It’s not clear to me why this would be faster than polars, duckdb, vaex or clickhouse. They seem to be taking the same approach of multithreading, optimizing the plan, using arrow, optimizing the core functions like group by.