Datasets:

jncarlo
/

tinymix

"""Parallel, resource-bounded build for big sizes: python build_pool.py 1000M [options]
Same filters and safety nets as build.py (process, probe and the relaxation ladder come from there).
What is different:
  - Jobs. A source whose estimated time (from work/100M/stats.json) exceeds --job-hours is split into
    shards (whole dataset files, or every k-th row). Jobs run longest first in at most --workers
    processes, at below-normal priority and with one thread each, so the machine stays usable.
  - Memory. Dedup keeps 8-byte keys per document (the md5 key and the LSH band keys build.Dedup
    uses) instead of datasketch's index, and each job only holds its own. A global pass over the key
    files with numpy then drops documents that repeat across jobs; the first occurrence in MIX order
    wins, and a document that matches any earlier one is dropped (build.Dedup only compares against
    kept ones, so this is marginally stricter). Jobs collect --overshoot extra words so that pass
    rarely leaves a source short, and the excess is trimmed.
    The scheduler waits to start jobs while free RAM is below --min-free-gb and stops the youngest job
    when it falls below --kill-free-gb; a stopped job resumes later.
  - Robustness. Each job checkpoints every --ckpt-seconds, after flushing its data, so a hard kill
    loses at most that interval. On a network error it reopens the stream and skips the documents
    already seen (the streams are deterministic: fixed seed and shard). A job that crashes restarts
    from its checkpoint up to --retries times. Rerunning the same command resumes everything and
    skips finished jobs.
Output: work/<size>/<source>.jsonl and <source>.topup.jsonl plus stats.json, which assemble.py reads as
before. Per-job files (data, keys, checkpoints, logs) live in work/<size>/jobs/.
Testing: --scale 0.002 multiplies every quota (output in work/<size>_x0.002/); a small --job-hours
forces shards.
"""
import argparse
import hashlib
import itertools
import json
import math
import multiprocessing as mp
import os
import re
import shutil
import statistics
import sys
import time
import traceback
from collections import Counter
from pathlib import Path

# one thread per process for numpy, BLAS and Arrow; spawned workers inherit these
for _v in ("OMP_NUM_THREADS", "OPENBLAS_NUM_THREADS", "MKL_NUM_THREADS"):
    os.environ.setdefault(_v, "1")

import numpy as np
import psutil
from datasketch import MinHash, MinHashLSH
from wordfreq import top_n_list

import build
import mix
import sources

ROOT = build.ROOT
_LSH = MinHashLSH(threshold=0.8, num_perm=64)  # build.Dedup's parameters; only b and r are used
BANDS, ROWS = _LSH.b, _LSH.r
_MH = MinHash(num_perm=64)  # reused permutations; building them per document is most of MinHash's cost
PERMS, SCHEME = _MH.permutations, _MH.scheme
# odd 64-bit multipliers that fold the ROWS values of a band into one key
MIXER = np.random.RandomState(7).randint(1, 2**62, size=ROWS, dtype=np.int64).astype(np.uint64) | np.uint64(1)
REC = np.dtype([("x", "<u8"), ("b", "<u8", (BANDS,)), ("w", "<u4")])
# not streamed through sources._ds (so they cannot be sharded), or row order matters (poetry)
NO_SHARD = {"tinystories", "tinydialogues", "tinyhelen", "poetry"}
# Gutenberg rows are whole books in ~300 MB row groups: filling the shuffle buffer from datasets'
# default 10 files at once peaks at 6 GB, from one file at 1.7 GB (and is faster)
INPUT_SHARDS = {"gutenberg_books": 1}
# expected peak RAM per worker (GB), measured: a streamed parquet source with a 20k-row shuffle
# buffer sits near 1.4. The scheduler raises these to the peaks it observes during the run.
MEM_GB = {"gutenberg_books": 2.0}
MEM_DEFAULT_GB = 1.5


def doc_keys(text):
    """The md5 key and the LSH band keys build.Dedup would use for this text, as 64-bit ints."""
    x = int.from_bytes(hashlib.md5(re.sub(r"\W+", "", text.lower()).encode()).digest()[:8], "little")
    w = text.lower().split()
    mh = MinHash(num_perm=64, permutations=PERMS, scheme=SCHEME)
    mh.update_batch([" ".join(w[i:i + 5]).encode() for i in range(max(len(w) - 4, 1))])
    hv = mh.hashvalues[:BANDS * ROWS].astype(np.uint64).reshape(BANDS, ROWS)
    with np.errstate(over="ignore"):
        return x, (hv * MIXER).sum(axis=1, dtype=np.uint64)


class Keys:
    """build.Dedup on plain sets of 64-bit keys: same decisions, a fraction of the memory."""

