A small suite of microbenchmark queries targeting Vortex-specific scan paths, run on every PR commit (unlike the label-gated suites) with a high iteration count, so it is the most sensitive — and most frequently seen — benchmark comment.

init.sql generates a 25M-row two-column table; the numbered queries each pin down one scan behavior:

• 0_sum-with-filter.sql: a filtered aggregation that forces a linear scan today; once statistics are propagated to arrays it should use zone maps instead of decoding every row.
• 1_sum.sql: an unfiltered aggregation that should eventually be answered from footer statistics without decoding data.

CI runs this as the Vortex queries PR comment (see .github/workflows/sql-vortex-pr.yml), comparing DataFusion and DuckDB over Parquet and Vortex files with 100 iterations.

vx-bench run vortex --engine datafusion,duckdb --format parquet,vortex