When the sea lion learns to quack: DuckDB as a MariaDB storage engine
summary
An embedded DuckDB storage engine for MariaDB is implemented with a DuckDB manager singleton plus per-connection THD MVCC and delta appender buffer that batches writes into temporary DuckDB tables and merges on commit; full-query pushdown handles most SQL with lightweight rewrites. Benchmarks: TPC‑H SF1000 shows the engine close to DuckDB CLI (CLI wins some queries like Q29), ClickBench shows engine matches ClickHouse on hot cache but CLI dominates cold reads due to a 7GB on‑disk difference. Code is merged into LTS branches and packages ship end of July.
AI-written summary · about
chapters (9)
- 0:00Sea lion meets DuckDB
- 1:05OLAP goals in MariaDB
- 2:31DuckDB as storage engine
- 4:40SQL syntax differences
- 6:26TPC-H performance
- 7:30ClickBench vs ClickHouse
- 10:01Multi-tenant concurrency
- 12:47Plugin for DuckDB quacking
- 15:16Acknowledgements
related talks
40mDuckDB Meets Data Lakes
47mWhy DuckDB could be the engine for your Data Lake?
27mState of the Duck
38mFrom Big Data to DuckDB: Ryan Boyd (MotherDuck) on Hyper-Tenancy, Duck Lake, and Making Data Fun.
21mWhat can Postgres learn from DuckDB?
27mThe Duck and the DataFrame: A Data Engineer’s Journey with DuckDB
24mpg_duckdb: Ducking awesome analytics in Postgres
05mAnnouncing DuckDB Support for Delta Lake and a DuckDB Extension to Unity Catalog