Copy-on-Write (CoW)#
Note
Copy-on-Write is now the default with pandas 3.0.
Copy-on-Write was first introduced in version 1.5.0. Starting from version 2.0 most of the optimizations that become possible through CoW are implemented and supported. All possible optimizations are supported starting from pandas 2.1.
CoW will lead to more predictable behavior since it is not possible to update more than one object with one statement, e.g. indexing operations or methods won’t have side-effects. Additionally, through delaying copies as long as possible, the average performance and memory usage will improve.
Previous behavior#
pandas indexing behavior is tricky to understand. Some operations return views while other return copies. Depending on the result of the operation, mutating one object might accidentally mutate another:
In [1]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [2]: subset = df["foo"]
In [3]: subset.iloc[0] = 100
In [4]: df
Out[4]:
foo bar
0 100 4
1 2 5
2 3 6
Mutating subset, e.g. updating its values, also updated df. The exact behavior was
hard to predict. Copy-on-Write solves accidentally modifying more than one object,
it explicitly disallows this. df is unchanged:
In [1]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [2]: subset = df["foo"]
In [3]: subset.iloc[0] = 100
In [4]: df
Out[4]:
foo bar
0 1 4
1 2 5
2 3 6
The following sections will explain what this means and how it impacts existing applications.
Migrating to Copy-on-Write#
Copy-on-Write is the default and only mode in pandas 3.0. This means that users need to migrate their code to be compliant with CoW rules.
The default mode in pandas < 3.0 raises warnings for certain cases that will actively change behavior and thus change user intended behavior.
pandas 2.2 has a warning mode
pd.options.mode.copy_on_write = "warn"
that will warn for every operation that will change behavior with CoW. We expect this mode to be very noisy, since many cases that we don’t expect that they will influence users will also emit a warning. We recommend checking this mode and analyzing the warnings, but it is not necessary to address all of these warning. The first two items of the following lists are the only cases that need to be addressed to make existing code work with CoW.
The following few items describe the user visible changes:
Chained assignment will never work
loc should be used as an alternative. Check the
chained assignment section for more details.
Accessing the underlying array of a pandas object will return a read-only view
In [5]: ser = pd.Series([1, 2, 3])
In [6]: ser.to_numpy()
Out[6]: array([1, 2, 3])
This example returns a NumPy array that is a view of the Series object. This view can be modified and thus also modify the pandas object. This is not compliant with CoW rules. The returned array is set to non-writeable to protect against this behavior. Creating a copy of this array allows modification. You can also make the array writeable again if you don’t care about the pandas object anymore.
See the section about read-only NumPy arrays for more details.
Only one pandas object is updated at once
The following code snippet updated both df and subset without CoW:
In [1]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [2]: subset = df["foo"]
In [3]: subset.iloc[0] = 100
In [4]: df
Out[4]:
foo bar
0 100 4
1 2 5
2 3 6
This is not possible anymore with CoW, since the CoW rules explicitly forbid this.
This includes updating a single column as a Series and relying on the change
propagating back to the parent DataFrame.
This statement can be rewritten into a single statement with loc or iloc if
this behavior is necessary. DataFrame.where() is another suitable alternative
for this case.
Updating a column selected from a DataFrame with an inplace method will
also not work anymore.
In [7]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [8]: df["foo"].replace(1, 5, inplace=True)
In [9]: df
Out[9]:
foo bar
0 1 4
1 2 5
2 3 6
This is another form of chained assignment. This can generally be rewritten in 2 different forms:
In [10]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [11]: df.replace({"foo": {1: 5}}, inplace=True)
In [12]: df
Out[12]:
foo bar
0 5 4
1 2 5
2 3 6
A different alternative would be to not use inplace:
In [13]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [14]: df["foo"] = df["foo"].replace(1, 5)
In [15]: df
Out[15]:
foo bar
0 5 4
1 2 5
2 3 6
Constructors now copy NumPy arrays by default
The Series and DataFrame constructors now copies a NumPy array by default when not
otherwise specified. This was changed to avoid mutating a pandas object when the
NumPy array is changed inplace outside of pandas. You can set copy=False to
avoid this copy.
Description#
CoW means that any DataFrame or Series derived from another in any way always behaves as a copy. As a consequence, we can only change the values of an object through modifying the object itself. CoW disallows updating a DataFrame or a Series that shares data with another DataFrame or Series object inplace.
This avoids side-effects when modifying values and hence, most methods can avoid actually copying the data and only trigger a copy when necessary.
The following example will operate inplace:
In [16]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [17]: df.iloc[0, 0] = 100
In [18]: df
Out[18]:
foo bar
0 100 4
1 2 5
2 3 6
The object df does not share any data with any other object and hence no
copy is triggered when updating the values. In contrast, the following operation
triggers a copy of the data under CoW:
In [19]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [20]: df2 = df.reset_index(drop=True)
In [21]: df2.iloc[0, 0] = 100
In [22]: df
Out[22]:
foo bar
0 1 4
1 2 5
2 3 6
In [23]: df2
Out[23]:
foo bar
0 100 4
1 2 5
2 3 6
reset_index returns a lazy copy with CoW while it copies the data without CoW.
