ENH: Add future.python_scalars

.github/workflows/unit-tests.yml

@@ -31,6 +31,7 @@ jobs:
         # Prevent the include jobs from overriding other jobs
         pattern: [""]
         pandas_future_infer_string: ["1"]
+        pandas_future_python_scalars: ["0"]
         include:
           - name: "Downstream Compat"
             env_file: actions-313-downstream_compat.yaml
@@ -75,6 +76,10 @@ jobs:
             env_file: actions-313.yaml
             pandas_future_infer_string: "0"
             platform: ubuntu-24.04
+          - name: "PANDAS_FUTURE_PYTHON_SCALARS=1"
+            env_file: actions-313.yaml
+            pandas_future_python_scalars: "1"
+            platform: ubuntu-24.04
           - name: "Numpy Dev"
             env_file: actions-313-numpydev.yaml
             pattern: "not slow and not network and not single_cpu"
@@ -92,6 +97,7 @@ jobs:
       LC_ALL: ${{ matrix.lc_all || '' }}
       PANDAS_CI: '1'
       PANDAS_FUTURE_INFER_STRING: ${{ matrix.pandas_future_infer_string || '1' }}
+      PANDAS_FUTURE_PYTHON_SCALARS: ${{ matrix.pandas_future_python_scalars || '0' }}
       TEST_ARGS: ${{ matrix.test_args || '' }}
       PYTEST_WORKERS: 'auto'
       PYTEST_TARGET: ${{ matrix.pytest_target || 'pandas' }}

pandas/_config/__init__.py

@@ -35,6 +35,11 @@ def using_string_dtype() -> bool:
     return _mode_options["infer_string"]
 
 
+def using_python_scalars() -> bool:
+    _mode_options = _global_config["future"]
+    return _mode_options["python_scalars"]
+
+
 def is_nan_na() -> bool:
     _mode_options = _global_config["mode"]
     return _mode_options["nan_is_na"]

pandas/conftest.py

@@ -2097,6 +2097,11 @@ def using_infer_string() -> bool:
     return pd.options.future.infer_string is True
 
 
+@pytest.fixture
+def using_python_scalars() -> bool:
+    return pd.options.future.python_scalars is True
+
+
 _warsaws: list[Any] = ["Europe/Warsaw", "dateutil/Europe/Warsaw"]
 if pytz is not None:
     _warsaws.append(pytz.timezone("Europe/Warsaw"))

pandas/core/arraylike.py

@@ -15,6 +15,7 @@
 from pandas._libs import lib
 from pandas._libs.ops_dispatch import maybe_dispatch_ufunc_to_dunder_op
 
+from pandas.core.dtypes.cast import maybe_unbox_numpy_scalar
 from pandas.core.dtypes.generic import ABCNDFrame
 
 from pandas.core import roperator
@@ -529,4 +530,6 @@ def dispatch_reduction_ufunc(self, ufunc: np.ufunc, method: str, *inputs, **kwar
 
     # By default, numpy's reductions do not skip NaNs, so we have to
     #  pass skipna=False
-    return getattr(self, method_name)(skipna=False, **kwargs)
+    result = getattr(self, method_name)(skipna=False, **kwargs)
+    result = maybe_unbox_numpy_scalar(result)
+    return result

pandas/core/arrays/masked.py

@@ -12,7 +12,10 @@
 
 import numpy as np
 
-from pandas._config import is_nan_na
+from pandas._config import (
+    is_nan_na,
+    using_python_scalars,
+)
 
 from pandas._libs import (
     algos as libalgos,
@@ -27,7 +30,10 @@
 from pandas.errors import AbstractMethodError
 
 from pandas.core.dtypes.base import ExtensionDtype
-from pandas.core.dtypes.cast import maybe_downcast_to_dtype
+from pandas.core.dtypes.cast import (
+    maybe_downcast_to_dtype,
+    maybe_unbox_numpy_scalar,
+)
 from pandas.core.dtypes.common import (
     is_bool,
     is_integer_dtype,
@@ -1518,7 +1524,10 @@ def _reduce(
             if isna(result):
                 return self._wrap_na_result(name=name, axis=0, mask_size=(1,))
             else:
-                result = result.reshape(1)
+                if using_python_scalars():
+                    result = np.array([result])
+                else:
+                    result = result.reshape(1)
                 mask = np.zeros(1, dtype=bool)
                 return self._maybe_mask_result(result, mask)
 
@@ -1742,7 +1751,7 @@ def any(
 
         values = self._data.copy()
         np.putmask(values, self._mask, self.dtype._falsey_value)
-        result = values.any()
+        result = maybe_unbox_numpy_scalar(values.any())
         if skipna:
             return result
         else:
@@ -1828,7 +1837,7 @@ def all(
 
         values = self._data.copy()
         np.putmask(values, self._mask, self.dtype._truthy_value)
-        result = values.all(axis=axis)
+        result = maybe_unbox_numpy_scalar(values.all(axis=axis))
 
         if skipna:
             return result  # type: ignore[return-value]

pandas/core/config_init.py

@@ -900,5 +900,13 @@ def register_converter_cb(key: str) -> None:
         validator=is_one_of_factory([True, False]),
     )
 
