Backport PR pandas-dev#54110: CI: Fix pyarrow nightly build

Author: mroeschke

.github/actions/build_pandas/action.yml

@@ -8,6 +8,7 @@ runs:
       run: |
         micromamba info
         micromamba list
+        pip list --pre
       shell: bash -el {0}
 
     - name: Build Pandas

pandas/compat/__init__.py

@@ -31,6 +31,7 @@
     pa_version_under8p0,
     pa_version_under9p0,
     pa_version_under11p0,
+    pa_version_under13p0,
 )
 
 
@@ -161,6 +162,7 @@ def get_lzma_file() -> type[pandas.compat.compressors.LZMAFile]:
     "pa_version_under8p0",
     "pa_version_under9p0",
     "pa_version_under11p0",
+    "pa_version_under13p0",
     "IS64",
     "PY39",
     "PY310",

pandas/compat/pyarrow.py

@@ -7,16 +7,17 @@
 try:
     import pyarrow as pa
 
-    _pa_version = pa.__version__
-    _palv = Version(_pa_version)
+    _palv = Version(Version(pa.__version__).base_version)
     pa_version_under7p0 = _palv < Version("7.0.0")
     pa_version_under8p0 = _palv < Version("8.0.0")
     pa_version_under9p0 = _palv < Version("9.0.0")
     pa_version_under10p0 = _palv < Version("10.0.0")
     pa_version_under11p0 = _palv < Version("11.0.0")
+    pa_version_under13p0 = _palv < Version("13.0.0")
 except ImportError:
     pa_version_under7p0 = True
     pa_version_under8p0 = True
     pa_version_under9p0 = True
     pa_version_under10p0 = True
     pa_version_under11p0 = True
+    pa_version_under13p0 = True

pandas/core/arrays/arrow/array.py

@@ -40,6 +40,7 @@
     pa_version_under8p0,
     pa_version_under9p0,
     pa_version_under11p0,
+    pa_version_under13p0,
 )
 from pandas.util._decorators import doc
 from pandas.util._validators import validate_fillna_kwargs
@@ -1252,6 +1253,8 @@ def _reduce(self, name: str, *, skipna: bool = True, **kwargs):
 
         data_to_reduce = self._data
 
+        cast_kwargs = {} if pa_version_under13p0 else {"safe": False}
+
         if name in ["any", "all"] and (
             pa.types.is_integer(pa_type)
             or pa.types.is_floating(pa_type)
@@ -1327,9 +1330,15 @@ def pyarrow_meth(data, skip_nulls, **kwargs):
         if name in ["min", "max", "sum"] and pa.types.is_duration(pa_type):
             result = result.cast(pa_type)
         if name in ["median", "mean"] and pa.types.is_temporal(pa_type):
+            if not pa_version_under13p0:
+                nbits = pa_type.bit_width
+                if nbits == 32:
+                    result = result.cast(pa.int32(), **cast_kwargs)
+                else:
+                    result = result.cast(pa.int64(), **cast_kwargs)
             result = result.cast(pa_type)
         if name in ["std", "sem"] and pa.types.is_temporal(pa_type):
-            result = result.cast(pa.int64())
+            result = result.cast(pa.int64(), **cast_kwargs)
             if pa.types.is_duration(pa_type):
                 result = result.cast(pa_type)
             elif pa.types.is_time(pa_type):

pandas/tests/io/test_parquet.py

@@ -14,6 +14,7 @@
 from pandas.compat.pyarrow import (
     pa_version_under7p0,
     pa_version_under8p0,
+    pa_version_under13p0,
 )
 import pandas.util._test_decorators as td
 
@@ -998,14 +999,15 @@ def test_read_dtype_backend_pyarrow_config(self, pa, df_full):
 
         pa_table = pyarrow.Table.from_pandas(df)
         expected = pa_table.to_pandas(types_mapper=pd.ArrowDtype)
-        # pyarrow infers datetimes as us instead of ns
-        expected["datetime"] = expected["datetime"].astype("timestamp[us][pyarrow]")
-        expected["datetime_with_nat"] = expected["datetime_with_nat"].astype(
-            "timestamp[us][pyarrow]"
-        )
-        expected["datetime_tz"] = expected["datetime_tz"].astype(
-            pd.ArrowDtype(pyarrow.timestamp(unit="us", tz="Europe/Brussels"))
-        )
+        if pa_version_under13p0:
+            # pyarrow infers datetimes as us instead of ns
+            expected["datetime"] = expected["datetime"].astype("timestamp[us][pyarrow]")
+            expected["datetime_with_nat"] = expected["datetime_with_nat"].astype(
+                "timestamp[us][pyarrow]"
+            )
+            expected["datetime_tz"] = expected["datetime_tz"].astype(
+                pd.ArrowDtype(pyarrow.timestamp(unit="us", tz="Europe/Brussels"))
+            )
 
         check_round_trip(
             df,