reshard_parquet.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "pyarrow>=15.0",
# ]
# ///
"""Re-shard existing Parquet files into fewer, larger shards.

Reads the small shards from data/ and rewrites them with many more
rows per shard to minimize the total number of files.

Usage:
    uv run reshard_parquet.py [--input-dir PATH] [--output-dir PATH] [--rows-per-shard N]
"""

import argparse
import os
from pathlib import Path

import pyarrow.parquet as pq


DEFAULT_DIR = os.path.expanduser("~/parameter-golf/data")


def collect_shards(directory: Path, prefix: str) -> list[Path]:
    """Return sorted shard paths matching a split prefix."""
    return sorted(p for p in directory.glob(f"{prefix}-*.parquet"))


def reshard(shards: list[Path], split: str, out_dir: Path) -> None:
    """Stream input shards into a single output file using ParquetWriter (low memory)."""
    total_rows = sum(pq.read_metadata(s).num_rows for s in shards)
    print(f"  {split}: {total_rows:,} rows across {len(shards)} shards -> 1 output file")

    name = f"{split}-00000-of-00001.parquet"
    out_path = out_dir / name
    schema = pq.read_schema(shards[0])
    writer = pq.ParquetWriter(
        out_path, schema,
        compression="zstd",
        write_page_index=True,
        use_content_defined_chunking=True,
    )

    rows_written = 0
    for shard_path in shards:
        table = pq.read_table(shard_path)
        writer.write_table(table, row_group_size=200_000)
        rows_written += table.num_rows
        print(f"    appended {shard_path.name} ({rows_written:,}/{total_rows:,} rows)")
        del table

    writer.close()
    print(f"    wrote {name}")


def main() -> None:
    parser = argparse.ArgumentParser(description="Re-shard Parquet files into fewer, larger shards")
    parser.add_argument("--input-dir", default=DEFAULT_DIR, help="Directory with existing shards")
    parser.add_argument("--output-dir", default=None, help="Output directory (default: input-dir + '_resharded')")
    args = parser.parse_args()

    in_dir = Path(args.input_dir)
    out_dir = Path(args.output_dir) if args.output_dir else in_dir.parent / "data_resharded"
    out_dir.mkdir(parents=True, exist_ok=True)

    print(f"Input:  {in_dir}")
    print(f"Output: {out_dir}")
    print()

    for split in ("validation", "train"):
        shards = collect_shards(in_dir, split)
        if shards:
            reshard(shards, split, out_dir)

    print(f"\nDone! Output in {out_dir}")


if __name__ == "__main__":
    main()