Analyzing data using Polars in Python: an introduction

Short step-by-step guide using Polars for data analysis in Python.

James Birkenau


November 22, 2023


I’ve been working with data for a while, and I thought I had seen it all until I encountered Polars. This versatile library has become a cornerstone in my data analysis, handling large datasets with surprising ease. It’s not just fast; it’s mind-blowingly efficient compared to what I was used to with other tools. I’m continually impressed by how it simplifies the complex, and I’ve found myself reaching for it more and more. Let me share some insights on why Polars is reshaping my approach to data analysis and could very well change yours too.

Introduction to Polars and Its Significance in Data Analysis

Graphic showing python icons and data science icons

In the landscape of data analysis where Python reigns supreme, Polars emerges as a promising newcomer destined to handle large datasets with speed and efficiency. It’s a data processing library written in Rust, designed to provide lightning-fast performance for data manipulation. Unlike Pandas, which is almost synonymous with Python data analysis, Polars leverages Rust’s memory safety and speed, providing an API that feels familiar to Pandas users but promises better performance, especially when dealing with big data.

When I first stumbled upon Polars, I was skeptical. Could it really outshine the tried-and-tested Pandas? The answer isn’t simple, but it’s exciting - for large datasets, Polars is a game-changer.

import polars as pl

# Reading a CSV file
df = pl.read_csv('large_dataset.csv')

This snippet is your entry ticket into the world of Polars. With just a few lines, I was able to read in a large dataset that would typically have Pandas gasping for memory. But with Polars, it’s all smooth sailing.

The crux of Polars’ excellence lies in its two types of DataFrame representations: DataFrame and LazyFrame. It’s not just a matter of eager versus lazy evaluation; it’s a paradigm shift that optimizes your data analysis tasks efficiently behind the scenes.

To cast a bit of light on LazyFrame: imagine a situation where you’re only interested in the final result and not the intermediate steps that lead to it. LazyFrame embodies this principle – by deferring computations until necessary, it optimizes the overall workflow.

# Lazy evaluation in Polars
lazy_df = df.lazy().filter(pl.col("some_column") > 10)
result = lazy_df.collect()

By invoking lazy(), we switch to a lazy context where no computation is done until collect() is called. This avoids unnecessary computations and hence speeds up the process.

But Polars isn’t just about speed. Its significance lies in making data analysis more accessible and less resource-intensive. In my experience, I’ve found that its syntax is intuitive for those making the transition from Pandas.

# Simple columnar operation in Polars
df = df.with_column((pl.col("column_a") * 2).alias("doubled_column_a"))

In this example, column manipulation is a breeze. With Polars, I can perform operations on DataFrame columns using methods that feel very much in line with Pandas.

Moreover, Polars opens up advanced data analysis techniques by supporting features such as window functions, groupbys, and joins, which are critical when dealing with complex datasets. It can handle these with a degree of performance that often leaves traditional Python-based libraries in the dust.

# Using window functions in Polars
df = df.with_column(

What sets Polars apart is how it has integrated sophisticated data processing capabilities with an API that’s so user-friendly. Starting from basic data operations to advanced analytics, it provides a broad spectrum of options that cater to both novice and seasoned data analysts.

Importantly, getting the most out of Polars doesn’t require deep knowledge of Rust or systems programming. As a Pythonista, I found the transition seamless, adopting an incredibly powerful tool without stepping out of the comfort zone provided by Python.

In the living ecosystem of data analysis libraries, Polars has carved its niche, demonstrating that you can have speed, efficiency, and ease of use in a single package. While this introduction scratches the surface, it’s a strong foundation that paves the way for diving into more nuanced and advanced Polars functionalities in the sections that follow.

For those who crave a deeper dive into the technicalities and the development of Polars, the Github repository (Polars Git Repository) provides ample resources, from detailed documentation to active issues and feature requests. For a more academic understanding or research on the implications of using Rust for data science, the rapidly growing collection of articles and papers serves as an excellent starting point.

Getting Started with Polars in Python

Screenshot of a terminal with polars installation commands being executed

Embarking on the journey with Polars in Python, I quickly realized how intuitive and fast it is to manipulate and analysis large datasets. Let’s dive straight into setting up Polars and executing some basic operations that are the bread and butter of any data analysis workflow.

