Polars

Polars is what pandas would be if rewritten from scratch with performance as the primary goal. A Rust-based DataFrame engine that's 5-30x faster than pandas, uses a fraction of the memory, and parallelizes automatically across all CPU cores. Lazy evaluation lets a query optimizer rewrite your code before execution.

At 37k stars, Polars has moved from "interesting alternative" to "default choice for new data projects." The Arrow-backed columnar storage means zero-copy interop with other Arrow tools. Compared to pandas (ubiquitous but single-threaded), Polars wins on every performance benchmark. Compared to DuckDB (SQL-focused analytical engine), Polars is more DataFrame-native. Compared to Spark (distributed compute), Polars handles single-machine workloads without cluster overhead.

Use this for any new Python data project, especially with large datasets, ETL pipelines, or memory-constrained environments. Skip this if your team knows pandas cold and your datasets are small — the API differences aren't worth relearning for 10MB CSVs.

The catch: pandas compatibility layer exists but isn't perfect. Some pandas-dependent libraries (scikit-learn, seaborn) expect pandas DataFrames. And the API, while Pythonic, has its own idioms that take time to internalize. MIT license.