We use DuckDB to provide Python/Pandas access to data in our feature store. It helped us beat Databricks (45X), Vertex (13X), and Sagemaker (10X) in reading into Pandas from the Feature Store.
The recent support in DuckDB for "AS OF" has really helped our performance when doing point-in-time joins across tables.
https://www.hopsworks.ai/post/python-centric-feature-service...