First things first, we need to get Polars installed. Simply run:

pip install polars

Now, let’s start with importing Polars and creating our first DataFrame. Unlike Pandas, where I’d use pd.DataFrame(), Polars has its own syntax that’s just as straightforward.

import polars as pl

# Define a simple DataFrame
df = pl.DataFrame({
    "id":[1, 2, 3, 4],
    "names":["Alice", "Bob", "Charlie", "Dave"],
    "scores":[9.5, 7.0, 8.3, 5.4]

This code piece crafted a DataFrame - a fundamental object in Polars, akin to Excel spreadsheets or SQL tables. Next is to read and write data, without which, honestly, analysis is moot. Polars shines here with its ability to handle CSVs, Parquet, and other file formats efficiently.

# Writing DataFrame to CSV

# Reading from CSV
df = pl.read_csv("my_data.csv")

I regard column addition as if it’s child’s play with Polars; just select the column, perform an operation or append a series. Witness the simplicity:

# Adding a new column
df = df.with_column((pl.col("scores") * 10).alias("score_x10"))

And, selecting data in Polars? That’s a cakewalk, neatly performed with a fluent API. Here’s a filter operation typically representing a WHERE clause in SQL:

# Filter rows where scores are above 8
high_scores = df.filter(pl.col("scores") > 8)

Sorting, on my word, is ridiculously easy. Just decide the column and order, and it’s done.

# Sort DataFrame by scores
sorted_df = df.sort("scores")

During aggregations, which are ubiquitously essential, a ‘groupby’ comes into picture. I regularly sum up score data grouped by some categorical variable.

# Group by "names" and sum "scores"
grouped_df = df.groupby("names").agg(

But wait, imagine I was too hasty and had a typo in my column name. I’d ordinarily revisit the code, but with lazy evaluation, Polars lets me stack up transformations and delay execution, saving time when the datasets are colossal.

# Fixing column name using lazy evaluation
lazy_df = (df.lazy()
           .rename({"scores": "points"})

Granted these operations sound elementary, but grasping them kicks off the data analysis journey. Each function here, from read_csv to groupby, is pivotal in routine data operations. And paired with Polars’ speed, they make an analyst’s day remarkably more efficient.

Coming from using Pandas, the shift to Polars felt like swapping a hatchback for a sports car—both get the job done, but the latter does it with an exhilarating vroom. There’s a different kind of thrill in performing data wrangling at lightning-fast speeds.

The motley of simple operations showcased here marks the onset of a data analysis adventure with Polars. Wielding these tools, I could slice, dice, aggregate, and evaluate data with unparalleled efficiency, prepping me up for more intricate maneuvers down the road.

DataFrames and LazyFrames in Polars

Visualization comparing dataframe and lazyframe structures in polars

Polars is a fast data manipulation library written in Rust with a Python interface that’s designed for speed and efficiency. When I first discovered Polars, it challenged my understanding of how I could interact with large datasets, especially through its two core abstractions: DataFrame and LazyFrame.

I’m often dealing with huge datasets, and standard DataFrame operations consume a significant amount of memory and compute resources. With Polars, I can work fluidly even when my datasets are massive, thanks to LazyFrame.

import polars as pl

# Create a DataFrame
df = pl.DataFrame({
    "fruits": ["apple", "banana", "pear", "pineapple"],
    "baskets": [15, 32, 10, 5]

A DataFrame is immediately familiar if you’ve used Pandas. It’s an eager structure that holds data in-memory. When I conduct transformations or computations, they are performed instantly. For small to moderately sized datasets, this immediate feedback is quite useful.

However, the magic unfolds with LazyFrame. Once datasets start ballooning in size, the in-memory operations of DataFrames can become unwieldy. This is where Polars takes a page out of Dask’s playbook and incorporates lazy evaluation. Using LazyFrames not only made my coding more efficient but also dramatically reduced memory overhead.

# Lazy evaluation with LazyFrame
lf = df.lazy()

# Define computation graph
lf = lf.with_column((pl.col("baskets") * 2).alias("doubled_baskets"))

# Trigger computation
df_result = lf.collect()

In the example above, I create a LazyFrame using df.lazy(). This doesn’t start any computations yet. It simply creates a computation graph. I can then define various operations, like doubling the number of baskets for each fruit, and give it an alias. The computation isn’t performed until I call the collect() method, which is when the results are eagerly evaluated and returned as a DataFrame.

