build_window_index.py

#!/usr/bin/env python3
"""Build per-directory window index parquets for remote data parquets on HF.

For every directory of data parquets in a HF dataset repo, this script reads
only parquet footers plus the `duration` column (HTTP range requests, no audio
download), computes sliding-window offsets per row, and writes one index
parquet per directory:

    source_path   string          e.g. hf://datasets/<repo>/<dir>/<file>.parquet
    row_group     int32           parquet row-group ordinal inside source_path
    row_in_group  int32           row ordinal inside that row group
    valid_offsets list<float64>   window start seconds within the session
    num_windows   int32           len(valid_offsets)

Offsets follow the NeMo semantic: every full window that fits (stepping by
--shift-sec); a session shorter than one window still yields a single 0.0
offset so the loader can pad it.

Example:
    python build_window_index.py \
        --repo-id tsw0411/Real_sd_ds_0701 --output-dir data/index --upload
"""

from __future__ import annotations

import argparse
import logging
import time
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path

import pyarrow as pa
import pyarrow.parquet as pq
from huggingface_hub import HfApi, HfFileSystem

LOG = logging.getLogger("build_window_index")

INDEX_SCHEMA = pa.schema(
    [
        pa.field("source_path", pa.string()),
        pa.field("row_group", pa.int32()),
        pa.field("row_in_group", pa.int32()),
        pa.field("valid_offsets", pa.list_(pa.float64())),
        pa.field("num_windows", pa.int32()),
    ]
)


def compute_valid_offsets(duration: float, window_sec: float, shift_sec: float) -> list[float]:
    """Every full window that fits; short sessions get a single padded window at 0."""
    offsets: list[float] = []
    offset = 0.0
    while offset + window_sec <= duration + 1e-6:
        offsets.append(round(offset, 6))
        offset += shift_sec
    if not offsets and duration > 0:
        offsets.append(0.0)
    return offsets


def index_one_parquet(
    fs: HfFileSystem,
    fs_path: str,
    window_sec: float,
    shift_sec: float,
    max_retries: int = 4,
) -> list[dict]:
    """Read footer + duration column of one remote parquet and index its rows.

    Transient network failures (SSL resets, closed HTTP clients) are retried
    with a fresh, uncached filesystem instance because a failed request can
    leave the shared HfFileSystem's HTTP client closed for all threads.
    """
    source_path = f"hf://{fs_path}"
    for attempt in range(max_retries + 1):
        try:
            records: list[dict] = []
            with fs.open(fs_path, "rb") as handle:
                parquet_file = pq.ParquetFile(handle)
                for group in range(parquet_file.num_row_groups):
                    durations = parquet_file.read_row_group(group, columns=["duration"])
                    for row_in_group, duration in enumerate(
                        durations.column("duration").to_pylist()
                    ):
                        offsets = compute_valid_offsets(float(duration), window_sec, shift_sec)
                        records.append(
                            {
                                "source_path": source_path,
                                "row_group": group,
                                "row_in_group": row_in_group,
                                "valid_offsets": offsets,
                                "num_windows": len(offsets),
                            }
                        )
            return records
        except Exception as exc:
            if attempt == max_retries:
                raise
            delay = min(30.0, 2.0**attempt)
            LOG.warning(
                "%s: read failed (%s: %s); retry %d/%d in %.0fs",
                fs_path,
                type(exc).__name__,
                exc,
                attempt + 1,
                max_retries,
                delay,
            )
            time.sleep(delay)
            fs = HfFileSystem(skip_instance_cache=True)
    raise AssertionError("unreachable")


def build_index_for_dir(
    fs: HfFileSystem,
    dir_name: str,
    fs_paths: list[str],
    output_path: Path,
    window_sec: float,
    shift_sec: float,
    workers: int,
) -> None:
    with ThreadPoolExecutor(max_workers=workers) as pool:
        per_file = list(
            pool.map(
                lambda path: index_one_parquet(fs, path, window_sec, shift_sec),
                sorted(fs_paths),
            )
        )
    records = [record for file_records in per_file for record in file_records]

    output_path.parent.mkdir(parents=True, exist_ok=True)
    pq.write_table(pa.Table.from_pylist(records, schema=INDEX_SCHEMA), output_path)
    LOG.info(
        "%s: %d files, %d rows, %d windows -> %s",
        dir_name,
        len(fs_paths),
        len(records),
        sum(record["num_windows"] for record in records),
        output_path,
    )