    def __init__(self):
        self.exact, self.bands = set(), [set() for _ in range(BANDS)]

    def load(self, path):
        if path.exists():
            recs = np.fromfile(path, dtype=REC)
            self.exact.update(recs["x"].tolist())
            for i, s in enumerate(self.bands):
                s.update(recs["b"][:, i].tolist())

    def seen(self, x, bands):
        if x in self.exact:
            return True
        self.exact.add(x)
        b = bands.tolist()
        if any(k in s for s, k in zip(self.bands, b)):
            return True
        for s, k in zip(self.bands, b):
            s.add(k)
        return False


def write_json(path, obj):
    tmp = path.with_suffix(path.suffix + ".tmp")
    tmp.write_text(json.dumps(obj), encoding="utf-8")
    for i in range(20):  # Windows refuses the replace while another process has the file open
        try:
            os.replace(tmp, path)
            return
        except PermissionError:
            time.sleep(0.5)
    os.replace(tmp, path)


def read_json(path):
    try:
        return json.loads(path.read_text(encoding="utf-8"))
    except (OSError, ValueError):
        return None


def apply_scale(scale, size):
    """--scale: multiply this size's quotas in memory (parent and every worker), for test runs."""
    if scale == 1:
        return
    i = mix.SIZES[size]
    for k, (g, q) in list(mix.MIX.items()):
        mix.MIX[k] = (g, tuple(v * scale if j == i else v for j, v in enumerate(q)))


def paths(jobs_dir, job):
    return {ext: jobs_dir / f"{job['id']}.{ext}" for ext in ("jsonl", "keys", "ckpt", "done", "log")}


def is_done(jobs_dir, job):
    d = read_json(paths(jobs_dir, job)["done"])
    return bool(d) and abs(d["quota_M"] - job["quota"] / 1e6) < 1e-6


# ---------------------------------------------------------------- worker side

def robust_stream(name, size, start, st, tries=8):
    """Yield the source's documents from index `start`; on a stream error reopen the stream and skip
    what was already yielded."""
    pos, fails = start, 0
    while True:
        try:
            for doc in itertools.islice(sources.REGISTRY[name](size), pos, None):
                pos += 1
                fails = 0
                yield doc
            return
        except Exception as e:
            fails += 1
            st["retries"] = st.get("retries", 0) + 1
            if fails > tries:
                raise
            wait = min(600, 15 * 2 ** fails)
            print(f"stream error at doc {pos} ({e!r}); retry {fails}/{tries} in {wait}s", flush=True)
            time.sleep(wait)


def consumed_docs(job, jobs_dir):
    """Where a top-up starts: after what its source's primary job read, when both read the same stream."""
    if not job["tag"] or list(job["shard"]) != [0, 1]:
        return 0
    d = read_json(jobs_dir / f"{job['name']}.s0of1.done")
    return d["scanned_docs"] if d else 0


