DuckDB vs Polars vs pandas on 30 million rows: 6.6 s vs 22.9 s vs 178 s

How fast is DuckDB compared with Polars and pandas on a 1.3 GB CSV?

A group-by on 30 million rows (1.3 GB CSV) took 6.6 s and 187 MB with DuckDB, 22.9 s and 522 MB with Polars, and 178 s and 568 MB with pandas on a MacBook Air.

Results

DuckDB	6.6 s, 187 MB RAM
Polars	22.9 s, 522 MB RAM
pandas	178 s, 568 MB RAM

How it was measured

SUM and COUNT grouped by country and month, reading the CSV directly.

SELECT country, date_trunc('month', created_at) m, count(*), sum(amount) FROM read_csv('orders.csv') GROUP BY 1,2

Why

DuckDB is a vectorized, parallel, columnar engine that streams the CSV.

When this does not apply

Synthetic data and one query; CSV date parsing dominates the Polars and pandas numbers.

Environment

MacBook Air (Apple silicon, 8 cores), macOS. Measured 2026-10-12.

Source: shorts/demo-ticket/cost in my content-engine repository. Treat any single run as a direction; re-run it on your hardware.

Parquet vs CSV: 3x smaller and 3.6x faster in DuckDB

By Josué Barros, senior software engineer (full-stack, data and AI).