Polars 2.0 arrives with major upgrades, including out-of-core spill-to-disk support, performance boosts, first-class SQL, a new Map data type, and faster, stricter type handling feedback.

The engine defaults now favor out-of-core and streaming: collecting a LazyFrame uses the streaming engine by default, with a 64GB disk budget and spill-to-disk kicking in around 80% RAM, and plans to extend out-of-core to joins and group-bys.

Looking ahead, the team is continuing work on out-of-core performance, large-scale CPU scaling, Polars Cloud, and GeoPolars, and invites users to report issues on GitHub.

Benchmarks show Polars outperforming or matching competitors in many tests, though scalability notes at very high core counts reveal overheads under investigation for the next release.

A stricter behavior model emphasizes fast upfront error reporting and opt-in implicit data-mismatch handling, with improved schema validation via collect_schema to catch issues without full data materialization.

SQL is now a first-class citizen, with major optimizer and engine enhancements such as join reordering, better subplan elimination, and dynamic predicates using bloom filters; in benchmarks Polars often leads on TPC-H and TPC-DS workloads across configurations against DuckDB and DataFusion.

Migration resources are provided, including links to the Polars 2.0 benchmark repo and an upgrade guide to help users transition smoothly.

A new Map datatype is supported, with Arrow MapType becoming Polars Map, enabling dictionary-like operations—key lookups, iteration, and contains_key checks directly on map columns.