CoW: add readonly flag to ExtensionArrays, return read-only EA/ndarray in .array/EA.to_numpy()

jorisvandenbossche · jorisvandenbossche · commit a9df51b017f0 · 2025-07-23T00:46:10.000+02:00
diff --git a/pandas/core/arrays/_mixins.py b/pandas/core/arrays/_mixins.py
@@ -252,6 +252,9 @@ def shift(self, periods: int = 1, fill_value=None) -> Self:
         return self._from_backing_data(new_values)
 
     def __setitem__(self, key, value) -> None:
+        if self._readonly:
+            raise ValueError("Cannot modify readonly array")
+
         key = check_array_indexer(self, key)
         value = self._validate_setitem_value(value)
         self._ndarray[key] = value
diff --git a/pandas/core/arrays/arrow/array.py b/pandas/core/arrays/arrow/array.py
@@ -1967,6 +1967,9 @@ def __setitem__(self, key, value) -> None:
         -------
         None
         """
+        if self._readonly:
+            raise ValueError("Cannot modify readonly array")
+
         # GH50085: unwrap 1D indexers
         if isinstance(key, tuple) and len(key) == 1:
             key = key[0]
diff --git a/pandas/core/arrays/base.py b/pandas/core/arrays/base.py
@@ -40,6 +40,7 @@
     validate_insert_loc,
 )
 
+from pandas.core.dtypes.astype import astype_is_view
 from pandas.core.dtypes.cast import maybe_cast_pointwise_result
 from pandas.core.dtypes.common import (
     is_list_like,
@@ -269,6 +270,8 @@ class ExtensionArray:
     #  strictly less than 2000 to be below Index.__pandas_priority__.
     __pandas_priority__ = 1000
 
+    _readonly = False
+
     # ------------------------------------------------------------------------
     # Constructors
     # ------------------------------------------------------------------------
@@ -482,6 +485,11 @@ def __setitem__(self, key, value) -> None:
         Returns
         -------
         None
+
+        Raises
+        ------
+        ValueError
+            If the array is readonly and modification is attempted.
         """
         # Some notes to the ExtensionArray implementer who may have ended up
         # here. While this method is not required for the interface, if you
@@ -501,8 +509,59 @@ def __setitem__(self, key, value) -> None:
         #   __init__ method coerces that value, then so should __setitem__
         # Note, also, that Series/DataFrame.where internally use __setitem__
         # on a copy of the data.
+        # Check if the array is readonly
+        if self._readonly:
+            raise ValueError("Cannot modify readonly array")
+
         raise NotImplementedError(f"{type(self)} does not implement __setitem__.")
 
+    @property
+    def readonly(self) -> bool:
+        """
+        Whether the array is readonly.
+
+        If True, attempts to modify the array via __setitem__ will raise
+        a ValueError.
+
+        Returns
+        -------
+        bool
+            True if the array is readonly, False otherwise.
+
+        Examples
+        --------
+        >>> arr = pd.array([1, 2, 3])
+        >>> arr.readonly
+        False
+        >>> arr.readonly = True
+        >>> arr[0] = 5
+        Traceback (most recent call last):
+        ...
+        ValueError: Cannot modify readonly ExtensionArray
+        """
+        return getattr(self, "_readonly", False)
+
+    @readonly.setter
+    def readonly(self, value: bool) -> None:
+        """
+        Set the readonly state of the array.
+
+        Parameters
+        ----------
+        value : bool
+            True to make the array readonly, False to make it writable.
+
+        Examples
+        --------
+        >>> arr = pd.array([1, 2, 3])
+        >>> arr.readonly = True
+        >>> arr.readonly
+        True
+        """
+        if not isinstance(value, bool):
+            raise TypeError("readonly must be a boolean")
+        self._readonly = value
+
     def __len__(self) -> int:
         """
         Length of this array
@@ -595,8 +654,14 @@ def to_numpy(
         result = np.asarray(self, dtype=dtype)
         if copy or na_value is not lib.no_default:
             result = result.copy()
+        elif self._readonly and astype_is_view(self.dtype, result.dtype):
+            # If the ExtensionArray is readonly, make the numpy array readonly too
+            result = result.view()
+            result.flags.writeable = False
+
         if na_value is not lib.no_default:
             result[self.isna()] = na_value  # type: ignore[index]
+
         return result
 