def run_job(job, size, jobs_dir, ckpt_seconds):
    """build.build_source for one job, with checkpoints and resume."""
    p = paths(jobs_dir, job)
    if is_done(jobs_dir, job):
        return
    name, quota = job["name"], job["quota"]
    genre = mix.MIX[name][0]
    base = mix.TIERS[size]
    vocab = set(top_n_list("en", base["vocab"]))
    dedup = Keys()
    for seed in job["seeds"]:  # a top-up must not repeat its own source's primary files
        dedup.load(Path(seed))
    st = read_json(p["ckpt"])
    # resuming skips documents by position, so it is only valid on the very same stream
    stream = [list(job["shard"]), INPUT_SHARDS.get(name)]
    if st and (abs(st["quota"] - quota) > 1 or st.get("stream") != stream):
        st = None  # options changed since the checkpoint: start over
    openers = Counter()
    if st:
        # drop whatever was written after the last checkpoint, then rebuild state from the rest
        with p["jsonl"].open("r+b") as f:
            f.truncate(st["bytes"])
        with p["keys"].open("r+b") as f:
            f.truncate(st["kept"] * REC.itemsize)
        dedup.load(p["keys"])
        for line in p["jsonl"].open(encoding="utf-8"):
            openers[" ".join(json.loads(line)["text"].lower().split()[:4])] += 1
        print(f"resume at doc {st['seen_docs']}, {st['got'] / 1e6:.2f}M/{quota / 1e6:.2f}M words", flush=True)
    else:
        step, probe_report = build.probe(name, size, genre, base, vocab, quota)
        print(f"probe: start at ladder step {step}; {probe_report[-1]}", flush=True)
        st = dict(quota=quota, stream=stream, step=step, probe=probe_report, relaxed=[dict(step=step, at_words=0)],
                  got=0, scanned=0, kept=0, seen_docs=consumed_docs(job, jobs_dir), rej={}, seconds=0, retries=0,
                  bytes=0)
        p["jsonl"].write_bytes(b"")
        p["keys"].write_bytes(b"")
    step = st["step"]
    tier, cov = build.at_step(base, step)
    got, scanned, kept, seen = st["got"], st["scanned"], st["kept"], st["seen_docs"]
    rej = Counter(st["rej"])
    win_scanned = win_kept = win_docs = 0
    abort, t0, last, elapsed0 = None, time.time(), time.time(), st["seconds"]
    with p["jsonl"].open("ab") as f, p["keys"].open("ab") as kf:

        def save():
            for h in (f, kf):
                h.flush()
                os.fsync(h.fileno())
            st.update(step=step, got=got, scanned=scanned, kept=kept, seen_docs=seen, rej=dict(rej),
                      seconds=round(elapsed0 + time.time() - t0), bytes=f.tell())
            write_json(p["ckpt"], st)

        stream = robust_stream(name, size, seen, st) if got < quota else ()
        for doc in stream:
            seen += 1
            win_docs += 1
            w = build.doc_words(doc)
            scanned += w
            win_scanned += w
            if scanned > quota * mix.MAX_SCAN:
                abort = "scan_cap"
                break
            docs, why = build.process(doc, genre, tier, vocab, name, cov)
            if why:
                rej[why] += 1
            for text in docs:
                op = " ".join(text.lower().split()[:4])
                if genre != "poetry" and openers[op] >= max(5, 0.005 * kept):
                    rej["opener_cap"] += 1
                    continue
                x, bands = doc_keys(text)
                if dedup.seen(x, bands):
                    rej["duplicate"] += 1
                    win_scanned -= len(text.split())  # says nothing about the yield of what is left
                    continue
                openers[op] += 1
                kept += 1
                n = len(text.split())
                got += n
                win_kept += n
                f.write((json.dumps({"text": text, "genre": genre, "src": name}) + "\n").encode("utf-8"))
                rec = np.zeros(1, REC)
                rec["x"], rec["b"], rec["w"] = x, bands, n
                kf.write(rec.tobytes())
            if got >= quota:
                break
            if time.time() - last >= ckpt_seconds:
                save()
                last = time.time()
                print(f"{seen} docs, kept {kept}, {got / 1e6:.2f}M/{quota / 1e6:.2f}M words, "
                      f"step {step}, top rejections {rej.most_common(3)}", flush=True)
            if win_docs >= mix.WINDOW:
                # same rule as build.py: climb a ladder step when the recent yield cannot finish
                # within the scan budget, give up at the top of the ladder
                recent = win_kept / win_scanned if win_scanned > 0 else 0
                need = (quota - got) / recent if recent else float("inf")
                if win_scanned > 0 and need > quota * mix.MAX_SCAN - scanned:
                    if step + 1 < len(mix.LADDER):
                        step += 1
                        tier, cov = build.at_step(base, step)
                        st["relaxed"].append(dict(step=step, at_words=got))
                        print(f"yield {100 * recent:.1f}% too low, relaxing to step {step}", flush=True)
                    else:
                        abort = "projected_shortfall"
                        break
                win_scanned = win_kept = win_docs = 0
        save()
    stats = dict(job=job["id"], source=name + job["tag"], genre=genre, target_M=job["target"] / 1e6,
                 quota_M=quota / 1e6, words_M=round(got / 1e6, 3), docs=kept, scanned_docs=seen,
                 scanned_words_M=round(scanned / 1e6, 1), seconds=st["seconds"],
                 stopped="quota" if got >= quota else (abort or "source_exhausted"),
                 relaxation=st["relaxed"], probe=st["probe"], rejections=dict(rej.most_common()),
                 stream_retries=st.get("retries", 0))
    write_json(p["done"], stats)
    p["ckpt"].unlink(missing_ok=True)
    print(f"done: {stats['words_M']}M words, stopped by {stats['stopped']}", flush=True)


