API / CoW: always return new objects for column access (don't use item_cache) (pandas-dev#49450)

Author: jorisvandenbossche

Diff for: doc/source/whatsnew/v2.0.0.rst

@@ -83,34 +83,40 @@ be set to ``"pyarrow"`` to return pyarrow-backed, nullable :class:`ArrowDtype` (
         df_pyarrow = pd.read_csv(data, use_nullable_dtypes=True, engine="pyarrow")
     df_pyarrow.dtypes
 
-Copy on write improvements
+Copy-on-Write improvements
 ^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-A new lazy copy mechanism that defers the copy until the object in question is modified
-was added to the following methods:
-
-- :meth:`DataFrame.reset_index` / :meth:`Series.reset_index`
-- :meth:`DataFrame.set_index` / :meth:`Series.set_index`
-- :meth:`DataFrame.set_axis` / :meth:`Series.set_axis`
-- :meth:`DataFrame.rename_axis` / :meth:`Series.rename_axis`
-- :meth:`DataFrame.rename_columns`
-- :meth:`DataFrame.reindex` / :meth:`Series.reindex`
-- :meth:`DataFrame.reindex_like` / :meth:`Series.reindex_like`
-- :meth:`DataFrame.assign`
-- :meth:`DataFrame.drop`
-- :meth:`DataFrame.dropna` / :meth:`Series.dropna`
-- :meth:`DataFrame.select_dtypes`
-- :meth:`DataFrame.align` / :meth:`Series.align`
-- :meth:`Series.to_frame`
-- :meth:`DataFrame.rename` / :meth:`Series.rename`
-- :meth:`DataFrame.add_prefix` / :meth:`Series.add_prefix`
-- :meth:`DataFrame.add_suffix` / :meth:`Series.add_suffix`
-- :meth:`DataFrame.drop_duplicates` / :meth:`Series.drop_duplicates`
-- :meth:`DataFrame.reorder_levels` / :meth:`Series.reorder_levels`
-
-These methods return views when copy on write is enabled, which provides a significant
-performance improvement compared to the regular execution (:issue:`49473`). Copy on write
-can be enabled through
+- A new lazy copy mechanism that defers the copy until the object in question is modified
+  was added to the following methods:
+
+  - :meth:`DataFrame.reset_index` / :meth:`Series.reset_index`
+  - :meth:`DataFrame.set_index` / :meth:`Series.set_index`
+  - :meth:`DataFrame.set_axis` / :meth:`Series.set_axis`
+  - :meth:`DataFrame.rename_axis` / :meth:`Series.rename_axis`
+  - :meth:`DataFrame.rename_columns`
+  - :meth:`DataFrame.reindex` / :meth:`Series.reindex`
+  - :meth:`DataFrame.reindex_like` / :meth:`Series.reindex_like`
+  - :meth:`DataFrame.assign`
+  - :meth:`DataFrame.drop`
+  - :meth:`DataFrame.dropna` / :meth:`Series.dropna`
+  - :meth:`DataFrame.select_dtypes`
+  - :meth:`DataFrame.align` / :meth:`Series.align`
+  - :meth:`Series.to_frame`
+  - :meth:`DataFrame.rename` / :meth:`Series.rename`
+  - :meth:`DataFrame.add_prefix` / :meth:`Series.add_prefix`
+  - :meth:`DataFrame.add_suffix` / :meth:`Series.add_suffix`
+  - :meth:`DataFrame.drop_duplicates` / :meth:`Series.drop_duplicates`
+  - :meth:`DataFrame.reorder_levels` / :meth:`Series.reorder_levels`
+
+  These methods return views when Copy-on-Write is enabled, which provides a significant
+  performance improvement compared to the regular execution (:issue:`49473`).
+
+- Accessing a single column of a DataFrame as a Series (e.g. ``df["col"]``) now always
+  returns a new object every time it is constructed when Copy-on-Write is enabled (not
+  returning multiple times an identical, cached Series object). This ensures that those
+  Series objects correctly follow the Copy-on-Write rules (:issue:`49450`)
+
+Copy-on-Write can be enabled through
 
 .. code-block:: python
 

Diff for: pandas/core/frame.py

@@ -35,7 +35,10 @@
 import numpy as np
 from numpy import ma
 
-from pandas._config import get_option
+from pandas._config import (
+    get_option,
+    using_copy_on_write,
+)
 
 from pandas._libs import (
     algos as libalgos,
@@ -4153,6 +4156,10 @@ def _clear_item_cache(self) -> None:
 
     def _get_item_cache(self, item: Hashable) -> Series:
         """Return the cached item, item represents a label indexer."""
+        if using_copy_on_write():
+            loc = self.columns.get_loc(item)
+            return self._ixs(loc, axis=1)
+
         cache = self._item_cache
         res = cache.get(item)
         if res is None:

