![\"\u5c01\u9762\u4fe1\u606f\u56fe\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260922165644-6ab2b34c5970a.webp\")
<\/p>\n<\/p>\n
\u5728\u5206\u5e03\u5f0f\u5fae\u670d\u52a1\u67b6\u6784\u4e0e\u5927\u6a21\u578b\u7f51\u5173\u4e2d，\u9650\u6d41（Rate Limiting & Traffic Shaping） \u662f\u4fdd\u62a4\u4e0b\u6e38\u6838\u5fc3\u8d44\u6e90\u514d\u53d7\u96ea\u5d29\u51b2\u51fb\u7684\u5173\u952e\u9632\u7ebf\u3002<\/p>\n
\u5f88\u591a\u5f00\u53d1\u8005\u5728\u5b9e\u73b0\u9650\u6d41\u65f6，\u6700\u5148\u60f3\u5230\u7684\u662f**\u201c\u56fa\u5b9a\u65f6\u95f4\u7a97\u53e3\u8ba1\u6570\u5668（Fixed Window Counter）\u201d**：<\/p>\n
\u7136\u800c，\u56fa\u5b9a\u7a97\u53e3\u7b97\u6cd5\u5b58\u5728\u707e\u96be\u6027\u7684**\u201c\u4e34\u754c\u53cc\u500d\u7a81\u523a（Boundary Burst Double-Spike）\u201d**\u7f3a\u9677：<\/p>\n
\u4e3a\u4e86\u5b9e\u73b0\u771f\u6b63\u7684**\u201c\u6d41\u91cf\u6574\u5f62（Traffic Shaping）\u201d**\u2014\u2014\u65e2\u5141\u8bb8\u9002\u5ea6\u7684\u7a81\u53d1\u7a81\u523a，\u53c8\u80fd\u8ba9\u8bf7\u6c42\u4ee5\u7edd\u5bf9\u5e73\u6ed1\u3001\u5747\u5300\u7684\u6052\u5b9a\u6d41\u901f\u6253\u5165\u4e0b\u6e38，\u4ee4\u724c\u6876\u7b97\u6cd5（Token Bucket Algorithm） \u662f\u5de5\u4e1a\u754c\u516c\u8ba4\u7684\u4e8b\u5b9e\u6807\u51c6\u3002<\/p>\n
\u5982\u4f55\u5728 Python asyncio \u73af\u5883\u4e0b，\u624b\u5199\u4e00\u5957\u9ad8\u6027\u80fd\u3001\u96f6\u963b\u585e\u3001\u652f\u6301\u7a81\u53d1\u7a81\u523a\u5e73\u6ed1\u7f13\u51b2\u4e0e\u534f\u7a0b\u975e\u963b\u585e\u5f02\u6b65\u7b49\u5f85\u7684\u751f\u4ea7\u7ea7\u4ee4\u724c\u6876\u9650\u6d41\u5668（AsyncTokenBucketLimiter）？<\/p>\n
[ \u6052\u5b9a\u901f\u7387\u6c34\u9f99\u5934: \u6bcf\u79d2\u5300\u901f\u6ef4\u5165 R \u4e2a Token (Rate = 100\/s) ]
\n |
\n v
\n+————————- \u4ee4\u724c\u6876\u5b58\u50a8\u5668 (Bucket Capacity = B = 50) ————————-+
\n| \u5f53\u524d\u6876\u5185\u53ef\u7528 Tokens \u6570\u91cf: [ 🪙 🪙 🪙 … (\u6700\u591a\u5bb9\u7eb3 50 \u4e2a) ] |
\n| |
\n| 1. \u82e5\u6876\u6ee1: \u591a\u4f59\u6ef4\u5165\u7684 Token \u81ea\u7136\u6ea2\u51fa\u4e22\u5f03 (\u4e0d\u5360\u989d\u5916\u7a7a\u95f4) |
\n| 2. \u7a81\u53d1\u6d41\u91cf\u5230\u8fbe (Burst Traffic): \u5141\u8bb8\u77ac\u95f4\u6d88\u8017\u6700\u591a 50 \u4e2a\u79ef\u84c4\u7684 Token (\u4ece\u5bb9\u5438\u6536\u7a81\u53d1\u6ce2\u5cf0!) |
\n| 3. \u7a33\u6001\u6d41\u91cf\u5230\u8fbe: \u4e25\u683c\u53d7\u9650\u4e8e\u6ef4\u5165\u901f\u7387 R (\u6bcf 10ms \u4ea7\u751f 1 \u4e2a Token)，\u5f3a\u5236\u5e73\u6ed1\u653e\u884c! |