def worker(job, size, scale, ckpt_seconds, jobs_dir):
    jobs_dir = Path(jobs_dir)
    sys.stdout = sys.stderr = paths(jobs_dir, job)["log"].open("a", encoding="utf-8", buffering=1)
    print(f"== {job['id']} pid {os.getpid()} {time.strftime('%Y-%m-%d %H:%M:%S')}", flush=True)
    psutil.Process().nice(psutil.BELOW_NORMAL_PRIORITY_CLASS if os.name == "nt" else 10)
    try:
        import pyarrow
        pyarrow.set_cpu_count(1)
        pyarrow.set_io_thread_count(2)
    except Exception:
        pass
    apply_scale(scale, size)
    sources.SHARD = tuple(job["shard"])
    sources.INPUT_SHARDS.update(INPUT_SHARDS)
    try:
        run_job(job, size, jobs_dir, ckpt_seconds)
    except Exception:
        traceback.print_exc()
        sys.exit(1)


# ---------------------------------------------------------------- parent side

class Log:
    def __init__(self, path):
        self.f = path.open("a", encoding="utf-8", buffering=1)

    def __call__(self, msg):
        line = f"{time.strftime('%H:%M:%S')} {msg}"
        print(line, flush=True)
        self.f.write(line + "\n")


def seconds_per_word():
    """Build seconds per target word for each source, measured on the 100M build."""
    s = read_json(ROOT / "work" / "100M" / "stats.json") or {}
    rates = {r["source"]: r["seconds"] / (r["target_M"] * 1e6)
             for r in s.get("sources", []) if r.get("target_M") and "seconds" in r}
    default = statistics.median(rates.values()) if rates else 1e-4
    return lambda name: rates.get(name, default)


def plan_jobs(targets, tag, args, seeds_for):
    """Split each (source, words) into shards of at most --job-hours estimated time."""
    rate, jobs = seconds_per_word(), []
    for name, target in targets:
        if target <= 0:
            continue
        est = rate(name) * target
        k = 1 if name in NO_SHARD else max(1, min(args.max_shards, math.ceil(est / (args.job_hours * 3600))))
        for i in range(k):
            jobs.append(dict(id=f"{name}{tag}.s{i}of{k}", name=name, tag=tag, shard=(i, k), target=target / k,
                             quota=target / k * (1 + args.overshoot), est=est / k, seeds=seeds_for(name)))
    return jobs


def run_pool(jobs, args, size, jobs_dir, log):
    """Run jobs in at most --workers processes, longest first, within the RAM and disk limits.
    Returns the jobs that failed more than --retries times."""
    pending = sorted((j for j in jobs if not is_done(jobs_dir, j)), key=lambda j: -j["est"])
    if not pending:
        return []
    total_est = sum(j["est"] for j in pending)
    log(f"{len(pending)} jobs to run ({len(jobs) - len(pending)} already done), about "
        f"{total_est / 3600:.1f} CPU hours by the 100M rates, {args.workers} workers")
    ctx = mp.get_context("spawn")
    running, attempts, failed = {}, Counter(), []
    # a job stopped twice for memory runs alone from then on, otherwise it could be stopped forever
    mem_stops, alone = Counter(), set()
    peak = Counter()  # observed peak RSS per source, GB
    hold = kill_hold = last_status = 0.0

    def need(job):
        return max(MEM_GB.get(job["name"], MEM_DEFAULT_GB), peak[job["name"]])

    def rss(proc):
        try:
            return psutil.Process(proc.pid).memory_info().rss / 1e9
        except psutil.Error:
            return 0.0

    def stop(jid, why):
        proc, job, _ = running.pop(jid)
        proc.terminate()
        proc.join()
        pending.insert(0, job)
        log(f"stopped {jid} ({why}); it resumes from its checkpoint")

