API: New copy / view semantics using Copy-on-Write

.github/workflows/ubuntu.yml

@@ -52,6 +52,10 @@ jobs:
             extra_apt: "language-pack-zh-hans"
             lang: "zh_CN.utf8"
             lc_all: "zh_CN.utf8"
+          - name: "Copy-on-Write"
+            env_file: actions-310.yaml
+            pattern: "not slow and not network and not single_cpu"
+            pandas_copy_on_write: "1"
           - name: "Data Manager"
             env_file: actions-38.yaml
             pattern: "not slow and not network and not single_cpu"
@@ -84,6 +88,7 @@ jobs:
       LC_ALL: ${{ matrix.lc_all || '' }}
       PANDAS_TESTING_MODE: ${{ matrix.pandas_testing_mode || '' }}
       PANDAS_DATA_MANAGER: ${{ matrix.pandas_data_manager || 'block' }}
+      PANDAS_COPY_ON_WRITE: ${{ matrix.pandas_copy_on_write || '0' }}
       TEST_ARGS: ${{ matrix.test_args || '' }}
       PYTEST_WORKERS: ${{ contains(matrix.pattern, 'not single_cpu') && 'auto' || '1' }}
       PYTEST_TARGET: ${{ matrix.pytest_target || 'pandas' }}

pandas/_libs/internals.pyx

@@ -1,4 +1,5 @@
 from collections import defaultdict
+import weakref
 
 cimport cython
 from cpython.slice cimport PySlice_GetIndicesEx
@@ -674,8 +675,9 @@ cdef class BlockManager:
         public list axes
         public bint _known_consolidated, _is_consolidated
         public ndarray _blknos, _blklocs
+        public list refs
 
-    def __cinit__(self, blocks=None, axes=None, verify_integrity=True):
+    def __cinit__(self, blocks=None, axes=None, refs=None, verify_integrity=True):
         # None as defaults for unpickling GH#42345
         if blocks is None:
             # This adds 1-2 microseconds to DataFrame(np.array([]))
@@ -687,6 +689,7 @@ cdef class BlockManager:
 
         self.blocks = blocks
         self.axes = axes.copy()  # copy to make sure we are not remotely-mutable
+        self.refs = refs
 
         # Populate known_consolidate, blknos, and blklocs lazily
         self._known_consolidated = False
@@ -795,12 +798,14 @@ cdef class BlockManager:
             ndarray blknos, blklocs
 
         nbs = []
+        nrefs = []
         for blk in self.blocks:
             nb = blk.getitem_block_index(slobj)
             nbs.append(nb)
+            nrefs.append(weakref.ref(blk))
 
         new_axes = [self.axes[0], self.axes[1]._getitem_slice(slobj)]
-        mgr = type(self)(tuple(nbs), new_axes, verify_integrity=False)
+        mgr = type(self)(tuple(nbs), new_axes, nrefs, verify_integrity=False)
 
         # We can avoid having to rebuild blklocs/blknos
         blklocs = self._blklocs
@@ -813,7 +818,7 @@ cdef class BlockManager:
     def get_slice(self, slobj: slice, axis: int = 0) -> BlockManager:
 
         if axis == 0:
-            new_blocks = self._slice_take_blocks_ax0(slobj)
+            new_blocks, new_refs = self._slice_take_blocks_ax0(slobj)
         elif axis == 1:
             return self._get_index_slice(slobj)
         else:
@@ -822,4 +827,4 @@ cdef class BlockManager:
         new_axes = list(self.axes)
         new_axes[axis] = new_axes[axis]._getitem_slice(slobj)
 