Since both objects, df and df2 share the same data, a copy is triggered
when modifying df2. The object df still has the same values as initially
while df2 was modified.
If the object df isn’t needed anymore after performing the reset_index operation,
you can emulate an inplace-like operation through assigning the output of reset_index
to the same variable:
In [24]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [25]: df = df.reset_index(drop=True)
In [26]: df.iloc[0, 0] = 100
In [27]: df
Out[27]:
foo bar
0 100 4
1 2 5
2 3 6
The initial object gets out of scope as soon as the result of reset_index is
reassigned and hence df does not share data with any other object. No copy
is necessary when modifying the object. This is generally true for all methods
listed in Copy-on-Write optimizations.
Previously, when operating on views, the view and the parent object was modified:
In [1]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [2]: subset = df["foo"]
In [3]: subset.iloc[0] = 100
In [4]: df
Out[4]:
foo bar
0 100 4
1 2 5
2 3 6
CoW triggers a copy when df is changed to avoid mutating view as well:
In [28]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [29]: view = df[:]
In [30]: df.iloc[0, 0] = 100
In [31]: df
Out[31]:
foo bar
0 100 4
1 2 5
2 3 6
In [32]: view
Out[32]:
foo bar
0 1 4
1 2 5
2 3 6
Chained Assignment#
Chained assignment references a technique where an object is updated through two subsequent indexing operations, e.g.
In [1]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [2]: df["foo"][df["bar"] > 5] = 100
In [3]: df
Out[3]:
foo bar
0 100 4
1 2 5
2 3 6
The column foo was updated where the column bar is greater than 5.
This violated the CoW principles though, because it would have to modify the
view df["foo"] and df in one step. Hence, chained assignment will
consistently never work and raise a ChainedAssignmentError warning
with CoW enabled:
In [33]: df = pd.DataFrame({"foo": [1, 2, 3], "bar": [4, 5, 6]})
In [34]: df["foo"][df["bar"] > 5] = 100
With copy on write this can be done by using loc.
In [35]: df.loc[df["bar"] > 5, "foo"] = 100
Read-only NumPy arrays#
Accessing the underlying NumPy array of a DataFrame will return a read-only array if the array shares data with the initial DataFrame:
The array is a copy if the initial DataFrame consists of more than one array:
In [36]: df = pd.DataFrame({"a": [1, 2], "b": [1.5, 2.5]})
In [37]: df.to_numpy()
Out[37]:
array([[1. , 1.5],
[2. , 2.5]])
The array shares data with the DataFrame if the DataFrame consists of only one NumPy array:
In [38]: df = pd.DataFrame({"a": [1, 2], "b": [3, 4]})
In [39]: df.to_numpy()
Out[39]:
array([[1, 3],
[2, 4]])
This array is read-only, which means that it can’t be modified inplace:
In [40]: arr = df.to_numpy()
In [41]: arr[0, 0] = 100
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[41], line 1
----> 1 arr[0, 0] = 100
ValueError: assignment destination is read-only
The same holds true for a Series, since a Series always consists of a single array.
There are two potential solutions to this:
Trigger a copy manually if you want to avoid updating DataFrames that share memory with your array.
Make the array writeable. This is a more performant solution but circumvents Copy-on-Write rules, so it should be used with caution.
In [42]: arr = df.to_numpy()
In [43]: arr.flags.writeable = True
In [44]: arr[0, 0] = 100
In [45]: arr
Out[45]:
array([[100, 3],
[ 2, 4]])
Patterns to avoid#
No defensive copy will be performed if two objects share the same data while you are modifying one object inplace.
In [46]: df = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6]})
In [47]: df2 = df.reset_index(drop=True)
In [48]: df2.iloc[0, 0] = 100
This creates two objects that share data and thus the setitem operation will trigger a
copy. This is not necessary if the initial object df isn’t needed anymore.
Simply reassigning to the same variable will invalidate the reference that is
held by the object.
In [49]: df = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6]})
In [50]: df = df.reset_index(drop=True)
In [51]: df.iloc[0, 0] = 100
No copy is necessary in this example. Creating multiple references keeps unnecessary references alive and thus will hurt performance with Copy-on-Write.
Copy-on-Write optimizations#
A new lazy copy mechanism that defers the copy until the object in question is modified
and only if this object shares data with another object. This mechanism was added to
methods that don’t require a copy of the underlying data. Popular examples are DataFrame.drop() for axis=1
and DataFrame.rename().
These methods return views when Copy-on-Write is enabled, which provides a significant performance improvement compared to the regular execution.