What particularly excites me about LazyFrames is the query optimization. Just like SQL query planners, Polars optimizes the computations under the hood, often figuring out a faster way to apply the sequence of operations before executing them. This optimization is largely opaque; I don’t have to worry about it, but I reap the performance benefits.

# Filter and sort with LazyFrame
lf = (
    .filter(pl.col("baskets") > 10)

# Execute the computation graph
df_filtered_sorted = lf.collect()

In this example, I filter rows and sort them without triggering actual computation until I call collect(). This feature is a game-changer when chaining multiple transformations. It’s like building a recipe before turning on the stove.

For anyone just starting with Polars, remember that LazyFrames are not just about deferring computation. They’re also about efficiency. I avoid unnecessary memory allocation and harness optimizations I would normally have to craft by hand.

As you delve into your data analysis endeavors with Polars, embrace both DataFrames and LazyFrames, leveraging each where it suits. Use DataFrames for quick exploration and immediate results, but switch to LazyFrames for complex workflows and larger datasets, ensuring your resource utilization is as efficient as the library’s design.

Here’s the link to Polars’ GitHub repository, where you can dive deeper into its capabilities: Polars GitHub

Remember, smart data structure selection is pivotal, and Polars hands you a powerful toolkit to juggle performance and usability, revolutionizing the way you work with data in Python.

Basic Data Manipulations with Polars

Table depicting basic data manipulation functions like sorting and filtering in action

Manipulating data is like kneading dough; the better you do it, the better your bread – or in this case, your analysis. Here’s how I roll with Polars, a lightning-fast DataFrame library in Python, to tackle basic data manipulations.

First off, let’s read data into a DataFrame. With Polars, it’s as simple as:

import polars as pl
df = pl.read_csv("path_to_your_file.csv")

Now, let’s say we want to peek at the first few rows to confirm everything looks as expected:


Next on the agenda is column selection. I often only need a subset of the data, so here’s how to pick what you need:

subset =['column_1', 'column_2'])

Need to filter rows? Here’s a crisp example of pulling out entries where a certain condition is met:

filtered_data = df.filter(pl.col('some_column') > 10)

Often, I’ll need to create or manipulate columns. Polars shines here with its expressive syntax. To add a new column based on existing ones:

df = df.with_column((pl.col('column_1') / pl.col('column_2')).alias('new_column'))

Grouping and aggregating data is a staple in data analysis. In Polars, aggregating based on groups is not just powerful, but also intuitive:

grouped_df = df.groupby('group_column').agg([

Dealing with missing data is a reality we all face. Let’s fill missing values with zeros:

df = df.fill_null(0)

Sorting data can unveil patterns – ascending or descending, you name it, and Polars does it:

sorted_df = df.sort(['column_to_sort_by'], reverse=True)

Joining tables is a common scenario, and Polars’ speedy join operations come in handy. Here’s how I’d perform a left join:

left_join_df = df.join(

Summing up, these basic manipulations form the bedrock of data analysis. By getting familiar with them in Polars, you’re setting the foundation for more intricate analyses and operations down the road.

And remember, the best way to become adept is to get your hands dirty with real data, so go ahead and play around with these snippets. You’ll find yourself slicing and dicing data like a pro before you know it!

Advanced Data Analysis Techniques with Polars

Flowchart demonstrating several advanced data analysis operations available in polars

Once you’ve got the basics of Polars under your belt, it’s time to dive into some more advanced techniques. The leap from basic to advanced data analysis is like moving from arithmetic to calculus – the fundamental logic remains, but the operations become powerful enough to unlock entirely new insights.

Let’s discuss how to perform grouped operations. Imagine you’re analyzing a dataset containing sales data, and you want to aggregate sales by product category. In Polars, you don’t have to break a sweat:

import polars as pl

# Assume `df` is your DataFrame loaded earlier and 'category' and 'sales' are columns in your DataFrame
grouped_df = df.groupby("category").agg(

This concise syntax using chaining commands is heaven-sent – quick to write, and it reads like a story: “group by category, then aggregate sales by sum and mean.”

