Match pandas semantics for null reduction results and mode ordering

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

@@ -1342,7 +1342,7 @@ def all(
         result = self._reduce(
             "all", skipna=True, min_count=min_count, **kwargs
         )
-        if np.isnan(result):
+        if result is pd.NA or np.isnan(result):
             # Empty after dropping NaN/nulls - return np.bool_
             result = np.bool_(True)
 
@@ -1365,20 +1365,40 @@ def any(
             )
         if self.size == 0:
             return False
-        if not skipna and (self.has_nulls() or self.nan_count > 0):
+        is_masked_dtype = is_pandas_nullable_extension_dtype(self.dtype)
+        if not skipna and (
+            self.nan_count > 0 or (not is_masked_dtype and self.has_nulls())
+        ):
+            # NaN values (and the NaN null sentinel of numpy dtypes) are
+            # truthy. For pandas nullable extension dtypes <NA> is not
+            # truthy; Kleene logic below decides between True and <NA>.
             return True
-        elif skipna and self.null_count == self.size:
+        if self.null_count == self.size:
+            if not skipna:
+                # All-null nullable column with skipna=False: Kleene
+                # any([NA, ...]) with no True values is <NA>.
+                return _get_nan_for_dtype(self.dtype)
             return False
 
         # For any(), we want NaN values to be treated as truthy.
         # Call _reduce() with skipna=True to get the boolean result.
         result = self._reduce(
             "any", skipna=True, min_count=min_count, **kwargs
         )
-        if np.isnan(result):
+        if result is pd.NA or np.isnan(result):
             # Empty after dropping NaN/nulls
             # If skipna=False, NaN values should be treated as truthy
             result = np.bool_(not skipna)
+
+        # For pandas nullable extension dtypes with skipna=False, a False
+        # result in the presence of nulls is <NA> under Kleene logic.
+        if (
+            not result
+            and not skipna
+            and self.null_count > 0
+            and is_masked_dtype
+        ):
+            return _get_nan_for_dtype(self.dtype)
         return result
 
     def dropna(self) -> Self:
@@ -2960,6 +2980,11 @@ def _reduce(
                 return col_dtype.type(0)
             if op == "product":
                 return col_dtype.type(1)
+            if is_pandas_nullable_extension_dtype(self.dtype):
+                # pandas returns <NA> for empty/all-null reductions of
+                # nullable dtypes even when the reduction result dtype is
+                # a plain numpy dtype (e.g. Int64.mean() -> float64).
+                return _get_nan_for_dtype(self.dtype)
             return _get_nan_for_dtype(col_dtype)
 
         # Perform the actual reduction

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

@@ -231,12 +231,13 @@ def all(
         self, skipna: bool = True, min_count: int = 0, **kwargs: Any
     ) -> ScalarLike:
         """Check if all string values are truthy (non-empty)."""
-        if skipna and self.null_count == self.size:
-            return True
-        elif not skipna and self.has_nulls():
-            # pandas 3 treats the NaN null sentinel as truthy, matching
-            # numpy semantics, so all(skipna=False) returns True when all
-            # values are null.
+        if self.null_count == self.size:
+            # With skipna=True nulls are dropped, so all-null is vacuously
+            # True. pandas 3 treats the NaN null sentinel as truthy,
+            # matching numpy semantics, so all(skipna=False) is True too.
+            # A partially-null column must NOT short-circuit here: the
+            # result depends on the truthiness of the non-null strings
+            # (e.g. all([NaN, ""], skipna=False) is False).
             return True
         raise NotImplementedError("`all` not implemented for `StringColumn`")
 

python/cudf/cudf/core/series.py

@@ -2857,8 +2857,22 @@ def mode(self, dropna=True):
         if len(val_counts) > 0:
             val_counts = val_counts[val_counts == val_counts.iloc[0]]
 
+        # pandas sorts mode results on the underlying representation:
+        # NaT (INT64_MIN as i8) and the categorical null code (-1) sort
+        # before valid values, while float NaN and the <NA> of
+        # nullable/arrow dtypes (including arrow timestamps/durations)
+        # sort last.
+        na_position = (
+            "first"
+            if (
+                self.dtype.kind in "mM"
+                and isinstance(self.dtype, (np.dtype, pd.DatetimeTZDtype))
+            )
+            or isinstance(self.dtype, cudf.CategoricalDtype)
+            else "last"
+        )
         return Series._from_column(
-            val_counts.index.sort_values()._column,
+            val_counts.index.sort_values(na_position=na_position)._column,
             name=self.name,
             attrs=self.attrs,
         )