dbt without the warehouse (or the bill): DuckDB end to end

At a dbt Summit 2026 breakout session titled "dbt without the warehouse (or the bill): DuckDB end to end," Hannes Mühleisen, a co-founder of DuckDB, walked through why dbt v2 now ships with a built-in DuckDB adapter and demonstrated a fully offline dbt build against a real dataset.

Hannes had planned to make the case for pairing dbt with DuckDB, but said that argument was already settled once he learned dbt v2 ships with DuckDB support out of the box. He credited Josh, the original maintainer of the dbt DuckDB adapter built for dbt v1, with making the combination possible in the first place.

Connecting dbt to DuckDB requires only a small change to the profile: setting the output type to duckdb and pointing it at a file path, with no authentication or credit card required. Skipping the file path altogether creates a disposable in-memory database useful for testing.

DuckDB started eight years ago as what Mühleisen described as a "SQLite for analytics," and he positioned it as filling the gap between spreadsheet tools like Excel and distributed systems like Spark, since most real-world data does not actually need distributed compute.

Citing an analysis by Fivetran's George Fraser, Mühleisen said the median data warehouse query scans about 100 megabytes and the 99.9th percentile scans about 300 gigabytes, both sizes that fit comfortably on a single laptop.

Mühleisen highlighted several SQL conveniences built into DuckDB, including reading directly from CSV, Parquet, and Iceberg files without an import step, a GROUP BY ALL syntax that has since been adopted by Snowflake, Databricks, and Postgres and is headed into the SQL standard, and built-in UNION BY NAME, PIVOT, and summarize commands.

Mühleisen previewed DuckDB 2.0, expected around October 2026 and code-named Cinnamon Teal, which will bring a rewritten SQL parser, a variant data type for semi-structured data, and general availability for a client-server protocol called Quark.

In a live demo, Mühleisen disconnected his laptop from the internet entirely and ran a dbt build against roughly 10 gigabytes of compressed Wikipedia page-view CSV files stored locally, using window functions to rank articles by their change in traffic over time.

The build, including a window-function computation across millions of Wikipedia article titles and a group-by-all aggregation over roughly 300 million rows, finished in about three seconds, entirely offline and without any warehouse compute cost.

Mühleisen closed by framing DuckDB's mission as giving people confidence to work with any dataset on a single machine, arguing that the gap between what a laptop can handle and what most real-world datasets actually require has nearly closed.

Watch every keynote and breakout from dbt Summit 2026 on demand. Watch dbt Summit 2026 on demand →