benchingVortex

Benchmarks comparing query engines against the same dataset stored in three formats: CSV, Parquet, and Vortex.

The query is simple but realistic: count transaction failures grouped by day across a multi-file dataset (Q3/Q4 2025, one CSV per day).

---

Data

Path	Format	Generated from
data/	CSV (one file per day)	source
data_parquet/	Parquet (batched 10 CSVs per file)	convert_to_parquet.py
data_vortex/	Vortex	convert_to_vortex.py

Conversion pipeline: CSV → Parquet (via DuckDB) → Vortex (via vortex.io.write).

---

Benchmark scripts

CSV

Script	Engine
failures_by_day.py	DuckDB
failures_by_day_polars.py	Polars
failures_by_day_datafusion.py	DataFusion

Parquet (parquet_scripts/)

Script	Engine
failures_by_day_duckdb.py	DuckDB
failures_by_day_polars.py	Polars
failures_by_day_datafusion.py	DataFusion

Vortex (vortex_scripts/)

Script	Engine
failures_by_day_duckdb.py	DuckDB (via PyArrow bridge)
failures_by_day_polars.py	Polars (via PyArrow dataset)
failures_by_day_pyarrow.py	PyArrow + DuckDB
failures_by_day_vortex.py	Native Vortex scan + push-down filter

Note: DataFusion is not benchmarked against Vortex. DataFusion has no native Vortex reader, and bridging via PyArrow (the same approach used for DuckDB and Polars) has not been implemented yet.

---

Running

# setup
uv sync

# run a single benchmark
uv run python failures_by_day.py

# run all benchmarks and produce a grouped bar chart
uv run python run_benchmarks.py

---

Dependencies

• DuckDB
• Polars
• DataFusion
• Vortex
• PyArrow
• matplotlib (for run_benchmarks.py)