-        return type(self)(tuple(new_blocks), new_axes, verify_integrity=False)
+        return type(self)(tuple(new_blocks), new_axes, new_refs, verify_integrity=False)

pandas/compat/pickle_compat.py

@@ -224,7 +224,7 @@ def load_newobj(self):
         arr = np.array([], dtype="m8[ns]")
         obj = cls.__new__(cls, arr, arr.dtype)
     elif cls is BlockManager and not args:
-        obj = cls.__new__(cls, (), [], False)
+        obj = cls.__new__(cls, (), [], None, False)
     else:
         obj = cls.__new__(cls, *args)
 

pandas/conftest.py

@@ -1840,4 +1840,4 @@ def using_copy_on_write() -> Literal[False]:
     """
     Fixture to check if Copy-on-Write is enabled.
     """
-    return False
+    return pd.options.mode.copy_on_write and pd.options.mode.data_manager == "block"

pandas/core/config_init.py

@@ -539,6 +539,26 @@ def use_inf_as_na_cb(key) -> None:
     )
 
 
+# TODO better name?
+copy_on_write_doc = """
+: bool
+    Use new copy-view behaviour using Copy-on-Write. Defaults to False,
+    unless overridden by the 'PANDAS_COPY_ON_WRITE' environment variable (needs
+    to be set before pandas is imported).
+"""
+
+
+with cf.config_prefix("mode"):
+    cf.register_option(
+        "copy_on_write",
+        # Get the default from an environment variable, if set, otherwise defaults
+        # to False. This environment variable can be set for testing.
+        os.environ.get("PANDAS_COPY_ON_WRITE", "0") == "1",
+        copy_on_write_doc,
+        validator=is_bool,
+    )
+
+
 # user warnings
 chained_assignment = """
 : string

pandas/core/frame.py

@@ -3677,13 +3677,19 @@ def _get_column_array(self, i: int) -> ArrayLike:
         """
         Get the values of the i'th column (ndarray or ExtensionArray, as stored
         in the Block)
+
+        Warning! The returned array is a view but doesn't handle Copy-on-Write,
+        so this should be used with caution (for read-only purposes).
         """
         return self._mgr.iget_values(i)
 
     def _iter_column_arrays(self) -> Iterator[ArrayLike]:
         """
         Iterate over the arrays of all columns in order.
         This returns the values as stored in the Block (ndarray or ExtensionArray).
+
+        Warning! The returned array is a view but doesn't handle Copy-on-Write,
+        so this should be used with caution (for read-only purposes).
         """
         for i in range(len(self.columns)):
             yield self._get_column_array(i)
@@ -5070,7 +5076,7 @@ def set_axis(self, labels, axis: Axis = 0, inplace: bool = False):
         "labels",
         [
             ("method", None),
-            ("copy", True),
+            ("copy", None),
             ("level", None),
             ("fill_value", np.nan),
             ("limit", None),
@@ -5295,7 +5301,7 @@ def rename(
         index: Renamer | None = ...,
         columns: Renamer | None = ...,
         axis: Axis | None = ...,
-        copy: bool = ...,
+        copy: bool | None = ...,
         inplace: Literal[True],
         level: Level | None = ...,
         errors: IgnoreRaise = ...,
@@ -5310,7 +5316,7 @@ def rename(
         index: Renamer | None = ...,
         columns: Renamer | None = ...,
         axis: Axis | None = ...,
-        copy: bool = ...,
+        copy: bool | None = ...,
         inplace: Literal[False] = ...,
         level: Level | None = ...,
         errors: IgnoreRaise = ...,
@@ -5325,7 +5331,7 @@ def rename(
         index: Renamer | None = ...,
         columns: Renamer | None = ...,
         axis: Axis | None = ...,
-        copy: bool = ...,
+        copy: bool | None = ...,
         inplace: bool = ...,
         level: Level | None = ...,
         errors: IgnoreRaise = ...,
@@ -5339,7 +5345,7 @@ def rename(
         index: Renamer | None = None,
         columns: Renamer | None = None,
         axis: Axis | None = None,
-        copy: bool = True,
+        copy: bool | None = None,
         inplace: bool = False,
         level: Level | None = None,
         errors: IgnoreRaise = "ignore",
@@ -6294,7 +6300,7 @@ class    max    type
         if inplace:
             new_obj = self
         else:
-            new_obj = self.copy()
+            new_obj = self.copy(deep=None)
         if allow_duplicates is not lib.no_default:
             allow_duplicates = validate_bool_kwarg(allow_duplicates, "allow_duplicates")
 

pandas/core/generic.py

@@ -995,7 +995,7 @@ def _rename(
         index: Renamer | None = None,
         columns: Renamer | None = None,
         axis: Axis | None = None,
-        copy: bool_t = True,
+        copy: bool_t | None = None,
         inplace: bool_t = False,
         level: Level | None = None,
         errors: str = "ignore",
@@ -4101,6 +4101,12 @@ def _check_setitem_copy(self, t="setting", force=False):
         df.iloc[0:5]['group'] = 'a'
 
