保留正文与补采附件。源页面当前未重新核验；具体限制见交付清单。

Revisit consent button We value your privacy We use cookies to enhance your browsing experience, serve personalised ads or content, and analyse our traffic. By clicking "Accept All", you consent to our use of cookies. Customise Reject All Accept All Customise Consent Preferences We use cookies to help you navigate efficiently and perform certain functions. You will find detailed information about all cookies under each consent category below. The cookies that are categorised as "Necessary" are stored on your browser as they are essential for enabling the basic functionalities of the site. ... Show more NecessaryAlways Active Necessary cookies are required to enable the basic features of this site, such as providing secure log-in or adjusting your consent preferences. These cookies do not store any personally identifiable data.

Functional Functional cookies help perform certain functionalities like sharing the content of the website on social media platforms, collecting feedback, and other third-party features.

Analytics Analytical cookies are used to understand how visitors interact with the website. These cookies help provide information on metrics such as the number of visitors, bounce rate, traffic source, etc.

Performance Performance cookies are used to understand and analyse the key performance indexes of the website which helps in delivering a better user experience for the visitors.

Advertisement Advertisement cookies are used to provide visitors with customised advertisements based on the pages you visited previously and to analyse the effectiveness of the ad campaigns.

Uncategorised Other uncategorised cookies are those that are being analysed and have not been classified into a category as yet. No cookies to display. Reject All Save My Preferences Accept All Toggle SearchToggle Mobile Navigation Publish AI, ML & data-science insights to a global community of data professionals. Sign inSubmit an Article LatestEditor's PicksNewsletterShipAI

Write For TDS LinkedInXToggle Search Data Science

Modern DataFrames in Python: A Hands-On Tutorial with Polars and DuckDB

How I learned to handle growing datasets without slowing down my entire workflow Benjamin Nweke November 21, 20256 min read Photo by Rene Böhmer on Unsplash If you have worked with Python for data, you have probably experienced the frustration of waiting minutes for a Pandas operation to finish. At first, everything seems fine, but as your dataset grows and your workflows become more complex, your laptop suddenly feels like it's preparing for lift-off. A couple of months ago, I worked on a project analyzing e-commerce transactions with over 3 million rows of data. It was a pretty interesting experience, but most of the time, I watched simple groupby operations that normally ran in seconds suddenly stretch into minutes. At that point, I realized Pandas is amazing, but it is not always enough. This article explores modern alternatives to Pandas, including Polars and DuckDB, and examines how they can simplify and improve the handling of large datasets. For clarity, let me be upfront about a few things before we begin. This article is not a deep dive into Rust memory management or a proclamation that Pandas is obsolete. Instead, it is a practical, hands-on guide. You will see real examples, personal experiences, and actionable insights into workflows that can save you time and sanity. ··· Learn this step by step with the interactive AI and Data Scientist roadmap.

Why Pandas Can Feel Slow

Back when I was on the e-commerce project, I remember working with CSV files over two gigabytes, and every filter or aggregation in Pandas often took several minutes to complete. During that time, I would stare at the screen, wishing I could just grab a coffee or binge a few episodes of a show while the code ran. The main pain points I encountered were speed, memory, and workflow complexity. We all know how large CSV files consume enormous amounts of RAM, sometimes more than what my laptop could comfortably handle. On top of that, chaining multiple transformations also made code harder to maintain and slower to execute. Polars and DuckDB address these challenges in different ways. Polars, built in Rust, uses multi-threaded execution to process large datasets efficiently. DuckDB, on the other hand, is designed for analytics and executes SQL queries without needing you to load everything into memory. Basically, each of them has its own superpower. Polars is the speedster, and DuckDB is kind of like the memory magician. And the best part? Both integrate seamlessly with Python, allowing you to enhance your workflows without a complete rewrite.

Setting Up Your Environment

Before we start coding, make sure your environment is ready. For consistency, I used Pandas 2.2.0, Polars 0.20.0, and DuckDB 1.9.0. Pinning versions can save you headaches when following tutorials or sharing code. bash

pip install pandas==2.2.0 polars==0.20.0 duckdb==1.9.0

In Python, import the libraries: python

import pandas as pd
import polars as pl
import duckdb
import warnings
warnings.filterwarnings("ignore")

For example, I will use an e-commerce sales dataset with columns such as order ID, product ID, region, country, revenue, and date. You can download similar datasets from Kaggle or generate synthetic data.

