Polars Cheat Sheet: How to get fast, scalable DataFrame work by writing expressions

# Why Polars feels faster

# The core mental model: expressions + lazy execution

The simplest way to see the model is the difference between read_csv and scan_csv. read_csv reads a file into memory immediately. scan_csv reads only the header and returns a lazy frame. Every chain of transformations after scan_csv is a declarative description of intent. Nothing executes until you call collect. That delay gives the optimizer room to push filters down to the file and read only necessary columns.

When a dataset is larger than available memory, use collect(engine="streaming"). Instead of materializing the full frame, streaming collect processes the file in chunks so you can work with data that doesn't fit in memory.

# Use expressions, not imperative loops

Polars shines when you express transformations as expressions (select, filter, with_columns). Those expressions are combined and optimized before any data moves. Common verbs you'll use:

• select, filter, with_columns for column-level transformations
• group_by and agg for grouped summaries
• pivot and unpivot for reshaping
• when / then / otherwise for conditional logic

Because these are expressions, Polars can reorder and fuse operations to minimize intermediate data and I/O.

# Window-style computations with over

The over construct runs an aggregation per group but returns a value for every row. That means you can compute each region's share of its group total, or a rank within a category, using a single expression—no manual groupby-and-join required. It reads like a normal column expression and simplifies per-row group calculations.

# Important type behavior: null vs NaN

Polars uses two different states: null for missing values and NaN as a real floating-point value. They have different methods for handling and are not interchangeable. Expect confusion if you assume they behave the same.

# Joins, semi/anti joins, and namespaces

Polars supports the usual joins and also semi and anti joins, which filter rows based on another frame without widening your result. Type-specific namespaces—.str for string ops and.dt for datetime ops—expose focused operations for those types.

# Output paths and interoperability

# Practical takeaway

# Where to go next

Download the KDnuggets Polars cheat sheet for concise reference examples of scan_csv vs read_csv, over usage, the core verb set, join variants,.str/.dt operations, and output methods like sink_parquet and to_pandas.