The H2O.ai db-benchmark has been updated with new results, and the AWS EC2 instance used for benchmarking has been changed to ensure fairness across libraries. DuckDB is the fastest library for both join and group by queries at almost every data size. The benchmark was re-run on a c6id.metal machine to avoid issues with network storage and noisy neighbors. The previous machine had network storage and could suffer from variable performance, while the c6id.metal machine has local storage and avoids noisy neighbor problems. Additionally, the c6id.metal machine stresses parallel performance with its 128 cores. DuckDB performed exceptionally well, especially in query 5 of the advanced group by queries. It is also one of the few solutions to finish the 50GB join query. Some solutions experienced timeouts on the 50GB datasets.
https://duckdb.org/2023/11/03/db-benchmark-update.html