DuckLake vs Iceberg is usually argued on design. Iceberg keeps table metadata as files next to the data. DuckLake keeps it as rows in a SQL database. Both camps have a diagram that proves they are right.
I wanted a number instead. So I ran the same workload on both: 50 small commits, a delete, a column rename, time travel, then cleanup. I counted every file each one left on disk and timed every commit. Below is what each format cost, and which one I would pick for a team of two to ten people.
What is DuckLake?
DuckLake is an open table format from the people behind DuckDB. The data is Parquet, the same as Iceberg. The difference is where the table's bookkeeping lives.
In Iceberg, every commit writes new metadata files: a metadata.json, a manifest list and one or more manifests. A catalog then points at the newest metadata.json. Iceberg catalogs explained covers that part.
In DuckLake, the bookkeeping is a set of tables in a database you already know how to run: Postgres, SQLite, a DuckDB file, or any ACID SQL database with primary keys, per the FAQ. A commit is one transaction against that database. There are no metadata files at all.
On status: DuckLake 1.0 shipped on 13 April 2026 with DuckDB 1.5.2. The announcement calls it "a production-ready release with guaranteed backward-compatibility". In 1.0, data inlining is on by default, and the release added sorted tables, bucket partitioning that uses the same murmur3 hash as Iceberg, GEOMETRY and VARIANT types, and experimental Iceberg v3 deletion vectors. Clients exist for DataFusion, Spark, Trino and pandas. The release calendar plans 2.0 for 21 October 2026, and says the date is tentative.
The test
Same machine, local disk, same table: id, ts, kind, amount. Fifty appends of 1,000 rows each, then the same follow-up steps on both.
- Iceberg: PyIceberg 0.12.0 with its SQL catalog on SQLite. These are the numbers from my Iceberg vs Parquet test.
- DuckLake: DuckDB 1.5.6 with the
ducklakeextension, run once with a SQLite catalog and once with Postgres 16 in a throwaway container.
INSTALL ducklake; LOAD ducklake; ATTACH 'ducklake:sqlite:catalog.sqlite' AS lake (DATA_PATH 'data/'); USE lake; CREATE TABLE events (id BIGINT, ts TIMESTAMP, kind VARCHAR, amount DOUBLE); -- then 50 x: INSERT INTO events SELECT ... FROM range(start, start + 1000) One caveat up front. The writers differ: PyIceberg is Python, DuckDB is C++. So commit times compare two real tools, not two formats in a vacuum. The file counts are a property of the formats.
This is the first part. The full post — including the rest of the working details — is on my site: DuckLake vs Iceberg: which table format for a small team

