Lance
An open lakehouse file, table, and catalog format for multimodal AI data, built for fast random access and hybrid search on top of object storage.
Lance stores images, video, audio, text, and embeddings in one format so a lakehouse can sit directly on object storage instead of a separate database. It combines vector similarity search, full-text search, and SQL analytics over the same dataset with accelerated secondary indices, and its random access is reported to run much faster than Parquet or Iceberg without giving up scan performance.
Other features include adding columns with backfilled values without rewriting a table, automatic versioning with ACID transactions, time travel, tags and branches, and no extra infrastructure to run them. It integrates with Apache Arrow, Pandas, Polars, DuckDB, Spark, Ray, Trino, Flink, and open catalogs such as Apache Polaris and Unity Catalog.
Lance is written in Rust and ships a Python package (pylance).