tools/reencode_parquet.py

#!/usr/bin/env python3
"""Probe / dry-run / re-encode parquet files to dictionary + a chosen codec to shrink them.

Only the on-disk encoding changes; the logical data is preserved bit-for-bit, so HF
`load_dataset` features and config bundling are unaffected. Safe by design:
  * verifies EVERY file before replacing: full-table data+type equality (orig.equals,
    check_metadata=False) AND byte-identical 'huggingface' schema metadata AND the array
    column's ARROW:extension metadata (what lets datasets rebuild Array2D)
  * atomic: writes a temp then os.replace() — an interrupted run never leaves a half file
  * one file at a time (peak RAM ~= largest single file)
  * idempotent: in `reencode`, files already dictionary-encoded in the target codec are skipped

Modes (PATH may be a single .parquet file OR a folder, recursed for **/*.parquet):
  probe     read-only; measure target size and print the table. No verification, no write.
  dryrun    full verified pipeline (re-encode to temp + verify) but never replace. Prints table.
  reencode  same verified pipeline, then atomic in-place replace for files that shrink (>2%).

Every mode ends with a per-file + per-folder + grand-total size table.

Usage:
  python reencode_parquet.py probe    PATH [--codec ZSTD]
  python reencode_parquet.py dryrun   PATH [--codec ZSTD]
  python reencode_parquet.py reencode PATH [--codec ZSTD]

--codec defaults to SNAPPY (pyarrow's default; what this dataset is already encoded with).
"""
import argparse
import glob
import os
import tempfile

import pyarrow.parquet as pq

HF_KEY = b"huggingface"


def iter_parquet(path):
    """A single .parquet file -> [path]; a folder -> sorted **/*.parquet (excl. hidden)."""
    if os.path.isfile(path):
        return [path]
    files = glob.glob(os.path.join(path, "**", "*.parquet"), recursive=True)
    return sorted(f for f in files if "/." not in f and not os.path.basename(f).startswith("."))


def uncompressed_size(path):
    """Sum of total_uncompressed_size over every column of every row group (logical size)."""
    md = pq.ParquetFile(path).metadata
    total = 0
    for g in range(md.num_row_groups):
        rg = md.row_group(g)
        total += sum(rg.column(i).total_uncompressed_size for i in range(rg.num_columns))
    return total


def _big_column(path):
    """The biggest (uncompressed) column of the first row group — representative of encoding."""
    rg = pq.ParquetFile(path).metadata.row_group(0)
    cols = [rg.column(i) for i in range(rg.num_columns)]
    return max(cols, key=lambda c: c.total_uncompressed_size)


def _array_ext_meta(table):
    """Return the ARROW:extension metadata of the first extension column, or None."""
    for f in table.schema:
        m = f.metadata or {}
        if b"ARROW:extension:name" in m:
            return dict(m)
    return None


def write_temp(table, codec, dirpath):
    """Write `table` to a temp file in `dirpath` with `codec` + dictionary. Return temp path."""
    fd, tmp = tempfile.mkstemp(dir=dirpath, suffix=".reenc.tmp")
    os.close(fd)
    pq.write_table(table, tmp, compression=codec, use_dictionary=True)
    return tmp


def verify(orig, tmp_path):
    """Raise if the re-encoded temp is not bit-for-bit equal to `orig` (data + key metadata)."""
    back = pq.read_table(tmp_path)
    if not orig.equals(back, check_metadata=False):
        raise RuntimeError("DATA MISMATCH after re-encode")
    if (back.schema.metadata or {}).get(HF_KEY) != (orig.schema.metadata or {}).get(HF_KEY):
        raise RuntimeError("'huggingface' schema metadata changed")
    if _array_ext_meta(back) != _array_ext_meta(orig):
        raise RuntimeError("array extension metadata changed")


def process(path, mode, codec):
    """Return a row dict {path, current, uncompressed, target, status} for one file.

