Two SQL queries can look equivalent and return different answers. Query Witness finds a small, valid DuckDB database that shows the difference and saves it for replay.
Preparing an example…
A concrete example
An unassigned support ticket still exists. COUNT(*) counts it; COUNT(assignee_id) skips it because the assignee is NULL. That single row shows why the queries aren’t equivalent.
My contribution
I directed AI implementation agents and worked through examples to question the results. Separate reviews caught correctness problems we then fixed, including a timeout in the final comparison that could be reported as a completed result.
Kept deliberately small
The tool supports one table with up to two integer columns and a restricted SQL subset. It finds differences you can inspect. If the search runs out of time or attempts, that doesn’t prove the queries equivalent.
Python
DuckDB
SQLGlotReleases
All on GitHubReplay requires format_version to be the integer 1 (true and 1.0 are rejected). The witness size cap is 1 MiB of UTF-8 bytes. Schema types INT4 and INT32 are rejected; only INTEGER and INT remain. DuckDB Query interrupted is treated as cancellation, not a finished search.
Query Witness finds small, valid DuckDB databases where two queries disagree, reduces the data, and exports a replayable witness.