implementing-api-rate-limiting-and-throttling

实现 API 限流与节流

适用场景

• 保护身份验证端点，防范暴力破解和凭据填充攻击
• 防止自动化脚本和机器人滥用 API 并耗尽资源
• 为不同 API 使用方层级（免费、高级、企业）实施公平的使用配额
• 防御应用层拒绝服务攻击
• 满足要求实施 API 滥用防护措施的合规要求

不要仅将限流作为抵御攻击的唯一防线。应结合身份验证、授权和 WAF 规则。

前提条件

• Redis 6.0+，用于分布式限流计数器（单实例部署也可使用内存）
• API 框架（Express.js、FastAPI、Spring Boot 或 Django REST Framework）
• 用于监控限流指标的系统（Prometheus、CloudWatch、Datadog）
• 了解 API 的正常流量模式和峰值使用情况
• 用于验证限流行为的负载测试工具（k6、Gatling 或 Locust）

工作流

步骤 1：设计限流策略

按端点类别和用户层级定义限流值：

# Rate limit configuration
RATE_LIMITS = {
    # Authentication endpoints (most restrictive)
    "auth": {
        "login": {"requests": 5, "window_seconds": 60, "by": "ip"},
        "register": {"requests": 3, "window_seconds": 300, "by": "ip"},
        "forgot_password": {"requests": 3, "window_seconds": 3600, "by": "ip"},
        "verify_mfa": {"requests": 5, "window_seconds": 300, "by": "user"},
    },
    # Standard API endpoints
    "api": {
        "free": {"requests": 60, "window_seconds": 60, "by": "user"},
        "premium": {"requests": 300, "window_seconds": 60, "by": "user"},
        "enterprise": {"requests": 1000, "window_seconds": 60, "by": "user"},
    },
    # Resource-intensive endpoints
    "expensive": {
        "search": {"requests": 10, "window_seconds": 60, "by": "user"},
        "export": {"requests": 5, "window_seconds": 3600, "by": "user"},
        "bulk_import": {"requests": 2, "window_seconds": 3600, "by": "user"},
    },
    # Global limits
    "global": {
        "per_ip": {"requests": 1000, "window_seconds": 60, "by": "ip"},
        "per_user": {"requests": 5000, "window_seconds": 3600, "by": "user"},
    },
}

步骤 2：滑动窗口限流器（Redis）

import redis
import time
import hashlib
from functools import wraps
from flask import Flask, request, jsonify, g

app = Flask(__name__)
redis_client = redis.Redis(host='localhost', port=6379, db=0, decode_responses=True)

class SlidingWindowRateLimiter:
    """Sliding window rate limiter using Redis sorted sets."""

    def __init__(self, redis_conn):
        self.redis = redis_conn

    def is_allowed(self, key, max_requests, window_seconds):
        """Check if request is allowed and record it."""
        now = time.time()
        window_start = now - window_seconds
        pipe = self.redis.pipeline()

        # Remove expired entries
        pipe.zremrangebyscore(key, 0, window_start)
        # Count requests in current window
        pipe.zcard(key)
        # Add current request
        pipe.zadd(key, {f"{now}:{hashlib.md5(str(now).encode()).hexdigest()[:8]}": now})
        # Set TTL on the key
        pipe.expire(key, window_seconds + 1)

        results = pipe.execute()
        current_count = results[1]

        if current_count >= max_requests:
            # Calculate retry-after
            oldest = self.redis.zrange(key, 0, 0, withscores=True)
            if oldest:
                retry_after = int(oldest[0][1] + window_seconds - now) + 1
            else:
                retry_after = window_seconds
            return False, current_count, max_requests, retry_after

        return True, current_count + 1, max_requests, 0

rate_limiter = SlidingWindowRateLimiter(redis_client)

def rate_limit(max_requests, window_seconds, key_func=None):
    """Decorator for rate limiting API endpoints."""
    def decorator(f):
        @wraps(f)
        def wrapped(*args, **kwargs):
            # Determine the rate limit key
            if key_func:
                identifier = key_func()
            elif hasattr(g, 'user_id'):
                identifier = f"user:{g.user_id}"
            else:
                identifier = f"ip:{request.remote_addr}"

            key = f"ratelimit:{request.endpoint}:{identifier}"
            allowed, current, limit, retry_after = rate_limiter.is_allowed(
                key, max_requests, window_seconds)

