DuckLake

Docs · Specification · GitHub

DuckLake is a lakehouse format: Parquet data files on cheap storage (local disk, S3, GCS, ...) plus a transactional SQL database (the "catalog") holding all metadata — table/schema definitions, snapshots, statistics, partition info. Because the catalog is a real database (not just files), multiple DuckDB processes can concurrently read and write the same lakehouse — this is what the docs call "multiplayer DuckDB". On top of that it gets time travel, schema evolution, partitioning, and transactions for free.

Requires DuckDB v1.5.2+. Not supported: indexes, primary keys, foreign keys, or constraint enforcement — those live in the catalog database's own domain, not in DuckLake tables.

Install & attach