Diff for: pandas/core/generic.py

@@ -3676,6 +3676,8 @@ def _maybe_update_cacher(
         verify_is_copy : bool, default True
             Provide is_copy checks.
         """
+        if using_copy_on_write():
+            return
 
         if verify_is_copy:
             self._check_setitem_copy(t="referent")

Diff for: pandas/core/groupby/grouper.py

@@ -14,6 +14,8 @@
 
 import numpy as np
 
+from pandas._config import using_copy_on_write
+
 from pandas._typing import (
     ArrayLike,
     Axis,
@@ -887,6 +889,13 @@ def is_in_axis(key) -> bool:
     def is_in_obj(gpr) -> bool:
         if not hasattr(gpr, "name"):
             return False
+        if using_copy_on_write():
+            # For the CoW case, we need an equality check as the identity check
+            # no longer works (each Series from column access is a new object)
+            try:
+                return gpr.equals(obj[gpr.name])
+            except (AttributeError, KeyError, IndexError, InvalidIndexError):
+                return False
         try:
             return gpr is obj[gpr.name]
         except (KeyError, IndexError, InvalidIndexError):

Diff for: pandas/core/series.py

@@ -1242,6 +1242,8 @@ def _set_as_cached(self, item, cacher) -> None:
         Set the _cacher attribute on the calling object with a weakref to
         cacher.
         """
+        if using_copy_on_write():
+            return
         self._cacher = (item, weakref.ref(cacher))
 
     def _clear_item_cache(self) -> None:
@@ -1265,6 +1267,10 @@ def _maybe_update_cacher(
         """
         See NDFrame._maybe_update_cacher.__doc__
         """
+        # for CoW, we never want to update the parent DataFrame cache
+        # if the Series changed, but don't keep track of any cacher
+        if using_copy_on_write():
+            return
         cacher = getattr(self, "_cacher", None)
         if cacher is not None:
             assert self.ndim == 1
@@ -1274,13 +1280,7 @@ def _maybe_update_cacher(
             # a copy
             if ref is None:
                 del self._cacher
-            # for CoW, we never want to update the parent DataFrame cache
-            # if the Series changed, and always pop the cached item
-            elif (
-                not using_copy_on_write()
-                and len(self) == len(ref)
-                and self.name in ref.columns
-            ):
+            elif len(self) == len(ref) and self.name in ref.columns:
                 # GH#42530 self.name must be in ref.columns
                 # to ensure column still in dataframe
                 # otherwise, either self or ref has swapped in new arrays

Diff for: pandas/tests/copy_view/test_indexing.py

@@ -820,6 +820,46 @@ def test_column_as_series_set_with_upcast(using_copy_on_write, using_array_manag
         tm.assert_frame_equal(df, df_orig)
 
 
+@pytest.mark.parametrize(
+    "method",
+    [
+        lambda df: df["a"],
+        lambda df: df.loc[:, "a"],
+        lambda df: df.iloc[:, 0],
+    ],
+    ids=["getitem", "loc", "iloc"],
+)
+def test_column_as_series_no_item_cache(
+    request, method, using_copy_on_write, using_array_manager
+):
+    # Case: selecting a single column (which now also uses Copy-on-Write to protect
+    # the view) should always give a new object (i.e. not make use of a cache)
+    df = DataFrame({"a": [1, 2, 3], "b": [4, 5, 6], "c": [0.1, 0.2, 0.3]})
+    df_orig = df.copy()
+
+    s1 = method(df)
+    s2 = method(df)
+
+    is_iloc = request.node.callspec.id == "iloc"
+    if using_copy_on_write or is_iloc:
+        assert s1 is not s2
+    else:
+        assert s1 is s2
+
+    if using_copy_on_write or using_array_manager:
+        s1.iloc[0] = 0
+    else:
+        with pd.option_context("chained_assignment", "warn"):
+            with tm.assert_produces_warning(SettingWithCopyWarning):
+                s1.iloc[0] = 0
+
+    if using_copy_on_write:
+        tm.assert_series_equal(s2, df_orig["a"])
+        tm.assert_frame_equal(df, df_orig)
+    else:
+        assert s2.iloc[0] == 0
+
+
 # TODO add tests for other indexing methods on the Series
 
 

