feat(benchmarks): Add velox_tpch_data_gen utility for SF=N Parquet generation

[kgpai]

Summary

Adds velox_tpch_data_gen, a standalone CLI that generates TPC-H Parquet files at a configurable scale factor using Velox's own TpchConnector + Parquet writer. Output layout is Hive-style, matching what velox_tpch_benchmark --data_path=<dir> already expects, so the same directory is consumable by both the benchmark and other engines (DuckDB, Spark, etc.) for cross-engine comparison.

The implementation lifts the pattern from ParquetTpchTest::saveTpchTablesAsParquet but streams scan -> write through a multi-split pipeline so SF=10 (~10 GB) fits in memory. The original test pattern collects rows into a single in-memory RowVector, which only works at SF=0.01.

Why Velox-driven generation (vs. dbgen / DuckDB)

When this data feeds a Velox-vs-DuckDB perf comparison, having one engine write the files would bias the read path — DuckDB's writer might emit Parquet tuned for DuckDB's reader (page sizes, encodings, dictionary thresholds calibrated to its own decoder). Using Velox's writer keeps the read path neutral: both engines read files written by the engine under test, and we additionally exercise Velox's writer in the process.

Flags

Flag	Default	Notes
--output_dir	required	Created if missing. One subdirectory per table.
--scale_factor	1.0	Must be > 0. SF=1 → ~1 GB, SF=10 → ~10 GB.
--num_splits_per_table	8	Scan parallelism per table.
--num_drivers	4	Driver concurrency for the scan-then-write pipeline.

Usage

velox_tpch_data_gen \
  --output_dir=/data/tpch10 \
  --scale_factor=10 \
  --num_splits_per_table=8 \
  --num_drivers=8

Produces:

/data/tpch10/customer/*.parquet
/data/tpch10/lineitem/*.parquet
/data/tpch10/nation/*.parquet
/data/tpch10/orders/*.parquet
/data/tpch10/part/*.parquet
/data/tpch10/partsupp/*.parquet
/data/tpch10/region/*.parquet
/data/tpch10/supplier/*.parquet

Then:

velox_tpch_benchmark --data_path=/data/tpch10 --data_format=parquet

Test plan

• Builds clean on Linux with -DVELOX_ENABLE_BENCHMARKS=ON
• SF=1 generation produces files with row counts matching TPC-H spec (lineitem ~6M, orders ~1.5M, etc.)
• velox_tpch_benchmark --data_path=<dir> successfully reads the generated SF=1 files and runs all 22 queries
• DuckDB CLI can read the generated Parquet (interop check) — SELECT count(*) FROM read_parquet('/data/tpch1/lineitem/*.parquet')
• SF=10 generation succeeds within ~10 minutes on a 30GB-RAM / 8-core machine