            # Always set rate limit headers
            headers = {
                "X-RateLimit-Limit": str(limit),
                "X-RateLimit-Remaining": str(max(0, limit - current)),
                "X-RateLimit-Reset": str(int(time.time()) + window_seconds),
            }

            if not allowed:
                headers["Retry-After"] = str(retry_after)
                response = jsonify({
                    "error": "rate_limit_exceeded",
                    "message": "Too many requests. Please try again later.",
                    "retry_after": retry_after
                })
                response.status_code = 429
                for h, v in headers.items():
                    response.headers[h] = v
                return response

            response = f(*args, **kwargs)
            for h, v in headers.items():
                response.headers[h] = v
            return response
        return wrapped
    return decorator

# Apply rate limiting to endpoints
@app.route('/api/v1/auth/login', methods=['POST'])
@rate_limit(max_requests=5, window_seconds=60,
            key_func=lambda: f"ip:{request.remote_addr}")
def login():
    # Login logic
    return jsonify({"message": "Login successful"})

@app.route('/api/v1/users/me', methods=['GET'])
@rate_limit(max_requests=60, window_seconds=60)
def get_profile():
    # Profile logic
    return jsonify({"user": "data"})

@app.route('/api/v1/search', methods=['GET'])
@rate_limit(max_requests=10, window_seconds=60)
def search():
    # Search logic
    return jsonify({"results": []})

步骤 3：令牌桶限流器

import redis
import time

class TokenBucketRateLimiter:
    """Token bucket rate limiter allowing burst traffic within limits."""

    def __init__(self, redis_conn):
        self.redis = redis_conn

    def is_allowed(self, key, max_tokens, refill_rate, refill_interval=1):
        """
        Token bucket algorithm:
        - max_tokens: Maximum burst capacity
        - refill_rate: Tokens added per refill_interval
        - refill_interval: Seconds between refills
        """
        now = time.time()
        bucket_key = f"tb:{key}"

        # Lua script for atomic token bucket operation
        lua_script = """
        local key = KEYS[1]
        local max_tokens = tonumber(ARGV[1])
        local refill_rate = tonumber(ARGV[2])
        local refill_interval = tonumber(ARGV[3])
        local now = tonumber(ARGV[4])

        local bucket = redis.call('hmget', key, 'tokens', 'last_refill')
        local tokens = tonumber(bucket[1])
        local last_refill = tonumber(bucket[2])

        if tokens == nil then
            tokens = max_tokens
            last_refill = now
        end

        -- Refill tokens
        local elapsed = now - last_refill
        local refills = math.floor(elapsed / refill_interval)
        if refills > 0 then
            tokens = math.min(max_tokens, tokens + (refills * refill_rate))
            last_refill = last_refill + (refills * refill_interval)
        end

        local allowed = 0
        if tokens >= 1 then
            tokens = tokens - 1
            allowed = 1
        end

        redis.call('hmset', key, 'tokens', tokens, 'last_refill', last_refill)
        redis.call('expire', key, math.ceil(max_tokens / refill_rate * refill_interval) + 10)

        return {allowed, tokens, max_tokens}
        """

        result = self.redis.eval(lua_script, 1, bucket_key,
                                  max_tokens, refill_rate, refill_interval, now)
        allowed = bool(result[0])
        remaining = int(result[1])
        limit = int(result[2])

        return allowed, remaining, limit

步骤 4：按用户方案分层限流

from enum import Enum

class UserTier(Enum):
    FREE = "free"
    PREMIUM = "premium"
    ENTERPRISE = "enterprise"