         """
+        if (
+            config.get_option("mode.copy_on_write")
+            and config.get_option("mode.data_manager") == "block"
+        ):
+            return
+
         # return early if the check is not needed
         if not (force or self._is_copy):
             return
@@ -5197,7 +5203,7 @@ def reindex(self: NDFrameT, *args, **kwargs) -> NDFrameT:
         axes, kwargs = self._construct_axes_from_arguments(args, kwargs)
         method = missing.clean_reindex_fill_method(kwargs.pop("method", None))
         level = kwargs.pop("level", None)
-        copy = kwargs.pop("copy", True)
+        copy = kwargs.pop("copy", None)
         limit = kwargs.pop("limit", None)
         tolerance = kwargs.pop("tolerance", None)
         fill_value = kwargs.pop("fill_value", None)
@@ -5222,9 +5228,7 @@ def reindex(self: NDFrameT, *args, **kwargs) -> NDFrameT:
             for axis, ax in axes.items()
             if ax is not None
         ):
-            if copy:
-                return self.copy()
-            return self
+            return self.copy(deep=copy)
 
         # check if we are a multi reindex
         if self._needs_reindex_multi(axes, method, level):
@@ -6201,7 +6205,7 @@ def astype(
         return cast(NDFrameT, result)
 
     @final
-    def copy(self: NDFrameT, deep: bool_t = True) -> NDFrameT:
+    def copy(self: NDFrameT, deep: bool_t | None = True) -> NDFrameT:
         """
         Make a copy of this object's indices and data.
 

pandas/core/internals/array_manager.py

@@ -527,6 +527,11 @@ def copy(self: T, deep=True) -> T:
         -------
         BlockManager
         """
+        if deep is None:
+            # ArrayManager does not yet support CoW, so deep=None always means
+            # deep=True for now
+            deep = True
+
         # this preserves the notion of view copying of axes
         if deep:
             # hit in e.g. tests.io.json.test_pandas
@@ -591,6 +596,11 @@ def _reindex_indexer(
 
         pandas-indexer with -1's only.
         """
+        if copy is None:
+            # ArrayManager does not yet support CoW, so deep=None always means
+            # deep=True for now
+            copy = True
+
         if indexer is None:
             if new_axis is self._axes[axis] and not copy:
                 return self

pandas/core/internals/blocks.py

@@ -838,17 +838,22 @@ def _slice(
 
         return self.values[slicer]
 
-    def set_inplace(self, locs, values: ArrayLike) -> None:
+    def set_inplace(self, locs, values: ArrayLike, copy: bool = False) -> None:
         """
         Modify block values in-place with new item value.
 
+        If copy=True, first copy the underlying values in place before modifying
+        (for Copy-on-Write).
+
         Notes
         -----
         `set_inplace` never creates a new array or new Block, whereas `setitem`
         _may_ create a new array and always creates a new Block.
 
         Caller is responsible for checking values.dtype == self.dtype.
         """
+        if copy:
+            self.values = self.values.copy()
         self.values[locs] = values
 
     def take_nd(
@@ -1664,9 +1669,11 @@ def iget(self, i: int | tuple[int, int] | tuple[slice, int]):
                 raise IndexError(f"{self} only contains one item")
             return self.values
 
-    def set_inplace(self, locs, values: ArrayLike) -> None:
+    def set_inplace(self, locs, values: ArrayLike, copy: bool = False) -> None:
         # When an ndarray, we should have locs.tolist() == [0]
         # When a BlockPlacement we should have list(locs) == [0]
+        if copy:
+            self.values = self.values.copy()
         self.values[:] = values
 
     def _maybe_squeeze_arg(self, arg):