批量验证码流式处理：Python 异步生成器实战

批量提交 500 个验证码任务后，收尾动作不该是 await asyncio.gather(*tasks) 等一个总结果，而是让每条 token 到手就流向下游。同一批任务耗时不齐：Turnstile 上限 <10 秒，reCAPTCHA v2 上限 <60 秒——等最慢的那条，先完成的几百条 token 只能空耗有效期。

先算一笔账：三种消费方式的取舍

方案	首条可用时间	内存	下游延迟
等全部完成	最慢任务返回后	全部驻留	高
边识别边产出	最快任务返回后	只持有一条	低
微批（10 条）	第一组凑齐后	持有 10 条	中

这批任务到底该不该流式处理

场景	建议做法
拿到 token 就提交表单	流式
全量结果导出成 CSV	收齐再写
任务之间有先后依赖	收齐按序处理
单批 1,000 条以上	流式，压住内存峰值

Python：用异步生成器边识别边产出

asyncio 加 aiohttp 最自然：谁先识别完谁先被 yield 出来，调用方一个 async for 就能消费；Semaphore 控住在途任务数，asyncio.wait(..., return_when=FIRST_COMPLETED) 负责唤醒完成的那条：

import asyncio
import aiohttp
import time

API_KEY = "YOUR_API_KEY"
SUBMIT_URL = "https://ocr.captchaai.com/in.php"
RESULT_URL = "https://ocr.captchaai.com/res.php"


async def submit_task(session, task_data):
    """Submit a single CAPTCHA task."""
    params = {
        "key": API_KEY,
        "method": task_data.get("method", "userrecaptcha"),
        "json": 1,
    }
    if params["method"] == "userrecaptcha":
        params["googlekey"] = task_data["sitekey"]
        params["pageurl"] = task_data["pageurl"]
    elif params["method"] == "turnstile":
        params["sitekey"] = task_data["sitekey"]
        params["pageurl"] = task_data["pageurl"]

    async with session.post(SUBMIT_URL, data=params) as resp:
        result = await resp.json(content_type=None)
        if result.get("status") != 1:
            return None, result.get("request", "unknown")
        return result["request"], None


async def poll_task(session, task_id, timeout=300):
    """Poll until solved or timeout."""
    start = time.monotonic()
    while time.monotonic() - start < timeout:
        await asyncio.sleep(5)
        params = {"key": API_KEY, "action": "get", "id": task_id, "json": 1}
        async with session.get(RESULT_URL, params=params) as resp:
            result = await resp.json(content_type=None)

        if result.get("request") == "CAPCHA_NOT_READY":
            continue
        if result.get("status") == 1:
            return result["request"], None
        return None, result.get("request", "unknown")

    return None, "TIMEOUT"


async def solve_one(session, index, task_data, semaphore):
    """Solve a single task within concurrency limits."""
    async with semaphore:
        start = time.monotonic()
        task_id, error = await submit_task(session, task_data)
        if error:
            return {"index": index, "status": "failed", "error": error, "time": 0}

        token, error = await poll_task(session, task_id)
        elapsed = time.monotonic() - start

        if token:
            return {"index": index, "status": "solved", "token": token, "time": round(elapsed, 1)}
        return {"index": index, "status": "failed", "error": error, "time": round(elapsed, 1)}


async def stream_results(tasks, max_concurrent=20):
    """
    Async generator that yields each result as it completes.
    Results arrive in completion order, not submission order.
    """
    semaphore = asyncio.Semaphore(max_concurrent)

    async with aiohttp.ClientSession() as session:
        pending = set()
        for i, task in enumerate(tasks):
            coro = solve_one(session, i, task, semaphore)
            pending.add(asyncio.ensure_future(coro))

        while pending:
            done, pending = await asyncio.wait(pending, return_when=asyncio.FIRST_COMPLETED)
            for future in done:
                yield future.result()


async def main():
    tasks = [
        {"sitekey": "SITE_KEY", "pageurl": f"https://example.com/page{i}"}
        for i in range(50)
    ]

    solved = 0
    failed = 0

    async for result in stream_results(tasks, max_concurrent=15):
        # Process each result immediately
        if result["status"] == "solved":
            solved += 1
            print(f"  [{solved + failed}/{len(tasks)}] Task {result['index']} SOLVED in {result['time']}s")

            # Use token immediately — don't wait for batch
            # await submit_form(result["token"])
            # await save_to_database(result)
        else:
            failed += 1
            print(f"  [{solved + failed}/{len(tasks)}] Task {result['index']} FAILED: {result['error']}")

    print(f"\nDone: {solved} solved, {failed} failed")


asyncio.run(main())