TIER_LIMITS = {
    UserTier.FREE: {
        "default": (60, 60),          # 60 req/min
        "search": (10, 60),           # 10 req/min
        "export": (5, 3600),          # 5 req/hour
        "daily_total": (1000, 86400), # 1000 req/day
    },
    UserTier.PREMIUM: {
        "default": (300, 60),
        "search": (50, 60),
        "export": (20, 3600),
        "daily_total": (10000, 86400),
    },
    UserTier.ENTERPRISE: {
        "default": (1000, 60),
        "search": (200, 60),
        "export": (100, 3600),
        "daily_total": (100000, 86400),
    },
}

def get_rate_limit_for_request(user_tier, endpoint_category="default"):
    """Get rate limit configuration based on user tier and endpoint."""
    tier_config = TIER_LIMITS.get(user_tier, TIER_LIMITS[UserTier.FREE])
    limit_config = tier_config.get(endpoint_category, tier_config["default"])
    return limit_config  # (max_requests, window_seconds)

class TieredRateLimitMiddleware:
    """Middleware that applies rate limits based on user subscription tier."""

    def __init__(self, app, redis_conn):
        self.app = app
        self.limiter = SlidingWindowRateLimiter(redis_conn)

    def __call__(self, environ, start_response):
        # Extract user info from request
        user_id = environ.get("HTTP_X_USER_ID")
        user_tier = UserTier(environ.get("HTTP_X_USER_TIER", "free"))
        endpoint = environ.get("PATH_INFO", "/")

        # Determine endpoint category
        category = "default"
        if "/search" in endpoint:
            category = "search"
        elif "/export" in endpoint:
            category = "export"

        max_requests, window = get_rate_limit_for_request(user_tier, category)
        key = f"tiered:{user_id or environ.get('REMOTE_ADDR')}:{category}"

        allowed, current, limit, retry_after = self.limiter.is_allowed(
            key, max_requests, window)

        if not allowed:
            status = "429 Too Many Requests"
            headers = [
                ("Content-Type", "application/json"),
                ("Retry-After", str(retry_after)),
                ("X-RateLimit-Limit", str(limit)),
                ("X-RateLimit-Remaining", "0"),
            ]
            start_response(status, headers)
            body = f'{{"error":"rate_limit_exceeded","retry_after":{retry_after},"tier":"{user_tier.value}"}}'
            return [body.encode()]

        return self.app(environ, start_response)

步骤 5：微服务分布式限流

# Centralized rate limiting service using Redis Cluster
import redis
from redis.cluster import RedisCluster

class DistributedRateLimiter:
    """Rate limiter for microservice architectures using Redis Cluster."""

    def __init__(self):
        self.redis = RedisCluster(
            startup_nodes=[
                {"host": "redis-node-1", "port": 6379},
                {"host": "redis-node-2", "port": 6379},
                {"host": "redis-node-3", "port": 6379},
            ],
            decode_responses=True
        )

    def check_and_increment(self, service_name, user_id, endpoint,
                             max_requests, window_seconds):
        """Atomic check-and-increment using Redis Lua script."""
        key = f"rl:{{{service_name}}}:{user_id}:{endpoint}"

        # Lua script ensures atomicity across the check and increment
        lua_script = """
        local key = KEYS[1]
        local max_requests = tonumber(ARGV[1])
        local window = tonumber(ARGV[2])
        local now = tonumber(ARGV[3])
        local window_start = now - window

        -- Remove old entries
        redis.call('zremrangebyscore', key, '-inf', window_start)

        -- Count current entries
        local count = redis.call('zcard', key)

        if count >= max_requests then
            -- Get oldest entry for retry-after calculation
            local oldest = redis.call('zrange', key, 0, 0, 'WITHSCORES')
            local retry_after = 0
            if #oldest > 0 then
                retry_after = math.ceil(tonumber(oldest[2]) + window - now)
            end
            return {0, count, retry_after}
        end

        -- Add new entry
        redis.call('zadd', key, now, now .. ':' .. math.random(100000))
        redis.call('expire', key, window + 1)

        return {1, count + 1, 0}
        """

        result = self.redis.eval(lua_script, 1, key,
                                  max_requests, window_seconds, time.time())
        return {
            "allowed": bool(result[0]),
            "current": int(result[1]),
            "retry_after": int(result[2]),
        }

