Two recent posts on the Percona blog caught my attention. In Replicating from InnoDB into a DuckDB storage engine and The DuckDB MySQL engine at 500 GB, Evgeniy Patlan wired DuckDB into mysqld as a storage engine, pointed row-based replication at it, and measured it at TPC-H scale factor 500. The numbers are striking: loads 25 times faster than InnoDB, one fifth of the disk, and the full 22-query TPC-H suite done in 186 seconds where InnoDB needed about 28 hours and never finished four of the queries.
The posts are also honest about the cost. The work is experimental. The worst bug found was silent data loss on every replicated transaction: the two-phase commit code for mixed-engine transactions dropped prepared DuckDB transactions …
[Read more]