Deep Dive into pandas Copy-on-Write Mode—Part II | by Patrick Hoefler | Aug, 2023

Explaining how Copy-on-Write optimizes performance

Photo by Joshua Brown on Unsplash


The first post explained how the Copy-on-Write mechanism works. It highlights some areas where copies are introduced into the workflow. This post will focus on optimizations that ensure that this won’t slow the average workflow down.

We utilize a technique that pandas internals use to avoid copying the whole DataFrame when it’s not necessary and thus, increase performance.

I am part of the pandas core team and was heavily involved in implementing and improving CoW so far. I am an open source engineer for Coiled where I work on Dask, including improving the pandas integration and ensuring that Dask is compliant with CoW.

Removal of defensive copies

Let’s start with the most impactful improvement. Many pandas methods performed defensive copies to avoid side effects to protect against inplace modifications later on.

df = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6]})
df2 = df.reset_index()
df2.iloc[0, 0] = 100

There is no need to copy the data in reset_index, but returning a view would introduce side effects when modifying the result, e.g. df would be updated as well. Hence, a defensive copy is performed in reset_index.

All these defensive copies are no longer there when Copy-on-Write is enabled. This affects many methods. A full list can be found here.

Additionally, selecting a columnar subset of a DataFrame will now always return a view instead of a copy as before.

Let’s look at what this means performance-wise when we combine some of these methods:

import pandas as pd
import numpy as np

N = 2_000_000
int_df = pd.DataFrame(
np.random.randint(1, 100, (N, 10)),
columns=[f"col_{i}" for i in range(10)],
float_df = pd.DataFrame(
np.random.random((N, 10)),
columns=[f"col_{i}" for i in range(10, 20)],
str_df = pd.DataFrame(
columns=[f"col_{i}" for i in range(20, 30)],

df = pd.concat([int_df, float_df, str_df], axis=1)

This creates a DataFrame with 30 columns, 3 different dtypes and 2 million rows. Let’s execute the following method chain on this DataFrame:

df.rename(columns={"col_1": "new_index"})
.assign(sum_val=df["col_1"] + df["col_2"])
.drop(columns=["col_10", "col_20"])
.astype({"col_5": "int32"})

All of these methods perform a defensive copy without CoW enabled.

Performance without CoW:

2.45 s ± 293 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Performance with CoW enabled:

13.7 ms ± 286 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

An improvement by roughly a factor of 200. I chose this example explicitly to illustrate the potential benefits of CoW. Not every method will get that much faster.

Optimizing copies triggered by inplace modifications

The previous section illustrated many methods where a defensive copy is no longer necessary. CoW guarantees that you can’t modify two objects at once. This means that we have to introduce a copy when the same data is referenced by two DataFrames. Let’s look at techniques to make these copies as efficient as possible.

The previous post showed that the following might trigger a copy:

df.iloc[0, 0] = 100

The copy is triggered if the data that is backing df is referenced by another DataFrame. We assume that our DataFrame has n integer columns, e.g. is backed by a single Block.

Image by author

Our Reference tracking object is also referencing another Block, so we can not modify the DataFrame inplace without modifying another object. A naive approach would be to copy the whole block and be done with it.

Image by author

This would set up a new reference tracking object and create a new Block that is backed by a fresh NumPy array. This Block doesn’t have any more references, so another operation would be able to modify it inplace again. This approach copies n-1 columns that we don’t necessarily have to copy. We utilize a technique we call Block splitting to avoid this.

Image by author

Internally, only the first column is copied. All other columns are taken as views on the previous array. The new Block does not share any references with other columns. The old Block still shares references with other objects since it is only a view on the previous values.

There is one disadvantage to this technique. The initial array has n columns. We created a view on columns 2 till n, but this keeps the whole array alive. We also added a new array with one column for the first column. This will keep a bit more memory alive than necessary.

This system directly translates to DataFrames with different dtypes. All Blocks that are not modified at all are returned as is and only Blocks that are modified inplace are split.

Image by author

We now set a new value into column n+1 the float Block to create a view on columns n+2 to m. The new Block will only back column n+1.

df.iloc[0, n+1] = 100.5
Image by author

Methods that can operate inplace

The indexing operations we looked at don’t generally create a new object; they modify the existing object inplace, including the data of said object. Another group of pandas methods does not touch the data of the DataFrame at all. One prominent example is rename. Rename only changes the labels. These methods can utilize the lazy-copy mechanism mentioned above.

There is another third group of methods that can actually be done inplace, like replace or fillna. These will always trigger a copy.

df2 = df.replace(...)

Modifying the data inplace without triggering a copy would modify df and df2, which violates CoW rules. This is one of the reasons why we consider keeping the inplace keyword for these methods.

df.replace(..., inplace=True)

This would get rid of this problem. It’s still an open proposal and might go into a different direction. That said, this only pertains to columns that are actually changed; all other columns are returned as views anyway. This means that only one column is copied if your value is only found in one column.


We investigate how CoW changes pandas internal behavior and how this will translate to improvements in your code. Many methods will get faster with CoW, while we will see a slowdown in a couple of indexing related operations. Previously, these operations always operated inplace, which might have produced side effects. These side effects are gone with CoW and a modification on one DataFrame object will never impact another.

The next post in this series will explain how you can update your code to be compliant with CoW. Also, we will explain which patterns to avoid in the future.

Thank you for reading. Feel free to reach out to share your thoughts and feedback about Copy-on-Write.

Source link

Leave a Comment