核心概念

术语	定义
滑动窗口	在滚动时间窗口内跟踪请求的限流算法，相较于固定窗口可实现更平滑的限流
令牌桶	以固定速率添加令牌，并在每次请求时消耗令牌的算法；允许在桶容量范围内进行受控突发
固定窗口	最简单的限流方式，按固定时间窗口（例如，每分钟）统计请求数，但容易在窗口边界出现突发流量
429 Too Many Requests	表示客户端超出限流值的 HTTP 状态码，通常附带 Retry-After 标头
Retry-After 标头	告知客户端重试前应等待多少秒的 HTTP 响应标头，对于行为良好的 API 客户端至关重要
分布式限流	使用共享状态（Redis、Memcached）在多个服务实例间进行限流，以维护准确的全局计数器

工具与系统

• Redis：用于存储分布式限流计数器的内存数据存储，通过 Lua 脚本执行原子操作
• Kong Rate Limiting Plugin：API 网关插件，支持使用 Redis 后端实现固定窗口和滑动窗口限流
• express-rate-limit：Express.js 中间件，可通过 Redis、Memcached 或内存存储实现简单限流
• Flask-Limiter：Flask 扩展，支持多种后端，并可按端点配置限流值
• Envoy Rate Limit Service：面向基于 Envoy 的服务网格架构的集中式限流服务

常见场景

场景：为公共 API 实现速率限制

背景：一家公司推出了包含免费版、高级版和企业版的公共 API。API 必须防止滥用，同时为付费客户提供公平的访问权限。该 API 在 AWS ALB 后方运行于 6 个实例上。

方案：

1. 部署 Redis Cluster（3 个节点），用于存储分布式速率限制状态
2. 使用 Redis 有序集合和 Lua 脚本实现滑动窗口速率限制，确保原子性
3. 配置各层级的限制：免费版（60 次请求/分钟）、高级版（300 次请求/分钟）、企业版（1000 次请求/分钟）
4. 为身份验证端点设置更严格的限制（每个 IP 5 次请求/分钟），不受层级影响
5. 为资源密集型端点设置限制（搜索：免费版 10 次请求/分钟；导出：5 次请求/小时）
6. 在每个响应中设置速率限制响应标头（X-RateLimit-Limit、X-RateLimit-Remaining、X-RateLimit-Reset）
7. 超出限制时，返回包含 Retry-After 标头和 JSON 错误正文的 429 响应
8. 设置 Prometheus 指标以统计速率限制触发次数，并配置 CloudWatch 警报以检测异常模式

注意事项：

• 在实例上使用内存中的速率限制而不共享状态，导致客户端可以通过访问不同服务器绕过限制
• 未将身份验证端点的速率限制与常规 API 限制分开实施
• 使用固定窗口，导致窗口边界处出现突发流量（短时间内达到 2 倍限制）
• 未在成功响应中包含速率限制标头，导致客户端无法了解自己的配额
• 未验证 X-Forwarded-For 是否来自负载均衡器，就信任它来识别 IP

输出格式

## Rate Limiting Implementation Report

**API**: Public API v2
**Algorithm**: Sliding Window (Redis Sorted Sets)
**Backend**: Redis Cluster (3 nodes)
**Deployment**: 6 API instances behind AWS ALB

### Rate Limit Configuration

| Tier | Default | Search | Export | Auth (per IP) |
|------|---------|--------|--------|---------------|
| Free | 60/min | 10/min | 5/hour | 5/min |
| Premium | 300/min | 50/min | 20/hour | 5/min |
| Enterprise | 1000/min | 200/min | 100/hour | 10/min |

### Validation Results (k6 load test)

- Free tier: Rate limited at 61st request (correct)
- Premium tier: Rate limited at 301st request (correct)
- Cross-instance: Rate limiting consistent across all 6 instances
- Redis failover: Rate limiting degrades gracefully (allows traffic) when Redis is unreachable
- Retry-After header: Accurate within 1 second of actual reset time
- Response overhead: < 2ms added latency per request for rate limit check