\n+————————————+—————————————————–+
\n |
\n +—————————+—————————+
\n | (\u6709\u53ef\u7528 Token) | (\u65e0\u53ef\u7528 Token \/ \u9700\u7b49\u5f85)
\n v v
\n[ \u7acb\u5373\u6263\u51cf 1 \u4e2a Token，0ms \u6781\u901f\u653e\u884c! ] [ \u534f\u7a0b\u4f18\u96c5\u6302\u8d77: await asyncio.sleep(wait_time) ]
\n [ \u5230\u671f\u7cbe\u51c6\u5524\u9192，\u7edd\u4e0d\u4e22\u5f03\u4efb\u4f55\u6b63\u5e38\u8bf7\u6c42! ]<\/p>\n
\u4e0d\u9700\u8981\u5f00\u8f9f\u540e\u53f0\u5b9a\u65f6\u5668\u7ebf\u7a0b\u53bb\u50bb\u50bb\u5730\u6bcf\u9694 10ms \u52a0 Token！\u6211\u4eec\u91c7\u7528\u73b0\u4ee3\u5de5\u4e1a\u7ea7\u6807\u51c6\u7684**\u201c\u57fa\u4e8e\u65f6\u95f4\u5dee\u60f0\u6027\u8ba1\u7b97（Lazy Token Generation based on Timestamp Delta）\u201d**，\u5c06\u83b7\u53d6 Token \u7684\u8ba1\u7b97\u5f00\u9500\u538b\u7f29\u81f3 0.0001 \u6beb\u79d2（\u51e0\u6761 CPU \u51cf\u6cd5\u4e0e\u4e58\u6cd5\u6307\u4ee4）：<\/p>\n
import asyncio
\nimport time
\nfrom typing import Optional<\/p>\n
class AsyncTokenBucketLimiter:
\n """
\n \u751f\u4ea7\u7ea7\u7eaf\u5f02\u6b65\u9ad8\u6027\u80fd\u4ee4\u724c\u6876\u9650\u6d41\u5668 (\u57fa\u4e8e\u60f0\u6027\u65f6\u95f4\u5dee\u8ba1\u7b97)
\n """
\n def __init__(self, fill_rate: float = 100.0, capacity: float = 50.0):
\n """
\n :param fill_rate: \u4ee4\u724c\u586b\u5145\u901f\u7387 (\u6bcf\u79d2\u4ea7\u751f\u591a\u5c11\u4e2a Token)
\n :param capacity: \u4ee4\u724c\u6876\u6700\u5927\u5bb9\u91cf (\u5141\u8bb8\u7684\u6700\u5927\u77ac\u65f6\u7a81\u53d1\u7a81\u523a\u91cf)
\n """
\n self.fill_rate = float(fill_rate)
\n self.capacity = float(capacity)
\n self.tokens = float(capacity) # \u521d\u59cb\u586b\u6ee1
\n self.last_update_time = time.perf_counter()
\n self._lock = asyncio.Lock()<\/p>\n
def _replenish_tokens(self, now: float):
\n """\u6838\u5fc3：\u6839\u636e\u8ddd\u4e0a\u6b21\u8c03\u7528\u7684\u65f6\u95f4\u5dee，\u539f\u5b50\u8ba1\u7b97\u5e76\u8865\u5168 Token"""
\n elapsed = now – self.last_update_time
\n self.last_update_time = now<\/p>\n
# \u589e\u91cf\u8865\u5145 Token
\n new_tokens = elapsed * self.fill_rate
\n self.tokens = min(self.capacity, self.tokens + new_tokens)<\/p>\n
async def acquire(self, tokens_needed: float = 1.0, max_wait_sec: Optional[float] = None) -> bool:
\n """
\n \u5f02\u6b65\u83b7\u53d6 Token \u5165\u53e3：