Next up, joins. You have two sets of data and you’re looking to merge them to cross-reference information.

# Other DataFrame to join with, let's say it contains product details
df_products = pl.DataFrame({...})

# Performing a left join on 'product_id' column
df_joined = df.join(df_products, on="product_id", how="left")

Look at how seamless that is, just laying down the ‘how’ and the ‘on’ parameters, and you’re all set.

Moving forward, window functions can compute cumulative statistics, moving averages, or ranks within a specific window of the dataset. You can calculate a rolling average for instance:

# Calculate a 7-day rolling average for sales

Your code not only instructed Polars to calculate the average, but also effortlessly set the window size.

Another power move with Polars is utilizing expressions. An expression allows you to build a computation that you apply to a DataFrame later. They’re reusable and compose well:

sales_above_average = pl.col("sales") > df.get_column("sales").mean()

# Applying the expression to filter data
df_filtered = df.filter(sales_above_average)

By creating sales_above_average as an expression, you can apply it wherever necessary without rewriting the logic.

Lastly, don’t befuddle if you need conditional logic. Polars has a function similar to SQL’s CASE WHEN, which is when().then().otherwise(). Incredibly useful for creating new columns based on conditions:

    pl.when(pl.col("sales") > 10000)

Can you appreciate how this mirrors natural language? You’ve told Polars: If sales are above 10,000, label it High, otherwise Low, and name this new finding ‘sales_category’.

With advanced operations in Polars, you’re not just shuffling data around—you’re crafting a narrative with it. The story you’re telling is driven by the interplay of powerful, almost intuitive commands. And remember, there’s a vibrant community and troves of resources out there. You might want to check out the Polars documentation for more in-depth examples and the Polars GitHub repository for the latest updates.

Polars shines when you stretch it to its limits, and there’s a certain thrill to writing code that’s both expressive and efficient. It’s a gateway to viewing your data from angles and depths previously unimaginable and it’s all at your fingertips. Happy analyzing!

Performance Benchmarks: Polars vs Pandas vs Dask

Bar chart showing the speed comparison of polars pandas and dask for common operations

In the realm of data analysis with Python, three libraries emerge as primary contenders when it comes to handling large datasets: Pandas, Dask, and Polars. Each has its unique strengths, and performance benchmarks can reveal insightful differences that are crucial when we take scalability and efficiency into account.

Pandas, the go-to library for data manipulation, has been the cornerstone of many data analysts’ toolkits. However, it’s well-known that Pandas can struggle with large datasets, often consuming significant amounts of memory and CPU time for operations.

As I ventured into larger datasets, I stumbled upon Dask, which extends Pandas to larger-than-memory computations by parallelizing operations and managing memory across clusters. To my excitement, this meant I could handle larger data volumes without upgrading my hardware.

Nevertheless, the real game-changer was discovering Polars, a library optimized for performance with a focus on speed and memory efficiency. It uses Apache Arrow as its memory model, enabling fast data access and computation. When I first ran Polars on a dataset that would typically choke my Pandas workflow, the speedup was astonishing.

Let’s do a quick comparison starting with creating a DataFrame of random data in Pandas:

import pandas as pd
import numpy as np

# Generating random data with Pandas
df_pandas = pd.DataFrame({
    'a': np.random.rand(1000000),
    'b': np.random.rand(1000000)

Attempting the same with Polars, you’ll notice the similar syntax:

import polars as pl

# Generating random data with Polars
df_polars = pl.DataFrame({
    'a': np.random.rand(1000000),
    'b': np.random.rand(1000000)

With Dask, the code changes slightly due to its lazy evaluation, creating a parallelized DataFrame:

import dask.dataframe as dd

# Generating random data with Dask
ddf = dd.from_pandas(df_pandas, npartitions=4)

Now for a simple operation, like calculating the mean of a column:

# Pandas
%timeit df_pandas['a'].mean()

# Polars
%timeit df_polars['a'].mean()

# Dask (Note: Dask needs to compute since it is lazy)
%timeit ddf['a'].mean().compute()

In my experience, Polars often delivers results significantly faster than Pandas and even outperforms Dask in certain scenarios, especially when dealing with large datasets on single-machine setups.

