SBBD 2026 · lead author · best-short-papers session
Rampart
Reproducible benchmarking of data processing paradigms: DuckDB, Polars and Dask processing the same data and models, with automated anti-leakage verification.
- ETL integrity validated by requiring bitwise-identical predictions across engines: Δ must equal exactly 0.0
- Hashed data snapshots separate upstream data revisions from code changes
- Pinned Docker digests and explicit core budgets: every published latency is conditional on a recorded environment
- Demonstrated on Brazilian municipal school-dropout prediction (INEP + World Bank, 2007–2024)
duckdb · polars · dask