     # ------------------------------------------------------------------------
diff --git a/pandas/core/arrays/datetimelike.py b/pandas/core/arrays/datetimelike.py
@@ -368,7 +368,12 @@ def __array__(
 
         if copy is True:
             return np.array(self._ndarray, dtype=dtype)
-        return self._ndarray
+
+        result = self._ndarray
+        if self._readonly:
+            result = result.view()
+            result.flags.writeable = False
+        return result
 
     @overload
     def __getitem__(self, key: ScalarIndexer) -> DTScalarOrNaT: ...
diff --git a/pandas/core/arrays/interval.py b/pandas/core/arrays/interval.py
@@ -729,6 +729,9 @@ def __getitem__(self, key: PositionalIndexer) -> Self | IntervalOrNA:
         return self._simple_new(left, right, dtype=self.dtype)  # type: ignore[arg-type]
 
     def __setitem__(self, key, value) -> None:
+        if self._readonly:
+            raise ValueError("Cannot modify readonly array")
+
         value_left, value_right = self._validate_setitem_value(value)
         key = check_array_indexer(self, key)
 
diff --git a/pandas/core/arrays/masked.py b/pandas/core/arrays/masked.py
@@ -23,6 +23,7 @@
 from pandas.errors import AbstractMethodError
 from pandas.util._decorators import doc
 
+from pandas.core.dtypes.astype import astype_is_view
 from pandas.core.dtypes.base import ExtensionDtype
 from pandas.core.dtypes.common import (
     is_bool,
@@ -290,6 +291,9 @@ def _validate_setitem_value(self, value):
         raise TypeError(f"Invalid value '{value!s}' for dtype '{self.dtype}'")
 
     def __setitem__(self, key, value) -> None:
+        if self._readonly:
+            raise ValueError("Cannot modify readonly array")
+
         key = check_array_indexer(self, key)
 
         if is_scalar(value):
@@ -520,6 +524,9 @@ def to_numpy(
             with warnings.catch_warnings():
                 warnings.filterwarnings("ignore", category=RuntimeWarning)
                 data = self._data.astype(dtype, copy=copy)
+            if self._readonly and astype_is_view(self.dtype, dtype):
+                data = data.view()
+                data.flags.writeable = False
         return data
 
     @doc(ExtensionArray.tolist)
@@ -596,7 +603,12 @@ def __array__(
         if copy is False:
             if not self._hasna:
                 # special case, here we can simply return the underlying data
-                return np.array(self._data, dtype=dtype, copy=copy)
+                result = np.array(self._data, dtype=dtype, copy=copy)
+                # If the ExtensionArray is readonly, make the numpy array readonly too
+                if self._readonly:
+                    result = result.view()
+                    result.flags.writeable = False
+                return result
             raise ValueError(
                 "Unable to avoid copy while creating an array as requested."
             )
diff --git a/pandas/core/arrays/numpy_.py b/pandas/core/arrays/numpy_.py
@@ -12,7 +12,10 @@
 from pandas._libs.tslibs import is_supported_dtype
 from pandas.compat.numpy import function as nv
 