+    cf.register_option(
+        "python_scalars",
+        False if os.environ.get("PANDAS_FUTURE_PYTHON_SCALARS", "0") == "0" else True,
+        "Whether to return Python scalars instead of NumPy or PyArrow scalars. "
+        "Currently experimental, setting to True is not recommended for end users.",
+        validator=is_one_of_factory([True, False]),
+    )
+
 # GH#59502
 cf.deprecate_option("future.no_silent_downcasting", Pandas4Warning)

pandas/core/dtypes/cast.py

@@ -20,6 +20,7 @@
 
 from pandas._config import (
     is_nan_na,
+    using_python_scalars,
     using_string_dtype,
 )
 
@@ -1434,6 +1435,16 @@ def construct_1d_arraylike_from_scalar(
     return subarr
 
 
+def maybe_unbox_numpy_scalar(value):
+    result = value
+    if using_python_scalars() and isinstance(value, np.generic):
+        if isinstance(result, np.longdouble):
+            result = float(result)
+        else:
+            result = value.item()
+    return result
+
+
 def _maybe_box_and_unbox_datetimelike(value: Scalar, dtype: DtypeObj):
     # Caller is responsible for checking dtype.kind in "mM"
 

pandas/core/indexes/base.py

@@ -92,6 +92,7 @@
     common_dtype_categorical_compat,
     find_result_type,
     infer_dtype_from,
+    maybe_unbox_numpy_scalar,
     np_can_hold_element,
 )
 from pandas.core.dtypes.common import (
@@ -7532,7 +7533,7 @@ def min(self, axis: AxisInt | None = None, skipna: bool = True, *args, **kwargs)
             # quick check
             first = self[0]
             if not isna(first):
-                return first
+                return maybe_unbox_numpy_scalar(first)
 
         if not self._is_multi and self.hasnans:
             # Take advantage of cache
@@ -7543,7 +7544,7 @@ def min(self, axis: AxisInt | None = None, skipna: bool = True, *args, **kwargs)
         if not self._is_multi and not isinstance(self._values, np.ndarray):
             return self._values._reduce(name="min", skipna=skipna)
 
-        return nanops.nanmin(self._values, skipna=skipna)
+        return maybe_unbox_numpy_scalar(nanops.nanmin(self._values, skipna=skipna))
 
     def max(self, axis: AxisInt | None = None, skipna: bool = True, *args, **kwargs):
         """
@@ -7596,18 +7597,18 @@ def max(self, axis: AxisInt | None = None, skipna: bool = True, *args, **kwargs)
             # quick check
             last = self[-1]
             if not isna(last):
-                return last
+                return maybe_unbox_numpy_scalar(last)
 
         if not self._is_multi and self.hasnans:
             # Take advantage of cache
             mask = self._isnan
             if not skipna or mask.all():
-                return self._na_value
+                return maybe_unbox_numpy_scalar(self._na_value)
 
         if not self._is_multi and not isinstance(self._values, np.ndarray):
             return self._values._reduce(name="max", skipna=skipna)
 
-        return nanops.nanmax(self._values, skipna=skipna)
+        return maybe_unbox_numpy_scalar(nanops.nanmax(self._values, skipna=skipna))
 
     # --------------------------------------------------------------------
 

pandas/core/interchange/column.py

@@ -7,6 +7,8 @@
 
 import numpy as np
 
+from pandas._config import using_python_scalars
+
 from pandas._libs.lib import infer_dtype
 from pandas._libs.tslibs import iNaT
 from pandas.errors import NoBufferPresent
@@ -232,7 +234,10 @@ def null_count(self) -> int:
         """
         Number of null elements. Should always be known.
         """
-        return self._col.isna().sum().item()
+        result = self._col.isna().sum()
+        if not using_python_scalars():
+            result = result.item()
+        return result
 
     @property
     def metadata(self) -> dict[str, pd.Index]:

pandas/core/series.py

@@ -73,6 +73,7 @@
     find_common_type,
     infer_dtype_from,
     maybe_box_native,
+    maybe_unbox_numpy_scalar,
 )
 from pandas.core.dtypes.common import (
     is_dict_like,
@@ -2014,7 +2015,7 @@ def count(self) -> int:
         >>> s.count()
         2
         """
-        return notna(self._values).sum().astype("int64")
+        return maybe_unbox_numpy_scalar(notna(self._values).sum().astype("int64"))
 
     def mode(self, dropna: bool = True) -> Series:
         """
@@ -7356,7 +7357,7 @@ def _reduce(
 
         if isinstance(delegate, ExtensionArray):
             # dispatch to ExtensionArray interface
-            return delegate._reduce(name, skipna=skipna, **kwds)
+            result = delegate._reduce(name, skipna=skipna, **kwds)
 
         else:
             # dispatch to numpy arrays
@@ -7370,7 +7371,10 @@ def _reduce(
                     f"Series.{name} does not allow {kwd_name}={numeric_only} "
                     "with non-numeric dtypes."
                 )
-            return op(delegate, skipna=skipna, **kwds)
+            result = op(delegate, skipna=skipna, **kwds)
+
+        result = maybe_unbox_numpy_scalar(result)
+        return result
 
