scripts/update_danbooru.py

import argparse
import queue
import threading
import time
from pathlib import Path
from time import sleep

import requests
from _danbooru import DEFAULT_RATE, RateLimiter, get_json
from _env import credentials
from _paths import DANBOORU_DB_PATH
from rich.progress import (
    BarColumn,
    Progress,
    TextColumn,
    TimeElapsedColumn,
    TimeRemainingColumn,
)
from sqlalchemy import text

from db import get_engine, t_posts

url_queue: queue.Queue[str | None] = queue.Queue(maxsize=64)
data_queue: queue.Queue[list[dict[str, object]] | None] = queue.Queue(maxsize=64)
# 彻底抓不到的页。list.append 在 CPython 下是原子的,worker 线程可以直接写。
failures: list[str] = []

parser = argparse.ArgumentParser()
parser.add_argument("--start_id", type=int, default=0)
parser.add_argument("--api_key", type=str)
parser.add_argument("--login", type=str)
parser.add_argument("--workers", type=int, default=16)
parser.add_argument(
    "--rate",
    type=float,
    default=DEFAULT_RATE,
    help=f"起步请求速率(req/s,默认 {DEFAULT_RATE},全部 worker 共享)。撞到 429 会自动减半再慢慢恢复。",
)
parser.add_argument("--db_path", type=str, default=str(DANBOORU_DB_PATH))
parser.add_argument(
    "--follow",
    action="store_true",
    help="追平后不退出,每 2 秒轮询新帖。默认追平即退出 —— 定时任务链需要这个脚本会结束。",
)
parser.add_argument(
    "--update",
    action="store_true",
    help="全量更新模式：从头遍历所有 id 刷新 score/fav_count 等变化的数值，"
    "而非仅抓取新帖。已消失的条目会保留原值，不会从数据库删除。",
)
args = parser.parse_args()

API_KEY, LOGIN = credentials(args.api_key, args.login)
# 一个发牌器管住所有 worker。限速是按账号算的,16 个线程各自限速等于超发 16 倍。
limiter = RateLimiter(args.rate)
# 只回显 login。匿名和已认证的限流差一个数量级,跑之前得能一眼看出走的哪条路。
print(f"syncing posts -- {f'authenticated as {LOGIN}' if LOGIN else 'anonymous (heavier rate limiting)'}")

BASE_API_URL = "https://danbooru.donmai.us/posts.json?page=a{}&limit=200{}"
LATEST_ID_URL = "https://danbooru.donmai.us/posts.json?limit=1{}"

latest_id = 1


def auth_query() -> str:
    # 仅在提供凭据时附加认证参数，否则匿名访问；拼接 api_key=None 会触发 401
    if API_KEY and LOGIN:
        return f"&api_key={API_KEY}&login={LOGIN}"
    return ""


def fetch_latest_id(session: requests.Session) -> None:
    global latest_id
    try:
        rows = get_json(session, LATEST_ID_URL.format(auth_query()), limiter=limiter)
    except RuntimeError as exc:
        # 只是进度条的上界,拿不到就沿用旧值继续跑,不值得为它中断同步。
        print(f"\ncould not refresh the latest id ({exc}); keeping {latest_id}")
        return
    if rows and "id" in rows[0]:
        latest_id = int(rows[0]["id"])


def update_latest_id() -> None:
    session = requests.Session()
    while True:
        fetch_latest_id(session)
        time.sleep(60)


def fetch_data() -> None:
    session = requests.Session()
    while True:
        url = url_queue.get()
        if url is None:
            url_queue.task_done()
            break
        try:
            data_queue.put(get_json(session, url, limiter=limiter))
        except RuntimeError as exc:
            # 记下来,不要跳过就算了。这一页是 200 条投稿,少了在成品里完全看不出来,
            # 而增量模式下次从数据库最大 id 起跑,永远不会回头补这个洞。
            failures.append(f"{url}: {exc}")
            print(f"\nGAVE UP on {url}: {exc}")
        url_queue.task_done()


def save_data(engine_path: str, required_columns: list[str]) -> None:
    engine = get_engine(engine_path)
    while True:
        data = data_queue.get()
        if data is None:
            data_queue.task_done()
            break
        for column in required_columns:
            for row in data:
                if column not in row:
                    row[column] = None
        with engine.connect() as conn:
            conn.execute(t_posts.insert().prefix_with("OR REPLACE"), data)
            conn.commit()
        data_queue.task_done()


