)()
 
     def astype(self, dtype):
+        if self.dtype == dtype:
+            return self
         return Scalar(self.value, dtype)

python/cudf/cudf/core/series.py

@@ -216,17 +216,25 @@ def __setitem__(self, key, value):
         if (
             not isinstance(
                 self._frame._column.dtype,
-                (cudf.core.dtypes.DecimalDtype, cudf.CategoricalDtype),
+                (
+                    # Casting only for string and non-decimal numeric
+                    # columns
+                    cudf.core.dtypes.DecimalDtype,
+                    cudf.CategoricalDtype,
+                    cudf.ListDtype,
+                    cudf.StructDtype,
+                ),
             )
             and hasattr(value, "dtype")
             and _is_non_decimal_numeric_dtype(value.dtype)
         ):
             # normalize types if necessary:
-            if not is_integer(key):
-                to_dtype = np.result_type(
-                    value.dtype, self._frame._column.dtype
-                )
-                value = value.astype(to_dtype)
+            # In contrast to Column.__setitem__ (which downcasts the value to
+            # the dtype of the column) here we upcast the series to the
+            # larger data type mimicking pandas
+            to_dtype = np.result_type(value.dtype, self._frame._column.dtype)
+            value = value.astype(to_dtype)
+            if to_dtype != self._frame._column.dtype:
                 self._frame._column._mimic_inplace(
                     self._frame._column.astype(to_dtype), inplace=True
                 )

python/cudf/cudf/tests/test_setitem.py

@@ -297,3 +297,45 @@ def test_series_slice_setitem_struct():
     actual[0:3] = cudf.Scalar({"a": {"b": 5050}, "b": 101})
 
     assert_eq(actual, expected)
+
+
+@pytest.mark.parametrize("dtype", [np.int32, np.int64, np.float32, np.float64])
+@pytest.mark.parametrize("indices", [0, [1, 2]])
+def test_series_setitem_upcasting(dtype, indices):
+    sr = pd.Series([0, 0, 0], dtype=dtype)
+    cr = cudf.from_pandas(sr)
+    assert_eq(sr, cr)
+    new_value = np.float64(10.5)
+    col_ref = cr._column
+    sr[indices] = new_value
+    cr[indices] = new_value
+    if PANDAS_GE_150:
+        assert_eq(sr, cr)
+    else:
+        # pandas bug, incorrectly fails to upcasting from float32 to float64
+        assert_eq(sr.values, cr.values)
+    if dtype == np.float64:
+        # no-op type cast should not modify backing column
+        assert col_ref == cr._column
+
+
+# TODO: these two tests could perhaps be changed once specifics of
+# pandas compat wrt upcasting are decided on; this is just baking in
+# status-quo.
+def test_series_setitem_upcasting_string_column():
+    sr = pd.Series([0, 0, 0], dtype=str)
+    cr = cudf.from_pandas(sr)
+    new_value = np.float64(10.5)
+    sr[0] = str(new_value)
+    cr[0] = new_value
+    assert_eq(sr, cr)
+
+
+def test_series_setitem_upcasting_string_value():
+    sr = cudf.Series([0, 0, 0], dtype=int)
+    # This is a distinction with pandas, which let's you instead make an
+    # object column with ["10", 0, 0]
+    sr[0] = "10"
+    assert_eq(pd.Series([10, 0, 0], dtype=int), sr)
+    with pytest.raises(ValueError):
+        sr[0] = "non-integer"