Backport PR pandas-dev#53232 on branch 2.0.x (BUG: sort_values raising for dictionary arrow dtype) (pandas-dev#53241)

BUG: sort_values raising for dictionary arrow dtype (pandas-dev#53232)

(cherry picked from commit e1b657a)

Author: phofl

doc/source/whatsnew/v2.0.2.rst

@@ -30,6 +30,7 @@ Bug fixes
 - Bug in :func:`api.interchange.from_dataframe` was unnecessarily raising on bitmasks (:issue:`49888`)
 - Bug in :func:`merge` when merging on datetime columns on different resolutions (:issue:`53200`)
 - Bug in :meth:`DataFrame.convert_dtypes` ignores ``convert_*`` keywords when set to False ``dtype_backend="pyarrow"`` (:issue:`52872`)
+- Bug in :meth:`DataFrame.sort_values` raising for PyArrow ``dictionary`` dtype (:issue:`53232`)
 - Bug in :meth:`Series.describe` treating pyarrow-backed timestamps and timedeltas as categorical data (:issue:`53001`)
 - Bug in :meth:`Series.rename` not making a lazy copy when Copy-on-Write is enabled when a scalar is passed to it (:issue:`52450`)
 - Bug in :meth:`pd.array` raising for ``NumPy`` array and ``pa.large_string`` or ``pa.large_binary`` (:issue:`52590`)

pandas/core/arrays/arrow/array.py

@@ -269,7 +269,10 @@ def _from_sequence(cls, scalars, *, dtype: Dtype | None = None, copy: bool = Fal
                 # GH50430: let pyarrow infer type, then cast
                 scalars = pa.array(scalars, from_pandas=True)
         if pa_dtype:
-            scalars = scalars.cast(pa_dtype)
+            if pa.types.is_dictionary(pa_dtype):
+                scalars = scalars.dictionary_encode()
+            else:
+                scalars = scalars.cast(pa_dtype)
         arr = cls(scalars)
         if pa.types.is_duration(scalars.type) and scalars.null_count > 0:
             # GH52843: upstream bug for duration types when originally
@@ -868,7 +871,10 @@ def factorize(
         else:
             data = self._data
 
-        encoded = data.dictionary_encode(null_encoding=null_encoding)
+        if pa.types.is_dictionary(data.type):
+            encoded = data
+        else:
+            encoded = data.dictionary_encode(null_encoding=null_encoding)
         if encoded.length() == 0:
             indices = np.array([], dtype=np.intp)
             uniques = type(self)(pa.chunked_array([], type=encoded.type.value_type))

pandas/tests/extension/test_arrow.py

@@ -1831,6 +1831,20 @@ def test_searchsorted_with_na_raises(data_for_sorting, as_series):
         arr.searchsorted(b)
 
 
+def test_sort_values_dictionary():
+    df = pd.DataFrame(
+        {
+            "a": pd.Series(
+                ["x", "y"], dtype=ArrowDtype(pa.dictionary(pa.int32(), pa.string()))
+            ),
+            "b": [1, 2],
+        },
+    )
+    expected = df.copy()
+    result = df.sort_values(by=["a", "b"])
+    tm.assert_frame_equal(result, expected)
+
+
 @pytest.mark.parametrize("pat", ["abc", "a[a-z]{2}"])
 def test_str_count(pat):
     ser = pd.Series(["abc", None], dtype=ArrowDtype(pa.string()))