    Modes: probe (measure only), dryrun (measure + verify), reencode (measure + verify + replace).
    """
    current = os.path.getsize(path)
    uncompr = uncompressed_size(path)

    # reencode is idempotent: skip files already dict-encoded in the target codec.
    if mode == "reencode":
        big = _big_column(path)
        if big.has_dictionary_page and big.compression == codec:
            return dict(path=path, current=current, uncompressed=uncompr,
                        target=current, status="skipped (already)")

    orig = pq.read_table(path)
    dirpath = os.path.dirname(path) or "."
    tmp = write_temp(orig, codec, dirpath)
    try:
        if mode != "probe":
            verify(orig, tmp)
        target = os.path.getsize(tmp)
        beneficial = target < current * 0.98  # only worth it if it saves >2%

        if mode == "reencode" and beneficial:
            os.replace(tmp, path)  # atomic in-place swap
            tmp = None
            return dict(path=path, current=current, uncompressed=uncompr,
                        target=target, status="written")

        status = ("would shrink" if beneficial else "no gain") if mode != "probe" else "probed"
        return dict(path=path, current=current, uncompressed=uncompr,
                    target=target, status=status)
    finally:
        if tmp and os.path.exists(tmp):
            os.remove(tmp)


def _mb(n):
    return f"{n / 1e6:9.1f} MB"


def _savings(current, target):
    saved = current - target
    pct = (1 - target / current) * 100 if current else 0
    return f"{saved / 1e6:9.1f} MB ({pct:4.0f}%)"


def print_table(rows):
    """Per-file rows grouped by parent folder, a subtotal per folder, then a grand total."""
    name_w = max([len(os.path.basename(r["path"])) for r in rows] + [12])
    header = (f"  {'file'.ljust(name_w)} | {'current':>12} | {'uncompr':>12} | "
              f"{'target':>12} | {'savings':>18}")
    sep = "  " + "-" * (len(header) - 2)

    folders = {}
    for r in rows:
        folders.setdefault(os.path.dirname(r["path"]) or ".", []).append(r)

    g_cur = g_unc = g_tgt = 0
    print(header)
    print(sep)
    for folder in sorted(folders):
        fr = folders[folder]
        f_cur = f_unc = f_tgt = 0
        for r in fr:
            print(f"  {os.path.basename(r['path']).ljust(name_w)} | "
                  f"{_mb(r['current'])} | {_mb(r['uncompressed'])} | {_mb(r['target'])} | "
                  f"{_savings(r['current'], r['target'])}   {r['status']}")
            f_cur += r["current"]
            f_unc += r["uncompressed"]
            f_tgt += r["target"]
        print(f"  {('Folder ' + folder).ljust(name_w)} | "
              f"{_mb(f_cur)} | {_mb(f_unc)} | {_mb(f_tgt)} | {_savings(f_cur, f_tgt)}")
        print(sep)
        g_cur += f_cur
        g_unc += f_unc
        g_tgt += f_tgt

    print(f"  {'GRAND TOTAL'.ljust(name_w)} | "
          f"{_mb(g_cur)} | {_mb(g_unc)} | {_mb(g_tgt)} | {_savings(g_cur, g_tgt)}")


def main():
    ap = argparse.ArgumentParser(description=__doc__,
                                 formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("mode", choices=["probe", "dryrun", "reencode"])
    ap.add_argument("path", help="a .parquet file or a folder (recursed for **/*.parquet)")
    ap.add_argument("--codec", default="SNAPPY", help="target compression codec (default: SNAPPY)")
    args = ap.parse_args()
    codec = args.codec.upper()

    files = iter_parquet(args.path)
    if not files:
        raise SystemExit(f"no parquet files under {args.path}")

    print(f"MODE: {args.mode.upper()}   codec: {codec}   ({len(files)} files)\n")
    rows = []
    for f in files:
        rows.append(process(f, args.mode, codec))
    print()
    print_table(rows)


if __name__ == "__main__":
    main()