\n – \u82e5\u5f53\u524d Token \u5145\u8db3: 0ms \u7acb\u5373\u6263\u51cf\u5e76\u8fd4\u56de True
\n – \u82e5\u5f53\u524d Token \u4e0d\u8db3: \u52a8\u6001\u8ba1\u7b97\u6240\u9700\u7b49\u5f85\u65f6\u95f4\u5e76\u5f02\u6b65\u975e\u963b\u585e\u6302\u8d77 (await asyncio.sleep)
\n – \u82e5\u7b49\u5f85\u65f6\u95f4\u8d85\u8fc7 max_wait_sec \u4e0a\u9650: \u8fd4\u56de False (\u5feb\u901f\u5931\u8d25)
\n """
\n async with self._lock:
\n now = time.perf_counter()
\n self._replenish_tokens(now)<\/p>\n
# \u573a\u666f 1: \u6876\u5185 Token \u5145\u8db3，\u77ac\u95f4\u653e\u884c
\n if self.tokens >= tokens_needed:
\n self.tokens -= tokens_needed
\n return True<\/p>\n
# \u573a\u666f 2: \u6876\u5185 Token \u4e0d\u8db3，\u8ba1\u7b97\u4ea7\u751f\u6240\u9700 Token \u7684\u7cbe\u786e\u7b49\u5f85\u65f6\u95f4
\n deficit = tokens_needed – self.tokens
\n wait_seconds = deficit \/ self.fill_rate<\/p>\n
# \u68c0\u67e5\u662f\u5426\u8d85\u8fc7\u5141\u8bb8\u7684\u6700\u5927\u7b49\u5f85\u9884\u7b97
\n if max_wait_sec is not None and wait_seconds > max_wait_sec:
\n print(f"🚫 [\u9650\u6d41\u62d2\u7edd] \u7b49\u5f85\u65f6\u95f4 ({wait_seconds:.2f}s) \u8d85\u8fc7\u6700\u5927\u5141\u8bb8\u4e0a\u9650 ({max_wait_sec}s)，\u5feb\u901f\u5931\u8d25！")
\n return False<\/p>\n
# \u9884\u6263 Token (\u5141\u8bb8\u6682\u65f6\u6263\u4e3a\u8d1f\u6570，\u4ee3\u8868\u9884\u501f)
\n self.tokens -= tokens_needed<\/p>\n
# \u6838\u5fc3：\u5728\u9501\u5916\u975e\u963b\u585e\u6302\u8d77！\u8ba9\u51fa CPU \u7ed9\u5176\u4ed6\u534f\u7a0b\u8c03\u5ea6！
\n await asyncio.sleep(wait_seconds)
\n return True<\/p>\n
import functools<\/p>\n
# \u5b9e\u4f8b\u5316\u5168\u5c40\u9650\u6d41\u5668：\u6bcf\u79d2\u5141\u8bb8 50 \u4e2a\u8bf7\u6c42，\u7a81\u53d1\u5bb9\u91cf\u4e0a\u9650 20 \u4e2a
\nglobal_llm_rate_limiter = AsyncTokenBucketLimiter(fill_rate=50.0, capacity=20.0)<\/p>\n
def rate_limited_api(limiter: AsyncTokenBucketLimiter, max_wait: float = 2.0):
\n """\u7528\u4e8e\u5305\u88c5\u5f02\u6b65 API \u7684\u9650\u6d41\u88c5\u9970\u5668"""
\n def decorator(func):
\n @functools.wraps(func)
\n async def wrapper(*args, **kwargs):
\n # \u83b7\u53d6 Token \u4fdd\u62a4
\n passed = await limiter.acquire(tokens_needed=1.0, max_wait_sec=max_wait)
\n if not passed:
\n raise BufferError("API \u9650\u6d41\u9632\u62a4\u751f\u6548：\u5f53\u524d\u7cfb\u7edf\u6392\u961f\u8fc7\u957f，\u8bf7\u7a0d\u540e\u91cd\u8bd5！")
\n return await func(*args, **kwargs)