Loading Data

Loading data efficiently sets the tone for the rest of your workflow. I remember a project where the CSV file had nearly 5 million rows. Pandas handled it, but the load times were long, and the repeated reloads during testing were painful. It was one of those moments where you wish your laptop had a “fast forward” button. Switching to Polars and DuckDB completely improved everything, and suddenly, I could access and manipulate the data almost instantly, which honestly made the testing and iteration processes far more enjoyable. With Pandas: python

df_pd = pd.read_csv("sales.csv")
print(df_pd.head(3))

With Polars: python

df_pl = pl.read_csv("sales.csv")
print(df_pl.head(3))

With DuckDB: python

con = duckdb.connect()
df_duck = con.execute("SELECT * FROM 'sales.csv'").df()
print(df_duck.head(3))

DuckDB can query CSVs directly without loading the entire datasets into memory, making it much easier to work with large files.

Filtering Data

The problem here is that filtering in Pandas can be slow when dealing with millions of rows. I once needed to analyze European transactions in a massive sales dataset. Pandas took minutes, which slowed down my analysis. With Pandas: python

filtered_pd = df_pd[df_pd.region == "Europe"]

Polars is faster and can process multiple filters efficiently: python

filtered_pl = df_pl.filter(pl.col("region") == "Europe")

DuckDB uses SQL syntax: python

filtered_duck = con.execute("""
    SELECT *
    FROM 'sales.csv'
    WHERE region = 'Europe'
""").df()

Now you can filter through large datasets in seconds instead of minutes, leaving you more time to focus on the insights that really matter.

Aggregating Large Datasets Quickly

Aggregation is often where Pandas starts to feel slow. Imagine calculating total revenue per country for a marketing report. In Pandas: python

agg_pd = df_pd.groupby("country")["revenue"].sum().reset_index()

In Polars: python

agg_pl = df_pl.groupby("country").agg(pl.col("revenue").sum())

In DuckDB: python

agg_duck = con.execute("""
    SELECT country, SUM(revenue) AS total_revenue
    FROM 'sales.csv'
    GROUP BY country
""").df()

I remember running this aggregation on a 10 million-row dataset. In Pandas, it took nearly half an hour. Polars completed the same operation in under a minute. The sense of relief was almost like finishing a marathon and realizing your legs still work.

Joining Datasets at Scale

Joining datasets is one of those things that sounds simple until you are actually knee-deep in the data. In real projects, your data usually lives in multiple sources, so you have to combine them using shared columns like customer IDs. I learned this the hard way while working on a project that required combining millions of customer orders with an equally large demographic dataset. Each file was big enough on its own, but merging them felt like trying to force two puzzle pieces together while your laptop begged for mercy. Pandas took so long that I began timing the joins the same way people time how long it takes their microwave popcorn to finish. Spoiler: the popcorn won every time. Polars and DuckDB gave me a way out. With Pandas: python

merged_pd = df_pd.merge(pop_df_pd, on="country", how="left")

Polars: python

merged_pl = df_pl.join(pop_df_pl, on="country", how="left")

DuckDB: python

merged_duck = con.execute("""
    SELECT *
    FROM 'sales.csv' s
    LEFT JOIN 'pop.csv' p
    USING (country)
""").df()

Joins on large datasets that used to freeze your workflow now run smoothly and efficiently.