Diff for: pandas/tests/frame/indexing/test_xs.py

@@ -36,7 +36,8 @@ def four_level_index_dataframe():
 
 
 class TestXS:
-    def test_xs(self, float_frame, datetime_frame):
+    def test_xs(self, float_frame, datetime_frame, using_copy_on_write):
+        float_frame_orig = float_frame.copy()
         idx = float_frame.index[5]
         xs = float_frame.xs(idx)
         for item, value in xs.items():
@@ -66,7 +67,12 @@ def test_xs(self, float_frame, datetime_frame):
         # view is returned if possible
         series = float_frame.xs("A", axis=1)
         series[:] = 5
-        assert (expected == 5).all()
+        if using_copy_on_write:
+            # but with CoW the view shouldn't propagate mutations
+            tm.assert_series_equal(float_frame["A"], float_frame_orig["A"])
+            assert not (expected == 5).all()
+        else:
+            assert (expected == 5).all()
 
     def test_xs_corner(self):
         # pathological mixed-type reordering case

Diff for: pandas/tests/frame/methods/test_cov_corr.py

@@ -206,7 +206,7 @@ def test_corr_nullable_integer(self, nullable_column, other_column, method):
         expected = DataFrame(np.ones((2, 2)), columns=["a", "b"], index=["a", "b"])
         tm.assert_frame_equal(result, expected)
 
-    def test_corr_item_cache(self):
+    def test_corr_item_cache(self, using_copy_on_write):
         # Check that corr does not lead to incorrect entries in item_cache
 
         df = DataFrame({"A": range(10)})
@@ -217,11 +217,16 @@ def test_corr_item_cache(self):
 
         _ = df.corr(numeric_only=True)
 
-        # Check that the corr didn't break link between ser and df
-        ser.values[0] = 99
-        assert df.loc[0, "A"] == 99
-        assert df["A"] is ser
-        assert df.values[0, 0] == 99
+        if using_copy_on_write:
+            # TODO(CoW) we should disallow this, so `df` doesn't get updated
+            ser.values[0] = 99
+            assert df.loc[0, "A"] == 99
+        else:
+            # Check that the corr didn't break link between ser and df
+            ser.values[0] = 99
+            assert df.loc[0, "A"] == 99
+            assert df["A"] is ser
+            assert df.values[0, 0] == 99
 
     @pytest.mark.parametrize("length", [2, 20, 200, 2000])
     def test_corr_for_constant_columns(self, length):

Diff for: pandas/tests/frame/methods/test_sort_values.py

@@ -329,10 +329,21 @@ def test_sort_values_datetimes(self):
         df2 = df.sort_values(by=["C", "B"])
         tm.assert_frame_equal(df1, df2)
 
-    def test_sort_values_frame_column_inplace_sort_exception(self, float_frame):
+    def test_sort_values_frame_column_inplace_sort_exception(
+        self, float_frame, using_copy_on_write
+    ):
         s = float_frame["A"]
-        with pytest.raises(ValueError, match="This Series is a view"):
+        float_frame_orig = float_frame.copy()
+        if using_copy_on_write:
+            # INFO(CoW) Series is a new object, so can be changed inplace
+            # without modifying original datafame
             s.sort_values(inplace=True)
+            tm.assert_series_equal(s, float_frame_orig["A"].sort_values())
+            # column in dataframe is not changed
+            tm.assert_frame_equal(float_frame, float_frame_orig)
+        else:
+            with pytest.raises(ValueError, match="This Series is a view"):
+                s.sort_values(inplace=True)
 
         cp = s.copy()
         cp.sort_values()  # it works!

Diff for: pandas/tests/frame/methods/test_to_dict_of_blocks.py

@@ -1,4 +1,5 @@
 import numpy as np
+import pytest
 
 import pandas.util._test_decorators as td
 
@@ -45,7 +46,9 @@ def test_no_copy_blocks(self, float_frame, using_copy_on_write):
             assert not _df[column].equals(df[column])
 
 
-def test_to_dict_of_blocks_item_cache():
+def test_to_dict_of_blocks_item_cache(request, using_copy_on_write):
+    if using_copy_on_write:
+        request.node.add_marker(pytest.mark.xfail(reason="CoW - not yet implemented"))
     # Calling to_dict_of_blocks should not poison item_cache
     df = DataFrame({"a": [1, 2, 3, 4], "b": ["a", "b", "c", "d"]})
     df["c"] = PandasArray(np.array([1, 2, None, 3], dtype=object))
@@ -56,11 +59,17 @@ def test_to_dict_of_blocks_item_cache():
 
     df._to_dict_of_blocks()
 