两个细节：

结果按完成顺序到达，还原顺序靠每条自带的 index。
timeout=300 不能省，否则 pending 清不空，生成器会挂死。

安装依赖：

pip install aiohttp

国内可加 -i https://pypi.tuna.tsinghua.edu.cn/simple 走镜像。

Node.js：用 EventEmitter 把结果广播给下游

事件模型本身就是流：CaptchaStream 继承 EventEmitter，识别完一条就 emit("result")，全部结束再 emit("done")：

const { EventEmitter } = require("events");

const API_KEY = "YOUR_API_KEY";
const SUBMIT_URL = "https://ocr.captchaai.com/in.php";
const RESULT_URL = "https://ocr.captchaai.com/res.php";

class CaptchaStream extends EventEmitter {
  constructor(maxConcurrent = 15) {
    super();
    this.maxConcurrent = maxConcurrent;
    this.active = 0;
    this.queue = [];
    this.total = 0;
    this.completed = 0;
  }

  async submitAndPoll(index, taskData) {
    const params = new URLSearchParams({
      key: API_KEY,
      method: taskData.method || "userrecaptcha",
      googlekey: taskData.sitekey,
      pageurl: taskData.pageurl,
      json: "1",
    });

    const start = Date.now();
    const submitResp = await (await fetch(SUBMIT_URL, { method: "POST", body: params })).json();

    if (submitResp.status !== 1) {
      return { index, status: "failed", error: submitResp.request, time: 0 };
    }

    const taskId = submitResp.request;
    for (let i = 0; i < 60; i++) {
      await new Promise((r) => setTimeout(r, 5000));
      const url = `${RESULT_URL}?key=${API_KEY}&action=get&id=${taskId}&json=1`;
      const poll = await (await fetch(url)).json();

      if (poll.request === "CAPCHA_NOT_READY") continue;
      const elapsed = ((Date.now() - start) / 1000).toFixed(1);
      if (poll.status === 1) return { index, status: "solved", token: poll.request, time: elapsed };
      return { index, status: "failed", error: poll.request, time: elapsed };
    }
    return { index, status: "failed", error: "TIMEOUT", time: ((Date.now() - start) / 1000).toFixed(1) };
  }

  async processNext() {
    if (this.queue.length === 0 || this.active >= this.maxConcurrent) return;

    const { index, taskData } = this.queue.shift();
    this.active++;

    try {
      const result = await this.submitAndPoll(index, taskData);
      this.emit("result", result);
    } catch (err) {
      this.emit("result", { index, status: "failed", error: err.message });
    } finally {
      this.active--;
      this.completed++;

      if (this.completed === this.total) {
        this.emit("done");
      } else {
        this.processNext();
      }
    }
  }

  start(tasks) {
    this.total = tasks.length;
    this.queue = tasks.map((taskData, index) => ({ index, taskData }));

    // Launch initial batch
    const initial = Math.min(this.maxConcurrent, tasks.length);
    for (let i = 0; i < initial; i++) {
      this.processNext();
    }
    return this;
  }
}

// Usage
const tasks = Array.from({ length: 50 }, (_, i) => ({
  sitekey: "SITE_KEY",
  pageurl: `https://example.com/page${i}`,
}));

const stream = new CaptchaStream(15);
let solved = 0, failed = 0;

stream.on("result", (result) => {
  if (result.status === "solved") {
    solved++;
    console.log(`[${solved + failed}/${tasks.length}] Task ${result.index} SOLVED (${result.time}s)`);
    // Use token immediately
    // submitForm(result.token);
  } else {
    failed++;
    console.log(`[${solved + failed}/${tasks.length}] Task ${result.index} FAILED: ${result.error}`);
  }
});

stream.on("done", () => {
  console.log(`\nComplete: ${solved} solved, ${failed} failed`);
});

stream.start(tasks);

监听器里只做轻量动作，写库、提交表单再丢一层队列——慢监听器会拖住整条回调链。

并发数怎么定：线程数就是流的宽度

max_concurrent 应当等于套餐线程数，套餐内识别次数不限：

BASIC（$15/月）5 线程
STANDARD（$30/月）15 线程，即示例里的值
ADVANCE（$90/月）50 线程

设高了只是在服务端排队，首条反而更晚。吞吐还取决于类型：Turnstile 上限 <10 秒，15 线程一分钟至少过 90 条；reCAPTCHA v2 上限 <60 秒，保底约 15 条。国内站点更多是 GeeTest（极验）滑块，同一套框架通用，区别只在 method——CaptchaAI 支持 GeeTest v3。