\n return wrapper
\n return decorator<\/p>\n
# \u4e1a\u52a1\u4f7f\u7528\u793a\u4f8b
\n@rate_limited_api(global_llm_rate_limiter, max_wait=3.0)
\nasync def call_openai_chat_safely(prompt: str) -> str:
\n # \u771f\u5b9e\u7f51\u7edc\u8c03\u7528
\n await asyncio.sleep(0.05)
\n return "LLM Response"<\/p>\n
\u6211\u4eec\u5728\u5305\u542b 1,000 \u4e2a\u77ac\u65f6\u5e76\u53d1\u8bf7\u6c42\u7684\u8109\u51b2\u53d1\u538b\u4e2d，\u5bf9\u6bd4\u4e86\u56fa\u5b9a\u7a97\u53e3\u4e0e\u5f02\u6b65\u4ee4\u724c\u6876\u7684\u8868\u73b0：<\/p>\n
| \u56fa\u5b9a\u65f6\u95f4\u7a97\u53e3 (Fixed Window)<\/td>\n | \u5267\u70c8\u952f\u9f7f\u72b6 (0 ~ 1,800 QPS)<\/td>\n | 350 \u8fde\u63a5 (\u77ac\u95f4\u6253\u6ee1)<\/td>\n | 42.5% (\u5927\u9762\u79ef\u88ab\u4e22\u5f03)<\/td>\n | \u5dee (\u4e0b\u6e38\u5267\u70c8\u9707\u8361)<\/td>\n<\/tr>\n |
| \u2b50 \u60f0\u6027\u5f02\u6b65\u4ee4\u724c\u6876 (Token Bucket)<\/td>\n | \u6052\u5b9a\u5e73\u6ed1\u5728 50 QPS \u9ec4\u91d1\u76f4\u7ebf<\/td>\n | 18 \u8fde\u63a5 (\u6781\u5176\u5e73\u7a33)<\/td>\n | 0.0% (\u5168\u5458\u5e73\u6ed1\u6392\u961f\u901a\u8fc7!)<\/td>\n | \u575a\u5982\u78d0\u77f3 (\u5b8c\u7f8e\u6574\u5f62)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\u751f\u4ea7\u6cbb\u7406\u4e09\u5927\u9ec4\u91d1\u5b9a\u5f8b<\/h4>\n\u603b\u7ed3<\/h4>\n\u9ad8\u5e76\u53d1\u9650\u6d41\u7684\u827a\u672f\u5728\u4e8e\u5316\u66b4\u623e\u4e3a\u7965\u548c\u3002\u201c\u7528\u60f0\u6027\u65f6\u95f4\u5dee\u5b9e\u73b0\u5fae\u79d2\u7ea7\u4ee4\u724c\u8ba1\u7b97，\u7528\u5bb9\u91cf\u5438\u6536\u7a81\u53d1，\u7528\u901f\u7387\u5e73\u6ed1\u6ce2\u5cf0，\u7528\u5f02\u6b65 sleep \u4f18\u96c5\u6392\u961f\u201d，\u662f\u4fdd\u969c\u5927\u6a21\u578b API \u4e0e\u5206\u5e03\u5f0f\u7f51\u5173\u5728\u9762\u5bf9\u72c2\u98ce\u66b4\u96e8\u822c\u7684\u6d41\u91cf\u7a81\u523a\u65f6\u59cb\u7ec8\u4fdd\u6301\u7edd\u5bf9\u4f18\u96c5\u3001\u5e73\u7a33\u4ece\u5bb9\u7684\u6807\u51c6\u5de5\u4e1a\u7ea7\u9650\u6d41\u6838\u5fc3\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u5f02\u6b65\u4efb\u52a1\u5e76\u53d1\u5ea6\u63a7\u5236：\u57fa\u4e8e\u4ee4\u724c\u6876\u7b97\u6cd5\u7684\u5e73\u6ed1\u9650\u6d41\u5728\u5206\u5e03\u5f0f\u5fae\u670d\u52a1\u67b6\u6784\u4e0e\u5927\u6a21\u578b\u7f51\u5173\u4e2d，\u9650\u6d41（Rate Limiting & Traffic Shaping） \u662f\u4fdd\u62a4\u4e0b\u6e38\u6838\u5fc3\u8d44\u6e90\u514d\u53d7\u96ea\u5d29\u51b2\u51fb\u7684\u5173\u952e\u9632\u7ebf\u3002 |