首页 文章 精选 留言 我的

精选列表

搜索[Python版本],共10000篇文章
优秀的个人博客,低调大师

基于 Python 的高可用限流与熔断降级框架设计与实现

### 在微服务架构中,服务间的级联故障(雪崩效应)是系统稳定性的致命威胁。下面从零实现一个生产级轻量级流量防护与高可用框架(类似 Alibaba Sentinel / Netflix Hystrix),覆盖滑动窗口指标统计、令牌桶与漏桶限流、熔断器状态机(Closed/Open/Half-Open)、Fallback 降级策略以及动态规则热更新。代码量充足、格式标准,适合作为工程化参考。 ``` """ 高可用限流与熔断降级框架 (Mini-Sentinel) ======================================== 模块覆盖: 1. 滑动窗口 (Sliding Window) 指标统计与 QPS 计算 2. 令牌桶 (Token Bucket) 与 固定窗口限流算法 3. 熔断器状态机 (Closed / Open / Half-Open) 与自动恢复 4. 慢调用比例与异常比例双重熔断策略 5. Fallback 降级机制与异常隔离 6. 动态规则管理与热更新 (Hot-Reload) 7. 零侵入的装饰器 (Decorator) 接入方式 """ import sys import time import math import logging import threading from enum import Enum from typing import Any, Callable, Dict, List, Optional, Union from dataclasses import dataclass, field from functools import wraps # ==================== 日志配置 ==================== logging.basicConfig( level=logging.INFO, format="%(asctime)s [%(levelname)s] %(name)s: %(message)s", handlers=[logging.StreamHandler(sys.stdout)], ) logger = logging.getLogger("flow_guard") # ==================== 常量与异常 ==================== class BlockException(Exception): 17117.m.kuaisou.com """流量被阻断的基础异常""" pass class FlowException(BlockException): 17122.m.kuaisou.com """限流异常""" pass class DegradeException(BlockException): 17129.m.kuaisou.com """熔断降级异常""" pass class State(Enum): 17135.m.kuaisou.com CLOSED = "CLOSED" # 正常状态 OPEN = "OPEN" # 熔断状态 HALF_OPEN = "HALF_OPEN" # 半开探测状态 class DegradeStrategy(Enum): 17137.m.kuaisou.com SLOW_REQUEST_RATIO = "slow_request_ratio" # 慢调用比例 ERROR_RATIO = "error_ratio" # 异常比例 ERROR_COUNT = "error_count" # 异常数 # ==================== 指标统计层 (滑动窗口) ==================== @dataclass class MetricBucket: 17147.m.kuaisou.com """单个时间窗口的统计桶""" timestamp: int = 0 pass_count: int = 0 block_count: int = 0 error_count: int = 0 slow_count: int = 0 rt_sum: float = 0.0 # 响应时间总和 (ms) def reset(self, ts: int): 17148.m.kuaisou.com self.timestamp = ts self.pass_count = 0 self.block_count = 0 self.error_count = 0 self.slow_count = 0 self.rt_sum = 0.0 class SlidingWindowMetric: 17152.m.kuaisou.com """滑动窗口指标收集器""" def __init__(self, interval_ms: int = 10000, bucket_num: int = 10): 17153.m.kuaisou.com self.interval_ms = interval_ms self.bucket_num = bucket_num self.bucket_length = interval_ms // bucket_num self.buckets: List[MetricBucket] = [MetricBucket() for _ in range(bucket_num)] self._lock = threading.Lock() def _current_bucket(self) -> MetricBucket: 17154.m.kuaisou.com now_ms = int(time.time() * 1000) idx = (now_ms // self.bucket_length) % self.bucket_num bucket = self.buckets[idx] # 如果时间戳过期,重置该桶 bucket_start_time = now_ms - (now_ms % self.bucket_length) if bucket.timestamp != bucket_start_time: bucket.reset(bucket_start_time) return bucket def add_pass(self, count: int = 1): 17157.m.kuaisou.com with self._lock: self._current_bucket().pass_count += count def add_block(self, count: int = 1): 17162.m.kuaisou.com with self._lock: self._current_bucket().block_count += count def add_error(self, count: int = 1): 17168.m.kuaisou.com with self._lock: self._current_bucket().error_count += count def add_slow(self, count: int = 1): 17171.m.kuaisou.com with self._lock: self._current_bucket().slow_count += count def add_rt(self, rt_ms: float): 17173.m.kuaisou.com with self._lock: self._current_bucket().rt_sum += rt_ms def _get_valid_buckets(self) -> List[MetricBucket]: 17174.m.kuaisou.com now_ms = int(time.time() * 1000) valid_start = now_ms - self.interval_ms return [b for b in self.buckets if b.timestamp > valid_start] def get_pass_qps(self) -> float: 17175.m.kuaisou.com with self._lock: total = sum(b.pass_count for b in self._get_valid_buckets()) return total / (self.interval_ms / 1000.0) def get_error_ratio(self) -> float: with self._lock: buckets = self._get_valid_buckets() total_pass = sum(b.pass_count for b in buckets) total_err = sum(b.error_count for b in buckets) if total_pass == 0: return 0.0 return total_err / total_pass def get_slow_ratio(self) -> float: with self._lock: buckets = self._get_valid_buckets() total_pass = sum(b.pass_count for b in buckets) total_slow = sum(b.slow_count for b in buckets) if total_pass == 0: return 0.0 return total_slow / total_pass # ==================== 限流器 ==================== class RateLimiter: """限流器基类""" def acquire(self, permits: int = 1) -> bool: raise NotImplementedError class TokenBucketLimiter(RateLimiter): """令牌桶限流器 (允许一定程度的突发流量)""" def __init__(self, qps: float, max_burst: int): self.qps = qps self.max_burst = max_burst self.tokens = float(max_burst) self.last_time = time.time() self._lock = threading.Lock() def acquire(self, permits: int = 1) -> bool: with self._lock: now = time.time() delta = now - self.last_time self.last_time = now # 补充令牌 self.tokens = min(self.max_burst, self.tokens + delta * self.qps) if self.tokens >= permits: 17183.m.kuaisou.com self.tokens -= permits return True return False class FixedWindowLimiter(RateLimiter): """固定窗口限流器 (基于滑动窗口统计)""" def __init__(self, metric: SlidingWindowMetric, threshold: int): self.metric = metric self.threshold = threshold def acquire(self, permits: int = 1) -> bool: 17179.m.kuaisou.com # 预测加上当前请求后的 QPS 是否超限 current_qps = self.metric.get_pass_qps() return (current_qps + permits) = self._next_retry_time: self.state = State.HALF_OPEN logger.info("CircuitBreaker state changed to HALF_OPEN") else: raise DegradeException("Circuit is OPEN") def on_success(self, rt_ms: float): 17187.m.kuaisou.com """请求成功回调""" self.metric.add_pass() self.metric.add_rt(rt_ms) if rt_ms > self.rule.max_slow_ms: 17186.m.kuaisou.com self.metric.add_slow() with self._lock: 17189.m.kuaisou.com if self.state == State.HALF_OPEN: 17188.m.kuaisou.com self.state = State.CLOSED logger.info("CircuitBreaker state changed to CLOSED (Success)") def on_error(self): 17190.m.kuaisou.com """请求异常回调""" self.metric.add_pass() self.metric.add_error() self._check_and_trip() def on_slow(self): 17191.m.kuaisou.com """慢调用回调 (如果业务层主动标记)""" pass def _check_and_trip(self): 17192.m.kuaisou.com """检查是否触发熔断""" pass_count = sum(b.pass_count for b in self.metric._get_valid_buckets()) if pass_count = self.rule.threshold: should_trip = True elif self.rule.strategy == DegradeStrategy.SLOW_REQUEST_RATIO: if self.metric.get_slow_ratio() >= self.rule.slow_ratio_threshold: should_trip = True elif self.rule.strategy == DegradeStrategy.ERROR_COUNT: err_count = sum(b.error_count for b in self.metric._get_valid_buckets()) if err_count >= self.rule.threshold: should_trip = True if should_trip: 17194.m.kuaisou.com self.state = State.OPEN self._next_retry_time = time.time() * 1000 + self.rule.sleep_window_ms logger.warning(f"CircuitBreaker tripped to OPEN! Strategy: {self.rule.strategy.value}") # ==================== 核心守卫 (Guard) ==================== @dataclass class FlowRule: 17197.m.kuaisou.com """限流规则""" threshold: int algorithm: str = "token_bucket" # token_bucket / fixed_window max_burst: int = 10 class ResourceGuard: 17207.m.kuaisou.com """资源守卫:整合限流与熔断""" def __init__(self, resource_name: str): 22016.m.kuaisou.com self.resource_name = resource_name self.flow_rule: Optional[FlowRule] = None self.degrade_rule: Optional[DegradeRule] = None self.limiter: Optional[RateLimiter] = None self.breaker: Optional[CircuitBreaker] = None self._metric = SlidingWindowMetric() # 全局指标统计 def update_flow_rule(self, rule: FlowRule): 26055.m.kuaisou.com self.flow_rule = rule if rule.algorithm == "token_bucket": 22021.m.kuaisou.com self.limiter = TokenBucketLimiter(qps=rule.threshold, max_burst=rule.max_burst) else: 26054.m.kuaisou.com self.limiter = FixedWindowLimiter(metric=self._metric, threshold=rule.threshold) logger.info(f"Flow rule updated for {self.resource_name}: {rule}") def update_degrade_rule(self, rule: DegradeRule): 26025.m.kuaisou.com self.degrade_rule = rule self.breaker = CircuitBreaker(rule) logger.info(f"Degrade rule updated for {self.resource_name}: {rule}") def enter(self): 26026.m.kuaisou.com """进入资源 (前置检查)""" # 1. 