To delve deeper, say you want to merge two large DataFrames. Here’s where Polars shines with its join operations:

# Polars
%timeit df_larger = df_polars.join(df_polars, on='a')

# Pandas
%timeit df_larger_pandas = df_pandas.merge(df_pandas, on='a')

# Dask
%timeit df_larger_dask = ddf.merge(ddf, on='a').compute()

While Dask utilizes multiple cores and splits computation, the overhead of managing parallelism can make it less efficient for certain tasks compared to the ultra-optimized Polars. Plus, Polars is designed to minimize memory footprint, a critical factor when datasets grow.

One should note that benchmarks and performance can vary based on the operations, the size of the dataset, and the hardware used. Thus, while my anecdotes point towards the efficiency of Polars, it is essential to run your benchmarks to suit your specific use cases.

The choice between Pandas, Dask, and Polars depends on your data’s size, the complexity of operations, and your system’s capabilities. Switching between these libraries isn’t arduous, thanks to their somewhat similar API designs. For a budding data analyst or data scientist exploring Polars could be a game-changer by enabling faster and more efficient data processing.

Polars is under active development, with its repository at Polars GitHub teeming with updates and community contributions. There, you can dive into more extensive benchmarks, discussions, and examples which further affirm Polars as a promising tool for modern data analysis.

Case Study: Real-world Data Analysis Using Polars

Infographic presenting a step-by-step analysis of a dataset using polars functions

In the world of data analysis, we often hear about success stories of gaining insights and making impactful decisions. Let’s dive into a practical scenario where I used Polars, a fast DataFrame library for Python, to analyze real-world data.

Imagine we’ve got a dataset, sales_data.csv, which holds a year’s worth of sales data for an international retail company. Our goal is to extract meaningful patterns—such as top-selling products, trends over months, and performance by region.

First, we need to load the data. Polars makes it seamless with its read_csv function. I start by importing the library and reading the dataset into a DataFrame:

import polars as pl

# Load the sales data into a DataFrame
df = pl.read_csv('sales_data.csv')

With the data loaded, I notice some cleaning is necessary. Polars comes in handy for data preprocessing with its intuitive and concise syntax. Let’s say we need to parse the date column and remove any rows with missing values:

# Convert the 'date' column to datetime and drop missing values
df = df.with_column(
    pl.col('date').str.strptime(pl.Datetime, '%Y-%m-%d')

Analysis requires aggregation, and Polars performs it efficiently. To find the top-selling products, I use the groupby and agg functions:

# Group by 'product' and sum the 'sales' column
top_products = df.groupby('product').agg(
).sort('total_sales', reverse=True).limit(5)

Now it’s time for a more complex analysis: trend discovery over months. We often look for seasonality in sales, and Polars handles it gracefully. Let’s extract the month from the date and sum sales per month:

# Extract month from 'date' and calculate monthly sales
monthly_trends = df.with_column(

Looking at regional performance involves a combination of filtering, grouping, and sorting data. Polars’ expressive API makes such tasks straightforward. Below, I filter for a specific region, say ‘Europe’, group by ‘country’, and then sort by sales:

# Analysis for European countries
europe_sales = df.filter(
    pl.col('region') == 'Europe'
).sort('total_sales', reverse=True)

Visualizing the results can be equally effortless with Polars by making use of other libraries like matplotlib. However, sticking solely to Polars, one can still appreciate the results in a tabular format straight from the console.

While this real-world analysis scratches the surface of Polars’ capabilities, it underscores the library’s speed and ease of use. Not only does it outperform traditional tools in terms of performance, but its syntax is also delightfully Pythonic, making data analysis tasks less cumbersome and more enjoyable.

As we’ve seen, from simple cleaning to advanced aggregations, Polars offers a suite of powerful methods to manage and analyze data efficiently. Whether you’re wrangling small datasets or delving into massive data pools, Polars is an excellent choice, particularly for those who crave performance without sacrificing readability or ease of use. Although we didn’t delve into performance benchmarks in this section (that’s covered elsewhere in this article), my firsthand experience attests to Polars’ superior speed, especially when compared to Pandas.

By incorporating Polars into your data analysis toolkit, you too can streamline your workflows and uncover insights that steer your data narratives towards informed conclusions.