-from pandas.core.dtypes.astype import astype_array
+from pandas.core.dtypes.astype import (
+    astype_array,
+    astype_is_view,
+)
 from pandas.core.dtypes.cast import construct_1d_object_array_from_listlike
 from pandas.core.dtypes.common import pandas_dtype
 from pandas.core.dtypes.dtypes import NumpyEADtype
@@ -160,8 +163,19 @@ def __array__(
     ) -> np.ndarray:
         if copy is not None:
             # Note: branch avoids `copy=None` for NumPy 1.x support
-            return np.array(self._ndarray, dtype=dtype, copy=copy)
-        return np.asarray(self._ndarray, dtype=dtype)
+            result = np.array(self._ndarray, dtype=dtype, copy=copy)
+        else:
+            result = np.asarray(self._ndarray, dtype=dtype)
+
+        if (
+            self._readonly
+            and not copy
+            and (dtype is None or astype_is_view(self.dtype, dtype))
+        ):
+            result = result.view()
+            result.flags.writeable = False
+
+        return result
 
     def __array_ufunc__(self, ufunc: np.ufunc, method: str, *inputs, **kwargs):
         # Lightly modified version of
@@ -512,6 +526,9 @@ def to_numpy(
             result[mask] = na_value
         else:
             result = self._ndarray
+            if not copy and self._readonly:
+                result = result.view()
+                result.flags.writeable = False
 
         result = np.asarray(result, dtype=dtype)
 
diff --git a/pandas/core/arrays/period.py b/pandas/core/arrays/period.py
@@ -393,7 +393,11 @@ def __array__(
             # For NumPy 1.x compatibility we cannot use copy=None.  And
             # `copy=False` has the meaning of `copy=None` here:
             if not copy:
-                return np.asarray(self.asi8, dtype=dtype)
+                result = np.asarray(self.asi8, dtype=dtype)
+                if self._readonly:
+                    result = result.view()
+                    result.flags.writeable = False
+                return result
             else:
                 return np.array(self.asi8, dtype=dtype)
 
diff --git a/pandas/core/arrays/sparse/array.py b/pandas/core/arrays/sparse/array.py
@@ -562,7 +562,11 @@ def __array__(
             if copy is True:
                 return np.array(self.sp_values)
             else:
-                return self.sp_values
+                result = self.sp_values
+                if self._readonly:
+                    result = result.view()
+                    result.flags.writeable = False
+                return result
 
         if copy is False:
             raise ValueError(
@@ -591,6 +595,8 @@ def __array__(
         return out
 
     def __setitem__(self, key, value) -> None:
+        if self._readonly:
+            raise ValueError("Cannot modify readonly array")
         # I suppose we could allow setting of non-fill_value elements.
         # TODO(SparseArray.__setitem__): remove special cases in
         # ExtensionBlock.where
diff --git a/pandas/core/arrays/string_.py b/pandas/core/arrays/string_.py
@@ -769,6 +769,9 @@ def _maybe_convert_setitem_value(self, value):
         return value
 
     def __setitem__(self, key, value) -> None:
+        if self._readonly:
+            raise ValueError("Cannot modify readonly array")
+
         value = self._maybe_convert_setitem_value(value)
 
         key = check_array_indexer(self, key)
diff --git a/pandas/core/dtypes/astype.py b/pandas/core/dtypes/astype.py
@@ -24,8 +24,11 @@
     pandas_dtype,
 )
 from pandas.core.dtypes.dtypes import (
+    CategoricalDtype,
     ExtensionDtype,
+    IntervalDtype,
     NumpyEADtype,
+    PeriodDtype,
 )
 
 if TYPE_CHECKING:
@@ -283,6 +286,14 @@ def astype_is_view(dtype: DtypeObj, new_dtype: DtypeObj) -> bool:
         new_dtype = getattr(new_dtype, "numpy_dtype", new_dtype)
         return getattr(dtype, "unit", None) == getattr(new_dtype, "unit", None)
 
+    elif new_dtype == object and isinstance(dtype, (PeriodDtype, IntervalDtype)):
+        return False
+
+    elif isinstance(dtype, CategoricalDtype) and not isinstance(
+        new_dtype, CategoricalDtype
+    ):
+        return False
+
     numpy_dtype = getattr(dtype, "numpy_dtype", None)
     new_numpy_dtype = getattr(new_dtype, "numpy_dtype", None)
 
diff --git a/pandas/core/indexes/base.py b/pandas/core/indexes/base.py
@@ -4972,6 +4972,8 @@ def array(self) -> ExtensionArray:
             from pandas.core.arrays.numpy_ import NumpyExtensionArray
 
             array = NumpyExtensionArray(array)
+        array = array.view()
+        array._readonly = True
         return array
 