def drain(workers: list[threading.Thread]) -> None:
    """让在途的页真正落盘,再退出。

    这些线程是 daemon,进程一退就地消失。生产者停下时队列里通常还压着几十页 ——
    直接返回等于把它们扔掉,而且扔得悄无声息:日志上看是一次干净的同步,数据库里
    少了几千条。所以按 抓取 -> 保存 的顺序逐级投毒丸,每级都等它自己排空。
    """
    for _ in workers:
        url_queue.put(None)
    for worker in workers:
        worker.join()
    data_queue.put(None)
    data_queue.join()


def main() -> None:
    engine_path = Path(args.db_path)
    if engine_path.parent != Path("."):
        engine_path.parent.mkdir(parents=True, exist_ok=True)
    threads: list[threading.Thread] = []
    for _ in range(args.workers):
        thread = threading.Thread(target=fetch_data, daemon=True)
        thread.start()
        threads.append(thread)

    # 先同步取一次,再决定要不要开轮询线程。latest_id 初值是 1,交给后台线程去填
    # 就有竞态:非 follow 模式下 current_id(约 1190 万) >= 1 立刻成立,一条不抓就退出。
    fetch_latest_id(requests.Session())
    if args.follow:
        # 只有常驻模式才需要它。一次性同步对着固定快照跑,结果可复现。
        threading.Thread(target=update_latest_id, daemon=True).start()

    required_columns = [col.name for col in t_posts.columns if not col.primary_key]
    danbooru_engine = get_engine(str(engine_path))
    t_posts.metadata.create_all(danbooru_engine, tables=[t_posts])

    with danbooru_engine.connect() as conn:
        res = conn.execute(text("SELECT MAX(id) FROM posts"))
        row = res.fetchone()
        latest_db_id = row[0] if row and row[0] else 1

    # 起点优先级：显式 --start_id（续传） > --update（从头全量刷新） > 增量（数据库最大 id）
    if args.start_id != 0:
        start_id = args.start_id
    elif args.update:
        start_id = 1
    else:
        start_id = latest_db_id
    print("mode:", "update" if args.update else "incremental")
    print("start_id:", start_id)

    save_thread = threading.Thread(
        target=save_data,
        args=(str(engine_path), required_columns),
        daemon=True,
    )
    save_thread.start()

    try:
        with Progress(
            TextColumn("[progress.description]{task.description}"),
            BarColumn(),
            TextColumn("[progress.percentage]{task.percentage:>3.0f}%"),
            TimeElapsedColumn(),
            TimeRemainingColumn(),
            TextColumn("[{task.completed} / {task.total}]"),
        ) as progress:
            total_items = max(latest_id - start_id + 1, 0)
            task = progress.add_task("Processing", start=True, total=total_items)
            current_id = start_id

            while True:
                if current_id >= latest_id:
                    progress.update(task, completed=latest_id)
                    if not args.follow:
                        break
                    sleep(2)
                    continue

                url = BASE_API_URL.format(current_id, auth_query())
                # 不在这里 sleep:队列 maxsize=64 已经是背压,真正的节奏由 limiter
                # 决定。producer 再睡一次只会和 limiter 叠加成一个谁也说不清的速率。
                url_queue.put(url)
                current_id += 200
                progress.update(task, completed=current_id, total=latest_id)
    except KeyboardInterrupt:
        print("Stopped by user")
    drain(threads)
    with danbooru_engine.connect() as conn:
        rows = conn.execute(text("SELECT COUNT(*) FROM posts")).scalar()
        newest = conn.execute(text("SELECT MAX(id) FROM posts")).scalar()
    print(f"posts: {rows:,} rows, max id {newest:,} (site was at {latest_id:,})")
    print(f"final rate: {limiter.rate:.2f} req/s (started at {args.rate:g})")
    if failures:
        # 非零退出码:定时任务链上游得知道这次存档是有洞的。
        print(f"\n{len(failures)} page(s) were never fetched -- the archive has holes:")
        for failure in failures[:20]:
            print(f"  {failure}")
        if len(failures) > 20:
            print(f"  ... and {len(failures) - 20} more")
        print("Re-run with --start_id set below the lowest missing id to fill them.")
        raise SystemExit(1)
    raise SystemExit(0)


if __name__ == "__main__":
    main()