    def pick(free, reserved):
        """First pending job whose expected peak leaves the stop threshold clear, so a big job that
        does not fit yet does not hold back smaller ones; a job marked to run alone waits for the rest."""
        if alone & set(running):
            return None
        for j in pending:
            if j["id"] in alone:
                return j if not running else None
            if not running or free - reserved - need(j) >= args.kill_free_gb + 0.5:
                return j
        return None

    try:
        while pending or running:
            now = time.time()
            for jid, (proc, job, _) in list(running.items()):
                if proc.is_alive():
                    continue
                proc.join()
                del running[jid]
                d = read_json(paths(jobs_dir, job)["done"])
                if d:
                    log(f"done {jid}: {d['words_M']}M words in {d['seconds'] / 60:.0f} min, stopped by {d['stopped']}")
                    continue
                attempts[jid] += 1
                if attempts[jid] > args.retries:
                    failed.append(job)
                    log(f"FAILED {jid} after {attempts[jid]} attempts, see {paths(jobs_dir, job)['log']}")
                else:
                    pending.insert(0, job)
                    hold = max(hold, now + 30)
                    log(f"{jid} exited with code {proc.exitcode}; restart {attempts[jid]}/{args.retries}")
            free = psutil.virtual_memory().available / 1e9
            # memory a running job is still expected to take (it grows while its buffers fill)
            reserved = 0.0
            for proc, job, _ in running.values():
                r = rss(proc)
                peak[job["name"]] = max(peak[job["name"]], r)
                reserved += max(0.0, need(job) - r)
            disk = shutil.disk_usage(jobs_dir).free / 1e9
            if disk < args.min_disk_gb:
                for jid in list(running):
                    stop(jid, f"only {disk:.1f} GB free on disk")
                raise SystemExit(f"stopped: {disk:.1f} GB free on disk, below --min-disk-gb {args.min_disk_gb}. "
                                 "Free space and rerun the same command to resume.")
            # the last running job is never stopped: that would only trade progress for nothing
            if free < args.kill_free_gb and len(running) > 1 and now >= kill_hold:
                # one job at a time, then give the OS time to reclaim its memory before judging again
                jid = max(running, key=lambda j: running[j][2])
                stop(jid, f"free RAM {free:.1f} GB < {args.kill_free_gb}")
                mem_stops[jid] += 1
                if mem_stops[jid] >= 2 and jid not in alone:
                    alone.add(jid)
                    log(f"{jid} will run alone: it was stopped {mem_stops[jid]} times for memory")
                hold, kill_hold = now + 120, now + 30
            elif (len(running) < args.workers and pending and now >= hold and free >= args.min_free_gb
                  and (job := pick(free, reserved))):
                pending.remove(job)
                proc = ctx.Process(target=worker, args=(job, size, args.scale, args.ckpt_seconds, str(jobs_dir)))
                proc.start()
                running[job["id"]] = (proc, job, now)
                hold = now + args.stagger
                log(f"start {job['id']} (est {job['est'] / 60:.0f} min, ~{need(job):.1f} GB)")
            if now - last_status >= args.status_seconds:
                last_status = now
                parts = []
                for jid, (proc, job, _) in running.items():
                    ck = read_json(paths(jobs_dir, job)["ckpt"])
                    prog = f"{100 * ck['got'] / job['quota']:.0f}%" if ck else "probe"
                    parts.append(f"{jid} {prog} {rss(proc):.1f}GB")
                log(f"status: {len(pending)} pending, free RAM {free:.1f} GB | " + " | ".join(parts))
            time.sleep(2)
    except KeyboardInterrupt:
        for proc, _, _ in running.values():
            proc.terminate()
        log("interrupted; rerun the same command to resume")
        raise SystemExit(130)
    return failed