     @property
diff --git a/pandas/core/internals/blocks.py b/pandas/core/internals/blocks.py
@@ -2359,7 +2359,9 @@ def external_values(values: ArrayLike) -> ArrayLike:
     if isinstance(values, np.ndarray):
         values = values.view()
         values.flags.writeable = False
-
-    # TODO(CoW) we should also mark our ExtensionArrays as read-only
+    else:
+        # ExtensionArrays
+        values = values.view()
+        values._readonly = True
 
     return values
diff --git a/pandas/core/series.py b/pandas/core/series.py
@@ -820,7 +820,10 @@ def _references(self) -> BlockValuesRefs:
     @Appender(base.IndexOpsMixin.array.__doc__)  # type: ignore[prop-decorator]
     @property
     def array(self) -> ExtensionArray:
-        return self._mgr.array_values()
+        arr = self._mgr.array_values()
+        arr = arr.view()
+        arr._readonly = True
+        return arr
 
     def __len__(self) -> int:
         """
diff --git a/pandas/tests/extension/base/setitem.py b/pandas/tests/extension/base/setitem.py
@@ -454,3 +454,60 @@ def test_setitem_2d_values(self, data):
         df.loc[[0, 1], :] = df.loc[[1, 0], :].values
         assert (df.loc[0, :] == original[1]).all()
         assert (df.loc[1, :] == original[0]).all()
+
+    def test_readonly_property(self, data):
+        assert data.readonly is False
+
+        data.readonly = True
+        assert data.readonly is True
+
+        data_orig = data.copy()
+        assert data_orig.readonly is False
+
+        with pytest.raises(ValueError, match="Cannot modify readonly array"):
+            data[0] = data[1]
+
+        with pytest.raises(ValueError, match="Cannot modify readonly array"):
+            data[0:3] = data[1]
+
+        with pytest.raises(ValueError, match="Cannot modify readonly array"):
+            data[np.array([True] * len(data))] = data[1]
+
+        tm.assert_extension_array_equal(data, data_orig)
+
+    def test_readonly_propagates_to_numpy_array(self, data):
+        data.readonly = True
+
+        # when we ask for a copy, the result should never be readonly
+        arr = np.array(data)
+        assert arr.flags.writeable
+
+        # when we don't ask for a copy -> if the conversion is zero-copy,
+        # the result should be readonly
+        arr1 = np.asarray(data)
+        arr2 = np.asarray(data)
+        if np.shares_memory(arr1, arr2):
+            assert not arr1.flags.writeable
+        else:
+            assert arr1.flags.writeable
+
+    def test_readonly_propagates_to_numpy_array_method(self, data):
+        data.readonly = True
+
+        # when we ask for a copy, the result should never be readonly
+        arr = data.to_numpy(copy=True)
+        assert arr.flags.writeable
+
+        # when we don't ask for a copy -> if the conversion is zero-copy,
+        # the result should be readonly
+        arr1 = data.to_numpy(copy=False)
+        arr2 = data.to_numpy(copy=False)
+        if np.shares_memory(arr1, arr2):
+            assert not arr1.flags.writeable
+        else:
+            assert arr1.flags.writeable
+
+        # non-NA fill value should always result in a copy
+        if data.isna().any():
+            arr = data.to_numpy(copy=False, na_value=data[0])
+            assert arr.flags.writeable
diff --git a/pandas/tests/extension/date/array.py b/pandas/tests/extension/date/array.py
diff --git a/pandas/tests/extension/decimal/array.py b/pandas/tests/extension/decimal/array.py
diff --git a/pandas/tests/extension/json/array.py b/pandas/tests/extension/json/array.py