     @Appender(make_doc("any", ndim=1))
     # error: Signature of "any" incompatible with supertype "NDFrame"

pandas/tests/arrays/boolean/test_reduction.py

@@ -25,12 +25,15 @@ def data():
         ([False, False], False, False, False, False),
     ],
 )
-def test_any_all(values, exp_any, exp_all, exp_any_noskip, exp_all_noskip):
+def test_any_all(
+    values, exp_any, exp_all, exp_any_noskip, exp_all_noskip, using_python_scalars
+):
     # the methods return numpy scalars
-    exp_any = pd.NA if exp_any is pd.NA else np.bool_(exp_any)
-    exp_all = pd.NA if exp_all is pd.NA else np.bool_(exp_all)
-    exp_any_noskip = pd.NA if exp_any_noskip is pd.NA else np.bool_(exp_any_noskip)
-    exp_all_noskip = pd.NA if exp_all_noskip is pd.NA else np.bool_(exp_all_noskip)
+    if not using_python_scalars:
+        exp_any = pd.NA if exp_any is pd.NA else np.bool_(exp_any)
+        exp_all = pd.NA if exp_all is pd.NA else np.bool_(exp_all)
+        exp_any_noskip = pd.NA if exp_any_noskip is pd.NA else np.bool_(exp_any_noskip)
+        exp_all_noskip = pd.NA if exp_all_noskip is pd.NA else np.bool_(exp_all_noskip)
 
     for con in [pd.array, pd.Series]:
         a = con(values, dtype="boolean")
@@ -39,23 +42,30 @@ def test_any_all(values, exp_any, exp_all, exp_any_noskip, exp_all_noskip):
         assert a.any(skipna=False) is exp_any_noskip
         assert a.all(skipna=False) is exp_all_noskip
 
-        assert np.any(a.any()) is exp_any
-        assert np.all(a.all()) is exp_all
 
-
-def test_reductions_return_types(dropna, data, all_numeric_reductions):
+def test_reductions_return_types(
+    dropna, data, all_numeric_reductions, using_python_scalars
+):
     op = all_numeric_reductions
     s = pd.Series(data)
     if dropna:
         s = s.dropna()
 
-    if op in ("sum", "prod"):
-        assert isinstance(getattr(s, op)(), np.int_)
-    elif op == "count":
-        # Oddly on the 32 bit build (but not Windows), this is intc (!= intp)
-        assert isinstance(getattr(s, op)(), np.integer)
-    elif op in ("min", "max"):
-        assert isinstance(getattr(s, op)(), np.bool_)
+    if using_python_scalars:
+        expected = {
+            "sum": int,
+            "prod": int,
+            "count": int,
+            "min": bool,
+            "max": bool,
+        }.get(op, float)
     else:
-        # "mean", "std", "var", "median", "kurt", "skew"
-        assert isinstance(getattr(s, op)(), np.float64)
+        expected = {
+            "sum": np.int_,
+            "prod": np.int_,
+            "count": np.integer,
+            "min": np.bool_,
+            "max": np.bool_,
+        }.get(op, np.float64)
+    result = getattr(s, op)()
+    assert isinstance(result, expected), f"{type(result)} vs {expected}"

pandas/tests/arrays/floating/test_function.py

@@ -171,7 +171,7 @@ def test_floating_array_numpy_sum(values, expected):
 
 
 @pytest.mark.parametrize("op", ["sum", "min", "max", "prod"])
-def test_preserve_dtypes(op):
+def test_preserve_dtypes(op, using_python_scalars):
     df = pd.DataFrame(
         {
             "A": ["a", "b", "b"],
@@ -182,7 +182,10 @@ def test_preserve_dtypes(op):
 
     # op
     result = getattr(df.C, op)()
-    assert isinstance(result, np.float64)
+    if using_python_scalars:
+        assert isinstance(result, float)
+    else:
+        assert isinstance(result, np.float64)
 
     # groupby
     result = getattr(df.groupby("A"), op)()

pandas/tests/arrays/integer/test_dtypes.py

@@ -22,7 +22,7 @@ def test_dtypes(dtype):
 
 
 @pytest.mark.parametrize("op", ["sum", "min", "max", "prod"])
-def test_preserve_dtypes(op):
+def test_preserve_dtypes(op, using_python_scalars):
     # for ops that enable (mean would actually work here
     # but generally it is a float return value)
     df = pd.DataFrame(
@@ -35,7 +35,7 @@ def test_preserve_dtypes(op):
 
     # op
     result = getattr(df.C, op)()
-    if op in {"sum", "prod", "min", "max"}:
+    if op in {"sum", "prod", "min", "max"} and not using_python_scalars:
         assert isinstance(result, np.int64)
     else:
         assert isinstance(result, int)