| # FOSDEM PGDay 2027 — submission |
|
|
| Deadline: **2026-12-01 23:59:59 CET** (2 December 01:59 Moscow). Notification by 2026-12-20. |
| Talk: 45 minutes + questions, English only. No abstract length limit in the form (minimum 30 characters). |
|
|
| ## Title |
|
|
| When a rewrite looks faster but breaks the result: executable counterexamples for LLM-proposed SQL rewrites |
|
|
| ## Abstract |
|
|
| Large language models are increasingly used to propose SQL rewrites for performance. A rewrite can be syntactically valid, run faster, and still change the result: duplicate rows collapse, NULL semantics shift, ordering or LIMIT is lost. Measuring execution time alone cannot catch this, and formal equivalence checkers only cover a subset of SQL. |
|
|
| This talk presents an approach that checks rewrites against the states a schema actually allows, and records the evidence rather than a score. For each candidate rewrite the approach executes the reference and the candidate on a bank of database states generated under the real constraints, and when the results differ it stores the refuting state and the concrete rows that prove the difference. |
|
|
| Two findings from the pilot dataset. First, a model specialised in code was not safer than a general-purpose one: it rewrote more often and its rewrites were refuted more often. Second, three different models independently produced the same wrong rewrite, replacing `UNION ALL` with a disjunction `OR`: row multiplicities disappeared, and none of the randomly generated states caught it — only the states built for multiplicity, NULL handling and ties did. |
|
|
| The talk argues that practitioners adopting LLM-assisted rewriting need a validation step before trusting a speedup, and shows a reproducible protocol that runs against a real database. |
|
|
| ## What the audience gets |
|
|
| - A concrete failure mode that survives review because it is invisible in timing measurements. |
| - A reproducible procedure: state profiles, execution, and the rows that prove the difference. |
| - An open dataset and scripts where every published number can be recomputed. |
|
|
| ## About the speaker |
|
|
| Artem Yakovlev, student, Polytechnic College of Yaroslav-the-Wise Novgorod State University. |
| ORCID: 0009-0006-8617-7895. Works on SQL query optimisation and on validation of |
| machine-generated query rewrites; author of the AI Database Optimizer toolkit and of the |
| SREB dataset. |
|
|
| ## Submission notes (form field) |
|
|
| - **Track:** databases / query optimisation / tooling around query rewriting. |
| - **Travel and visa support:** a visa invitation letter would be needed for travel from Russia. The invitation policy mentions a valid paid registration while selected speakers receive a free ticket; please advise how to proceed in that case, and confirm when registration opens — the visa process has to start well before January. |
| - **Dataset:** the talk is based on a public dataset with a DOI (assigned on release) and on open scripts; the versioned release link will be included once published. A live demonstration of one counterexample takes about two minutes and needs only a laptop. |
| - **Recording:** please confirm whether the talk is recorded and published. |
| - **Materials:** slides can be provided before the event if recordings are made. |
|
|
| ## Internal checklist (not part of the form) |
|
|
| - [ ] registration opened and completed |
| - [ ] visa invitation requested |
| - [ ] dataset DOI resolves to the released version, not to master |
| - [ ] abstract character count checked against the live form |
| - [ ] bio and photo added to the speaker profile |
|
|