Lazy Evaluation in Polars

One thing I didn’t appreciate early in my data science journey was how much time gets wasted while running transformations line by line. Polars approaches this differently. It uses a technique called lazy evaluation, which essentially waits until you have completed defining your transformations before executing any operations. It examines the entire pipeline, determines the most efficient path, and executes everything simultaneously. It’s like having a friend who listens to your entire order before walking to the kitchen, instead of one who takes each instruction separately and keeps going back and forth. This TDS article indepthly explains lazy evaluation. Here’s what the flow looks like: Pandas: python

df = df[df["amount"] > 100]
df = df.groupby("segment").agg({"amount": "mean"})
df = df.sort_values("amount")

Polars Lazy Mode: python

import polars as pl

df_lazy = (
    pl.scan_csv("sales.csv")
      .filter(pl.col("amount") > 100)
      .groupby("segment")
      .agg(pl.col("amount").mean())
      .sort("amount")
)

result = df_lazy.collect()

The first time I used lazy mode, it felt strange not seeing instant results. But once I ran the final .collect(), the speed difference was obvious. Lazy evaluation won’t magically solve every performance issue, but it brings a level of efficiency that Pandas wasn’t designed for. ···

Conclusion and takeaways

Working with large datasets doesn’t have to feel like wrestling with your tools. Using Polars and DuckDB showed me that the problem wasn’t always the data. Sometimes, it was the tool I was using to handle it. If there is one thing you take away from this tutorial, let it be this: you don’t have to abandon Pandas, but you can reach for something better when your datasets start pushing their limits. Polars gives you speed as well as smarter execution, then DuckDB lets you query huge files like they’re tiny. Together, they make working with large data feel more manageable and less tiring. If you want to go deeper into the ideas explored in this tutorial, the official documentation of Polars and DuckDB are good places to start.

Register Now to Continue Reading

Unlock Towards Data Science with a Free Account

All articles on Towards Data Science are available to registered users. Create a free account today to access this article and thousands more. Register for freeSign in to your account

Written By

Benjamin Nweke

Benjamin Nweke is a Computer Science graduate and technical writer exploring AI, machine learning, and data science. He writes about building reliable AI systems, model evaluation, agentic architectures, and the practical challenges that appear when machine learning moves beyond experiments. His work focuses on making technically complex ideas clear, useful, and grounded in real implementation. See all from Benjamin Nweke Big DataData AnalysisData SciencePolarsPython Share This Article

Towards Data Science is a community publication. Submit your insights to reach our global audience and earn through the TDS Author Payment Program. Write for TDS

Related Articles

To Really Learn a New Topic, Take Your Time

Machine Learning Our weekly selection of must-read Editors' Picks and original features TDS Editors June 22, 20232 min read

Linear Regression in Python: Sklearn vs Excel

Machine Learning Is Microsoft excel a good alternative for a quick and approximate linear regression prediction business case? I think yes, but let's do a... Kaushik Choudhury August 6, 20206 min read

Data for Good - Somalia Drought Management

Data Science Applied data science and analytics to tackle global warming induced environmental challenges. End-to-end project with real world rainfall... Mohamed A. Warsame March 25, 202114 min read

5 Quick and Easy Tips to Become an Efficient Data Scientist in Both Python and R

Programming Having difficulty transitioning between using Python and R? Try these tips to make the transition smoother Madison Hunter November 8, 20229 min read

Pandas tips I wish I knew before

Programming How does pivot work? What is the main pandas building block? And more ... Roman Orac April 1, 20204 min read

The Streamlit Colour Picker: An Easy Way to Change Chart Colours on Your Streamlit Dashboard

Data Science Provide Your Dashboard Users with Greater Flexibility Andy McDonald December 20, 20225 min read

From Business Student to Data Scientist in Tech

Data Science A timeline of how I went from being scared of maths to becoming a full-fledged Data Scientist at a renowned Tech firm Khouloud El Alami July 3, 202310 min read

Comparing Documents With Similarity Metrics

Data Science An initial look into how math is leveraged to compare text in NLP Aashish Nair January 17, 20224 min read

Data Science Career Switch: Is Master's Degree Worth It?

Data Science Part 1: Mode of Learning (Online vs Physical) Nathan Timothy Handoko January 11, 20226 min read

Your home for data science and AI. The world's leading publication for data science, data analytics, data engineering, machine learning, and artificial intelligence professionals. Subscribe to Our NewsletterWrite For TDS•About•Advertise•Privacy Policy•Terms of Use © Insight Media Group, LLC 2026•The New Stack•roadmap.sh•Creator Network

Recommended Articles

Close