DuckDB

Hannes Mühleisen and Mark Raasveldt started DuckDB at CWI in Amsterdam as a research answer to a simple question: why does analytics always need a server? The first public release landed in 2019 under the MIT License, and the pitch was SQLite's model applied to OLAP. No daemon, no cluster, just a C++ library that loads into your process and runs columnar, vectorized queries against Parquet, CSV, and in-memory tables.

The project's early reputation rested on benchmarks that made analysts look twice. Scans over wide tables ran fast enough that people started calling it SQLite's analytical cousin, and the comparison stuck because it is accurate: SQLite handles transactions, DuckDB handles the billion-row join. Client libraries arrived for Python, R, Java, and more, and a WebAssembly build put the whole engine in a browser tab through the official web shell.

Adoption climbed steadily through the early 2020s, with monthly downloads passing six million by 2023. March 2025 brought a local UI with interactive notebooks for writing SQL and inspecting results. Then came the events that reshaped the project's trajectory: an announced 2.0 release for fall 2026 that adds client/server mode, and Amazon's agreement in August 2026 to acquire DuckLabs, the commercial team behind the engine, while DuckDB itself stays open source. [edit]

You use DuckDB through SQL, either from a client library inside Python or R, from the command line, or in the browser shell. The engine reads Parquet, CSV, and JSON directly without a load step, which is the feature that converts most newcomers: point a query at a folder of files and it just works. Columnar storage and vectorized execution keep scans fast, and the SQL dialect covers window functions, list and struct types, and table functions that most embedded engines skip.

Typical sessions involve joining a few large tables, aggregating, and writing results back out to Parquet. Analysts run it inside Jupyter notebooks, data engineers wire it into ETL jobs against S3, and the March 2025 UI added notebooks of its own for people who prefer a graphical surface. The v2.0 client/server mode, due fall 2026, extends the same engine to multiple concurrent users over a network connection. [edit]