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
    parser.add_argument("--repo-id", required=True, help="HF dataset repo, e.g. user/name")
    parser.add_argument("--output-dir", type=Path, required=True, help="Local staging directory")
    parser.add_argument("--window-sec", type=float, default=90.0)
    parser.add_argument("--shift-sec", type=float, default=8.0)
    parser.add_argument(
        "--test-suffix",
        default="_test",
        help="Directories ending with this suffix use --test-shift-sec (default: _test)",
    )
    parser.add_argument(
        "--test-shift-sec",
        type=float,
        default=None,
        help="Shift for test directories; defaults to --window-sec (non-overlapping windows)",
    )
    parser.add_argument("--workers", type=int, default=8, help="Concurrent remote reads per directory")
    parser.add_argument("--dirs", nargs="*", help="Only process these directory names")
    parser.add_argument("--force", action="store_true", help="Rebuild even if local index exists")
    parser.add_argument("--upload", action="store_true", help="Upload indexes to <repo>/index/")
    args = parser.parse_args()

    logging.basicConfig(level=logging.INFO, format="%(levelname)s %(message)s")

    fs = HfFileSystem()
    repo_prefix = f"datasets/{args.repo_id}"
    all_parquets = fs.glob(f"{repo_prefix}/**/*.parquet")

    by_dir: dict[str, list[str]] = {}
    for fs_path in all_parquets:
        relative = fs_path[len(repo_prefix) + 1 :]
        parts = relative.split("/")
        if parts[0] == "index" or len(parts) < 2:
            continue  # skip existing indexes and any root-level parquet
        by_dir.setdefault("/".join(parts[:-1]), []).append(fs_path)

    if args.dirs:
        missing = sorted(set(args.dirs) - set(by_dir))
        if missing:
            parser.error(f"Directories not found in repo: {missing}")
        by_dir = {name: by_dir[name] for name in args.dirs}

    test_shift_sec = args.test_shift_sec if args.test_shift_sec is not None else args.window_sec
    LOG.info(
        "Indexing %d directories (window=%ss, shift=%ss, %s-suffix shift=%ss)",
        len(by_dir),
        args.window_sec,
        args.shift_sec,
        args.test_suffix,
        test_shift_sec,
    )
    for dir_name in sorted(by_dir):
        output_path = args.output_dir / f"{dir_name.replace('/', '__')}.parquet"
        if output_path.exists() and not args.force:
            LOG.info("%s: index exists, skipping (use --force to rebuild)", dir_name)
            continue
        shift_sec = test_shift_sec if dir_name.endswith(args.test_suffix) else args.shift_sec
        LOG.info("%s: shift=%ss", dir_name, shift_sec)
        build_index_for_dir(
            fs=fs,
            dir_name=dir_name,
            fs_paths=by_dir[dir_name],
            output_path=output_path,
            window_sec=args.window_sec,
            shift_sec=shift_sec,
            workers=args.workers,
        )

    if args.upload:
        LOG.info("Uploading %s -> %s/index/", args.output_dir, args.repo_id)
        HfApi().upload_folder(
            folder_path=str(args.output_dir),
            path_in_repo="index",
            repo_id=args.repo_id,
            repo_type="dataset",
            commit_message="feat: add per-directory window index parquets",
        )
        LOG.info("Upload complete")


if __name__ == "__main__":
    main()