def global_pass(groups, jobs_dir, out_dir, log, write):
    """groups: [(output file, [job, ...])] in priority order. Drops documents whose md5 key or any band
    key appeared in an earlier document, then keeps each job's documents in order until its target.
    With `write`, writes the output files. Returns {job id: report}."""
    jobs = [j for _, js in groups for j in js if is_done(jobs_dir, j)]
    recs = [np.fromfile(paths(jobs_dir, j)["keys"], dtype=REC) for j in jobs]
    allr = np.concatenate(recs) if recs else np.zeros(0, REC)
    order = np.arange(len(allr))
    dup = np.zeros(len(allr), bool)
    for col in [allr["x"]] + [allr["b"][:, i] for i in range(BANDS)]:
        _, first, inv = np.unique(col, return_index=True, return_inverse=True)
        dup |= first[inv.ravel()] != order
    avail, off = {}, 0
    for j, r in zip(jobs, recs):
        avail[j["id"]] = int(r["w"][~dup[off:off + len(r)]].sum(dtype=np.int64))
        off += len(r)
    # the shards of a source share its target: what one lost to duplicates, another with spare words fills
    budget = {}
    for _, js in groups:
        short = sum(max(0, j["target"] - avail[j["id"]]) for j in js if j["id"] in avail)
        for j in js:
            if j["id"] in avail:
                add = min(max(0, avail[j["id"]] - j["target"]), short)
                budget[j["id"]] = j["target"] + add
                short -= add
    report, masks, off = {}, {}, 0
    for j, r in zip(jobs, recs):
        d, w = dup[off:off + len(r)], r["w"].astype(np.int64)
        off += len(r)
        keep = ~d
        kw = np.where(keep, w, 0)
        before = np.cumsum(kw) - kw
        trimmed = keep & (before >= budget[j["id"]])
        keep &= ~trimmed
        masks[j["id"]] = keep
        report[j["id"]] = dict(docs=int(keep.sum()), words=int(w[keep].sum()), dup_docs=int(d.sum()),
                               dup_words=int(w[d].sum()), trimmed_words=int(w[trimmed].sum()))
    log(f"global dedup over {len(allr)} docs: {int(dup.sum())} duplicates across jobs, "
        f"{sum(r['trimmed_words'] for r in report.values()) / 1e6:.2f}M words trimmed to quota")
    if write:
        wanted = {fname for fname, _ in groups}
        for fname, js in groups:
            tmp = out_dir / (fname + ".tmp")
            with tmp.open("wb") as dst:
                for j in js:
                    if j["id"] not in masks:
                        continue
                    mask, n = masks[j["id"]], 0
                    with paths(jobs_dir, j)["jsonl"].open("rb") as src:
                        for line in src:
                            if n >= len(mask):
                                break
                            if mask[n]:
                                dst.write(line)
                            n += 1
                    if n != len(mask):
                        raise RuntimeError(f"{j['id']}: {n} lines but {len(mask)} keys; delete its files and rerun")
            os.replace(tmp, out_dir / fname)
        # a top-up file from an earlier run that this plan no longer has would still be read by assemble.py
        for stale in out_dir.glob("*.jsonl"):
            if stale.name not in wanted:
                stale.replace(stale.with_suffix(".jsonl.stale"))
                log(f"renamed stale {stale.name} to .stale")
    return report


def summarize(groups, jobs_dir, report):
    out = []
    for fname, js in groups:
        ds = [read_json(paths(jobs_dir, j)["done"]) for j in js]
        ds = [d for d in ds if d]
        rs = [report[j["id"]] for j in js if j["id"] in report]
        rej = Counter()
        for d in ds:
            rej.update(d["rejections"])
        out.append(dict(
            source=fname[:-len(".jsonl")], genre=mix.MIX[js[0]["name"]][0],
            target_M=round(sum(j["target"] for j in js) / 1e6, 3),
            words_M=round(sum(r["words"] for r in rs) / 1e6, 3), docs=sum(r["docs"] for r in rs),
            collected_M=round(sum(d["words_M"] for d in ds), 3),
            cross_job_duplicates_M=round(sum(r["dup_words"] for r in rs) / 1e6, 3),
            trimmed_M=round(sum(r["trimmed_words"] for r in rs) / 1e6, 3),
            scanned_words_M=round(sum(d["scanned_words_M"] for d in ds), 1),
            seconds=sum(d["seconds"] for d in ds), shards=len(js), jobs_done=len(ds),
            stopped=sorted({d["stopped"] for d in ds}), stream_retries=sum(d["stream_retries"] for d in ds),
            rejections=dict(rej.most_common()),
            jobs=[{k: d[k] for k in ("job", "words_M", "stopped", "relaxation", "probe")} for d in ds]))
    return out


