Polars

Polars is a DataFrame library that leverages Rust's performance and Apache Arrow's columnar format. It's significantly faster than pandas for most operations, especially on large datasets, thanks to parallel execution and lazy evaluation.

Installation

# Python
pip install polars

# With all optional dependencies (Excel, SQL, cloud storage)
pip install 'polars[all]'

# Node.js
npm install nodejs-polars

Basic Operations

# basics.py: Create and manipulate DataFrames
import polars as pl

# Create from dict
df = pl.DataFrame({
    "name": ["Alice", "Bob", "Charlie", "Diana"],
    "age": [30, 28, 35, 42],
    "city": ["NYC", "London", "Paris", "NYC"],
    "salary": [85000, 72000, 90000, 110000],
})

# Basic operations
print(df.head(2))
print(df.describe())
print(df.shape)  # (4, 4)
print(df.columns)  # ['name', 'age', 'city', 'salary']

# Select columns
df.select("name", "salary")
df.select(pl.col("name"), pl.col("salary") / 1000)

# Filter rows
df.filter(pl.col("age") > 30)
df.filter((pl.col("city") == "NYC") & (pl.col("salary") > 80000))

# Sort
df.sort("salary", descending=True)

Expressions

# expressions.py: Polars expression system — the core of Polars
import polars as pl

df = pl.DataFrame({
    "product": ["A", "B", "A", "B", "A"],
    "revenue": [100, 200, 150, 300, 120],
    "cost": [60, 120, 80, 180, 70],
    "date": ["2026-01-01", "2026-01-01", "2026-01-02", "2026-01-02", "2026-01-03"],
})

# Computed columns with expressions
result = df.with_columns(
    profit=pl.col("revenue") - pl.col("cost"),
    margin=(pl.col("revenue") - pl.col("cost")) / pl.col("revenue") * 100,
    date_parsed=pl.col("date").str.to_date(),
)

# Multiple aggregations
summary = df.group_by("product").agg(
    total_revenue=pl.col("revenue").sum(),
    avg_revenue=pl.col("revenue").mean(),
    max_cost=pl.col("cost").max(),
    count=pl.len(),
)

# Window functions
df.with_columns(
    revenue_rank=pl.col("revenue").rank(descending=True).over("product"),
    cumulative=pl.col("revenue").cum_sum().over("product"),
    pct_of_total=pl.col("revenue") / pl.col("revenue").sum().over("product") * 100,
)

Lazy Evaluation

# lazy.py: Use lazy frames for optimized query plans
import polars as pl

# Lazy evaluation — build a query plan, execute once
result = (
    pl.scan_csv("sales_data.csv")  # Lazy read
    .filter(pl.col("status") == "completed")
    .with_columns(
        revenue=pl.col("quantity") * pl.col("unit_price"),
        order_date=pl.col("date").str.to_date(),
    )
    .filter(pl.col("order_date").dt.year() == 2026)
    .group_by("category")
    .agg(
        total_revenue=pl.col("revenue").sum(),
        order_count=pl.len(),
        avg_order=pl.col("revenue").mean(),
    )
    .sort("total_revenue", descending=True)
    .collect()  # Execute the optimized plan
)

# View the query plan before executing
lazy_df = pl.scan_csv("sales_data.csv").filter(pl.col("amount") > 100)
print(lazy_df.explain())  # Shows optimized plan with predicate pushdown

Joins

# joins.py: Join DataFrames efficiently
import polars as pl

orders = pl.DataFrame({
    "order_id": [1, 2, 3],
    "user_id": [10, 20, 10],
    "amount": [99.99, 249.50, 15.00],
})

users = pl.DataFrame({
    "user_id": [10, 20, 30],
    "name": ["Alice", "Bob", "Charlie"],
})

# Inner join
joined = orders.join(users, on="user_id", how="inner")

# Left join
all_orders = orders.join(users, on="user_id", how="left")

# Join with different column names
orders.join(users, left_on="user_id", right_on="user_id", how="inner")

I/O Operations

# io.py: Read and write various formats
import polars as pl

# CSV
df = pl.read_csv("data.csv")
df.write_csv("output.csv")

# Parquet (recommended for large datasets)
df = pl.read_parquet("data.parquet")
df.write_parquet("output.parquet", compression="zstd")

# JSON
df = pl.read_json("data.json")
df.write_json("output.json")

# From pandas
import pandas as pd
pandas_df = pd.read_sql("SELECT * FROM users", engine)
polars_df = pl.from_pandas(pandas_df)

# SQL databases
df = pl.read_database("SELECT * FROM orders WHERE amount > 100", connection)

# Scan (lazy) for large files — only reads what's needed
lazy = pl.scan_parquet("huge_dataset.parquet")
result = lazy.filter(pl.col("status") == "active").head(1000).collect()

Comparison with Pandas

# comparison.py: Common pandas patterns translated to Polars
import polars as pl

# pandas: df['new_col'] = df['a'] + df['b']
# Polars:
df = df.with_columns(new_col=pl.col("a") + pl.col("b"))

# pandas: df.groupby('cat').agg({'val': ['sum', 'mean']})
# Polars:
df.group_by("cat").agg(
    val_sum=pl.col("val").sum(),
    val_mean=pl.col("val").mean(),
)

# pandas: df.apply(lambda row: ..., axis=1)  # SLOW
# Polars: Use expressions instead (vectorized, parallel)
df.with_columns(
    label=pl.when(pl.col("score") > 90).then(pl.lit("A"))
         .when(pl.col("score") > 80).then(pl.lit("B"))
         .otherwise(pl.lit("C"))
)