Data lakes have become widely popular in managing enterprise data. Their widespread integration with query engines has allowed them to displace specialized data warehouses as the single source of truth for enterprise data. While the columnar storage format and block min-max indices allow query engines to achieve competitive performance on relational data analytics queries, they are not yet suitable for other search-oriented queries like full text and vector nearest neighbor search. We present Rottnest, a general system that builds additional lightweight indices on top of data lakes. We show that our system is more cost efficient compared to un-indexed data lakes or specialized databases across several orders of magnitude of total query loads and operating time horizons.