split_idx.py

"""Parallel split of sorted index parquets into parts < 50 GB for HTTP upload to HF.

Windows Xet upload stalls on large files; the HTTP fallback caps at 50 GB/file.
Parts preserve sorted order (contiguous row-group ranges), so DuckDB zone-map
pruning still works across parts. Multiprocessing (N_WORKERS) since pyarrow's
writer is mostly single-threaded. Resumable: existing completed parts are skipped.
"""
import multiprocessing as mp
import os
import time

import pyarrow.parquet as pq

BASE = os.path.dirname(os.path.abspath(__file__))
# budget on UNCOMPRESSED size; SNAPPY output lands well under the 50 GB cap
TARGET_UNCOMPRESSED = 45 * 1024**3
N_WORKERS = 5
BATCH_GROUPS = 100  # row groups read per batch (~1-3 GiB in RAM)

JOBS = ["idx_phone.parquet", "idx_aadhar.parquet"]


def plan_partitions(sizes, target):
    """Split row-group size list into (start, end) buckets each ~target bytes."""
    parts = []
    acc = 0
    start = 0
    for i, s in enumerate(sizes):
        acc += s
        if acc >= target and i - start > 200:  # at least 200 row groups per part
            parts.append((start, i + 1))
            start = i + 1
            acc = 0
    if start < len(sizes):
        parts.append((start, len(sizes)))
    return parts


def split_one(src, pi, a, b):
    out = f"{os.path.splitext(src)[0]}.{pi}.parquet"
    if os.path.exists(out) and os.path.getsize(out) > 0:
        print(f"  {out} exists, skip", flush=True)
        return
    tmp = out + ".tmp"
    if os.path.exists(tmp):
        os.remove(tmp)
    t0 = time.time()
    pf = pq.ParquetFile(src)
    with pq.ParquetWriter(tmp, pf.schema_arrow, compression="snappy",
                          use_dictionary=True, data_page_version="2.0",
                          version="2.6") as w:
        g = a
        while g < b:
            g2 = min(g + BATCH_GROUPS, b)
            w.write_table(pf.read_row_groups(list(range(g, g2))))
            g = g2
    os.rename(tmp, out)
    print(f"  DONE {out} ({os.path.getsize(out)/1073741824:.1f} GiB) "
          f"in {time.time()-t0:.0f}s", flush=True)


def worker(wid, tasklist):
    for src, pi, a, b in tasklist:
        print(f"[w{wid}] start {src} part {pi} (groups {a}-{b})", flush=True)
        try:
            split_one(src, pi, a, b)
        except Exception as e:
            print(f"[w{wid}] FAIL {src} part {pi}: {e}", flush=True)


def main():
    jobs = []  # (src, [(a,b), ...])
    for src in JOBS:
        if not os.path.exists(src):
            print(f"SKIP {src}: not found", flush=True)
            continue
        pf = pq.ParquetFile(src)
        sizes = [pf.metadata.row_group(i).total_byte_size
                 for i in range(pf.num_row_groups)]
        boundaries = plan_partitions(sizes, TARGET_UNCOMPRESSED)
        print(f"{src}: {pf.num_row_groups} row groups, "
              f"{sum(sizes)/1073741824:.1f} GiB uncompressed -> "
              f"{len(boundaries)} parts", flush=True)
        jobs.append((src, boundaries))

    flat = [(src, i, a, b)
            for src, bnd in jobs for i, (a, b) in enumerate(bnd)]
    procs = []
    for wi in range(N_WORKERS):
        mine = flat[wi::N_WORKERS]
        p = mp.Process(target=worker, args=(wi, mine))
        p.start()
        procs.append(p)
    for p in procs:
        p.join()
    print("ALL DONE", flush=True)


if __name__ == "__main__":
    main()