def main():
    phys = psutil.cpu_count(logical=False) or 2
    ap = argparse.ArgumentParser(description=__doc__.split("\n")[0])
    ap.add_argument("size", choices=list(mix.SIZES))
    ap.add_argument("--workers", type=int, default=max(1, phys // 2),
                    help=f"parallel jobs (default half the physical cores: {max(1, phys // 2)})")
    ap.add_argument("--min-free-gb", type=float, default=2.0, help="start a job only above this free RAM")
    ap.add_argument("--kill-free-gb", type=float, default=1.0, help="stop the youngest job below this free RAM")
    ap.add_argument("--min-disk-gb", type=float, default=10.0, help="stop everything below this free disk")
    ap.add_argument("--job-hours", type=float, default=1.5, help="split sources estimated above this")
    ap.add_argument("--max-shards", type=int, default=6)
    ap.add_argument("--overshoot", type=float, default=0.03, help="extra share collected before global dedup")
    ap.add_argument("--ckpt-seconds", type=int, default=120)
    ap.add_argument("--retries", type=int, default=3, help="restarts per job after a crash")
    ap.add_argument("--stagger", type=int, default=20, help="seconds between job starts")
    ap.add_argument("--status-seconds", type=int, default=300)
    ap.add_argument("--scale", type=float, default=1.0, help="multiply quotas, for test runs")
    ap.add_argument("--only", nargs="*", default=[], help="build only these sources (no fallbacks)")
    ap.add_argument("--allow-failed", action="store_true", help="go on to fallbacks even if a job failed")
    args = ap.parse_args()
    size = args.size
    apply_scale(args.scale, size)
    out_dir = ROOT / "work" / (size if args.scale == 1 else f"{size}_x{args.scale:g}")
    jobs_dir = out_dir / "jobs"
    jobs_dir.mkdir(parents=True, exist_ok=True)
    log = Log(out_dir / "pool.log")
    psutil.Process().nice(psutil.BELOW_NORMAL_PRIORITY_CLASS if os.name == "nt" else 10)
    t0, idx = time.time(), mix.SIZES[size]
    names = [n for n in mix.MIX if not args.only or n in args.only]
    log(f"== build {size} into {out_dir} ({' '.join(sys.argv[1:])})")

    prim = plan_jobs([(n, mix.MIX[n][1][idx] * 1e6) for n in names], "", args, lambda n: [])
    shards = Counter(j["name"] for j in prim)
    log("plan: " + ", ".join(f"{n} x{k}" for n, k in shards.items()))
    failed = run_pool(prim, args, size, jobs_dir, log)
    if failed and not args.allow_failed:
        raise SystemExit(f"{len(failed)} jobs failed: {[j['id'] for j in failed]}. Rerun the same command to "
                         "retry them, or add --allow-failed to fill their share from fallbacks.")
    groups = [(f"{n}.jsonl", [j for j in prim if j["name"] == n]) for n in names if shards[n]]

    fallback_log = []
    if not args.only:
        report = global_pass(groups, jobs_dir, out_dir, log, write=False)
        got = Counter()
        for j in prim:
            got[j["name"]] += report.get(j["id"], {}).get("words", 0)
        plan, fallback_log = build.plan_fallbacks(size, [dict(source=n, words_M=got[n] / 1e6) for n in names])
        log(f"fallback plan: {fallback_log}")
        topups = plan_jobs(list(plan.items()), ".topup", args,
                           lambda n: [str(paths(jobs_dir, j)["keys"]) for j in prim if j["name"] == n])
        failed = run_pool(topups, args, size, jobs_dir, log)
        if failed and not args.allow_failed:
            raise SystemExit(f"{len(failed)} top-up jobs failed: {[j['id'] for j in failed]}. Rerun to retry.")
        groups += [(f"{n}.topup.jsonl", [j for j in topups if j["name"] == n]) for n in plan]

    report = global_pass(groups, jobs_dir, out_dir, log, write=True)
    srcs = summarize(groups, jobs_dir, report)
    total = sum(s["words_M"] for s in srcs)
    stats = dict(total_M=round(total, 3), fallbacks=fallback_log, sources=srcs,
                 wall_seconds=round(time.time() - t0), options=vars(args))
    (out_dir / "stats.json").write_text(json.dumps(stats, indent=1))
    log(f"== total {total:.2f}M words in {(time.time() - t0) / 60:.0f} min wall")


if __name__ == "__main__":
    main()