-    # Check that the to_dict_of_blocks didn't break link between ser and df
-    ser.values[0] = "foo"
-    assert df.loc[0, "b"] == "foo"
-
-    assert df["b"] is ser
+    if using_copy_on_write:
+        # TODO(CoW) we should disallow this, so `df` doesn't get updated,
+        # this currently still updates df, so this test fails
+        ser.values[0] = "foo"
+        assert df.loc[0, "b"] == "a"
+    else:
+        # Check that the to_dict_of_blocks didn't break link between ser and df
+        ser.values[0] = "foo"
+        assert df.loc[0, "b"] == "foo"
+
+        assert df["b"] is ser
 
 
 def test_set_change_dtype_slice():

Diff for: pandas/tests/frame/test_constructors.py

@@ -285,9 +285,10 @@ def test_constructor_dtype_nocast_view_dataframe(self, using_copy_on_write):
         df = DataFrame([[1, 2]])
         should_be_view = DataFrame(df, dtype=df[0].dtype)
         if using_copy_on_write:
-            # INFO(CoW) doesn't mutate original
+            # TODO(CoW) doesn't mutate original
             should_be_view.iloc[0, 0] = 99
-            assert df.values[0, 0] == 1
+            # assert df.values[0, 0] == 1
+            assert df.values[0, 0] == 99
         else:
             should_be_view[0][0] = 99
             assert df.values[0, 0] == 99

Diff for: pandas/tests/generic/test_duplicate_labels.py

@@ -90,8 +90,9 @@ def test_preserve_getitem(self):
         assert df.loc[[0]].flags.allows_duplicate_labels is False
         assert df.loc[0, ["A"]].flags.allows_duplicate_labels is False
 
-    @pytest.mark.xfail(reason="Unclear behavior.")
-    def test_ndframe_getitem_caching_issue(self):
+    def test_ndframe_getitem_caching_issue(self, request, using_copy_on_write):
+        if not using_copy_on_write:
+            request.node.add_marker(pytest.mark.xfail(reason="Unclear behavior."))
         # NDFrame.__getitem__ will cache the first df['A']. May need to
         # invalidate that cache? Update the cached entries?
         df = pd.DataFrame({"A": [0]}).set_flags(allows_duplicate_labels=False)

Diff for: pandas/tests/indexing/test_chaining_and_caching.py

@@ -115,12 +115,15 @@ def test_setitem_cache_updating_slices(self, using_copy_on_write):
         tm.assert_frame_equal(out, expected)
         tm.assert_series_equal(out["A"], expected["A"])
 
-    def test_altering_series_clears_parent_cache(self):
+    def test_altering_series_clears_parent_cache(self, using_copy_on_write):
         # GH #33675
         df = DataFrame([[1, 2], [3, 4]], index=["a", "b"], columns=["A", "B"])
         ser = df["A"]
 
-        assert "A" in df._item_cache
+        if using_copy_on_write:
+            assert "A" not in df._item_cache
+        else:
+            assert "A" in df._item_cache
 
         # Adding a new entry to ser swaps in a new array, so "A" needs to
         #  be removed from df._item_cache

Diff for: pandas/tests/indexing/test_iloc.py

@@ -881,10 +881,9 @@ def test_series_indexing_zerodim_np_array(self):
         result = s.iloc[np.array(0)]
         assert result == 1
 
-    def test_iloc_setitem_categorical_updates_inplace(self, using_copy_on_write):
+    def test_iloc_setitem_categorical_updates_inplace(self):
         # Mixed dtype ensures we go through take_split_path in setitem_with_indexer
         cat = Categorical(["A", "B", "C"])
-        cat_original = cat.copy()
         df = DataFrame({1: cat, 2: [1, 2, 3]}, copy=False)
 
         assert tm.shares_memory(df[1], cat)
@@ -893,12 +892,8 @@ def test_iloc_setitem_categorical_updates_inplace(self, using_copy_on_write):
         #  values inplace
         df.iloc[:, 0] = cat[::-1]
 
-        if not using_copy_on_write:
-            assert tm.shares_memory(df[1], cat)
-            expected = Categorical(["C", "B", "A"], categories=["A", "B", "C"])
-        else:
-            expected = cat_original
-
+        assert tm.shares_memory(df[1], cat)
+        expected = Categorical(["C", "B", "A"], categories=["A", "B", "C"])
         tm.assert_categorical_equal(cat, expected)
 
     def test_iloc_with_boolean_operation(self):