Pandas has long been an indispensable tool for any developer thanks to a simple and understandable API, as well as a rich set of tools for cleaning, researching and analyzing data. And everything would be fine, but when it comes to data that does not fit into RAM or require complex calculations, pandas performance is not enough.

In this article, I will not describe qualitatively different approaches to data analysis, such as Spark or DataFlow. Instead, I will describe six interesting tools and demonstrate the results of their use:

Chapter 1:

  • Numba
  • Multiprocessing
  • Pandarallel

Chapter 2:

  • Swifter
  • Modin
  • Dask


This tool directly accelerates Python itself. Numba is a JIT compiler that likes loops, mathematical operations and Numpy, which is a Pandas core lib. Let's check in practice what advantages it gives.

We will simulate a typical situation - you need to add a new column by applying some function to the existing one using the apply method.

import pandas as pd
import numpy as np
import numba

# create a table of 100,000 rows and 4 columns filled with random numbers from 0 to 100
df = pd.DataFrame(np.random.randint(0,100,size=(100000, 4)),columns=['a', 'b', 'c', 'd'])

# function for creating new col
def multiply(x):
    return x * 5
# optimized version of this function
def multiply_numba(x):
    return x * 5

As you can see, you do not need to change anything in your code. Just add a decorator. Now look at the results:

# our function
In [1]: %timeit df['new_col'] = df['a'].apply(multiply)
23.9 ms ± 1.93 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

# pandas implementation
In [2]: %timeit df['new_col'] = df['a'] * 5
545 µs ± 21.4 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

# out function, optimized by numba
# we use vector of values so that numba itself performs the loop optimization
In [3]: %timeit df['new_col'] = multiply_numba(df['a'].to_numpy())
329 µs ± 2.37 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

The optimized version is ~ 70 times faster! However, in absolute terms, Pandas implementation is slightly behind, so let's take a more complex case. Lets define new functions:

# square the values and take their mean
def square_mean(row):
    row = np.power(row, 2)
    return np.mean(row)
# usage:
# df['new_col'] = df.apply(square_mean, axis=1)

# numba does not know how to work with pandas primitives (Dataframe, Series, etc.)
# so that we use a two-dimensional numpy array
def square_mean_numba(arr):
    res = np.empty(arr.shape[0])
    arr = np.power(arr, 2)
    for i in range(arr.shape[0]):
        res[i] = np.mean(arr[i])
    return res
# usage:
# df['new_col'] = square_mean_numba(df.to_numpy())

Let's build a graph showing the dependence of the calculation time on the number of rows in the data frame:



  • It is possible to achieve more than x1000 acceleration
  • Sometimes you need to rewrite the code to use Numba
  • Cannot be used everywhere, optimizing mathematical operations is the main case
  • Keep in mind that Numba does not support all features of python и numpy


The first thing that comes to mind when it comes to processing a large dataset is to parallelize all the calculations. This time we will not use any third-party libraries - only python tools.

We will use text processing as an example. Below I took dataset with news headlines. Like last time, we will try to speed up the apply method:

df = pd.read_csv('abcnews-date-text.csv', header=0)
# increase the dataset 10 times, adding copies to the end
df = pd.concat([df] * 10)
publish_date headline_text
0 20030219 aba decides against community broadcasting lic...
1 20030219 act fire witnesses must be aware of defamation
2 20030219 a g calls for infrastructure protection summit
3 20030219 air nz staff in aust strike for pay rise
4 20030219 air nz strike to affect australian travellers
# calculate the average word length in the title
def mean_word_len(line):
    # this cycle just complicates the task
    for i in range(6):
        words = [len(i) for i in line.split()]
        res = sum(words) / len(words)
    return res

def compute_avg_word(df):
    return df['headline_text'].apply(mean_word_len)

Concurrency will be provided by the following code:

from multiprocessing import Pool

# I have 4 cores
n_cores = 4
pool = Pool(n_cores)

def apply_parallel(df, func):
    # split dataframe
    df_split = np.array_split(df, n_cores)
    # calculate metrics for each and concatenate
    df = pd.concat(, df_split))
    return df
# df['new_col'] = apply_parallel(df, compute_avg_word)

Compare speed:



  • Works on the standard python library
  • We got x2-3 speedup
  • Using parallelization on small data is a bad idea, because the overhead of interprocess communication exceeds the time gain


Pandarallel is a small pandas library that adds the ability to work with multiple cores. Under the hood, it works on standard multiprocessing, so you should not expect an increase in speed compared to the previous approach, but everything is out of the box + some sugar in the form of a beautiful progress bar ;)


Let's start testing. Further I will use the same data and functions to process them as in the previous part. Set up pandarallel first - it’s very simple:

from pandarallel import pandarallel
# pandarallel will determine how many cores you have, but you can specify it yourself

Now It remains only to write an optimized version of our handler, which is also very simple - just replace apply withparallel_apply:


Compare speed:



  • A rather large overhead in about 0.5 seconds immediately catches your eye. Each time it is used, pandarallel first creates a pool of workers and then closes it. In the self-written version above, I created the pool 1 time, and then reused it, so the overhead was much lower
  • If you do not take into account the costs described above, then the acceleration is the same as in the previous version - about 2-3 times
  • Pandarallel also knows how to parallel_apply over grouped data (groupby), which is quite convenient. For a complete list of functionality and examples see here

In general, I would prefer this option to self-written, because for medium / large data volumes there is almost no difference in speed, and we get an extremely simple API and progress bar.

To be continued

In this part, we looked at 2 fairly simple approaches to pandas optimization - using jit compilation and parallelizing a task using several cores. In the next part I will talk about more interesting and complex tools, but for now, I advise you to test the tools yourself and make sure of their efficiency.

fast pandas

P.s Trust, but verify - all code used in the article (benchmarks and charts drawing), I posted on github