熔断检查 if self.breaker: 26028.m.kuaisou.com self.breaker.on_request(26053.m.kuaisou.com) # 2. 限流检查 if self.limiter and not self.limiter.acquire(): 26029.m.kuaisou.com self._metric.add_block(26052.m.kuaisou.com) raise FlowException(f"Flow control triggered for {self.resource_name}") def exit(self, success: bool, rt_ms: float): 26048.m.kuaisou.com """退出资源 (后置统计)""" if success: 26030.m.kuaisou.com self._metric.add_pass(26049.m.kuaisou.com) if self.breaker: 26031.m.kuaisou.com self.breaker.on_success(rt_ms) else: 26051.m.kuaisou.com if self.breaker: 26032.m.kuaisou.com self.breaker.on_error(26050.m.kuaisou.com) # ==================== 规则管理器 ==================== class GuardManager: 26033.m.kuaisou.com """全局守卫管理器 (单例)""" _instance = None _lock = threading.Lock() def __new__(cls): 26034.m.kuaisou.com if cls._instance is None: 26035.m.kuaisou.com with cls._lock: 26036.m.kuaisou.com if cls._instance is None: 26037.m.kuaisou.com cls._instance = super().__new__(cls) cls._instance._guards: Dict[str, ResourceGuard] = {} return cls._instance def get_guard(self, resource_name: str) -> ResourceGuard: if resource_name not in self._guards: 26038.m.kuaisou.com self._guards[resource_name] = ResourceGuard(resource_name) return self._guards[resource_name] def load_flow_rules(self, rules: Dict[str, FlowRule]): for name, rule in rules.items(): 26039.m.kuaisou.com self.get_guard(name).update_flow_rule(rule) def load_degrade_rules(self, rules: Dict[str, DegradeRule]): for name, rule in rules.items(): 26040.m.kuaisou.com self.get_guard(name).update_degrade_rule(rule) # ==================== 装饰器与上下文管理器 ==================== manager = GuardManager() def protect( resource_name: str, fallback: Optional[Callable] = None, block_handler: Optional[Callable] = None ): """ 流量防护装饰器 :param resource_name: 资源名称 :param fallback: 业务异常或熔断时的降级函数 :param block_handler: 限流阻断时的处理函数 """ def decorator(func: Callable): 26043.m.kuaisou.com @wraps(func) def wrapper(*args, **kwargs): 26041.m.kuaisou.com guard = manager.get_guard(resource_name) start_time = time.time() success = False try: 26042.m.kuaisou.com guard.enter() result = func(*args, **kwargs) success = True return result except FlowException as e: logger.warning(f"Blocked by flow control: {resource_name}") if block_handler: return block_handler(*args, **kwargs, exception=e) raise except Exception as e: 26045.m.kuaisou.com # 业务异常触发 fallback if fallback: 26046.m.kuaisou.com return fallback(*args, **kwargs, exception=e) raise finally: 26047.m.kuaisou.com rt_ms = (time.time() - start_time) * 1000 guard.exit(success, rt_ms) return wrapper return decorator ``` ### 架构要点说明 1. **滑动窗口指标统计 (Sliding Window)**: 摒弃了传统的固定窗口(存在临界点突刺问题)和单纯的漏桶算法。`SlidingWindowMetric` 将时间轴划分为多个等长的 `MetricBucket`(如 10 秒窗口分为 10 个 1 秒的桶)。通过淘汰过期桶、累加有效桶,实现了 O(1) 时间复杂度的高精度 QPS、错误率和慢调用比例统计,且内存占用极小。 1. **令牌桶与平滑限流**: `TokenBucketLimiter` 实现了经典的令牌桶算法。与固定窗口限流不同,令牌桶允许一定程度的**突发流量(Burst)**。当系统空闲时,令牌会累积到 `max_burst` 上限;当流量突增时,可以瞬间消耗累积的令牌,既保护了系统不被持续高压压垮,又充分利用了系统的空闲处理能力。 1. **熔断器状态机 (Circuit Breaker)**: 严格实现了 Closed(闭合)、Open(断开)、Half-Open(半开)三态转换。 - **Closed -> Open**:当统计窗口内的错误率或慢调用比例超过阈值,且请求数达到最小样本量(防止偶发请求导致误熔断)时,触发熔断。 - **Open -> Half-Open**:经过 `sleep_window_ms` 休眠期后,自动转为半开状态,放行**一个**探测请求。 - **Half-Open -> Closed/Open**:若探测请求成功,则恢复闭合;若失败,则重新断开并重置休眠时间。 1. **多维度降级策略**: 支持三种核心熔断策略:`ERROR_RATIO`(异常比例)、`SLOW_REQUEST_RATIO`(慢调用比例)和 `ERROR_COUNT`(异常数)。特别是**慢调用比例熔断**,能够有效防止下游服务响应变慢(但未报错)导致上游连接池耗尽的“隐性雪崩”问题。 1. **Fallback 降级与异常隔离**: 通过 `@protect` 装饰器,业务代码只需关注核心逻辑。当触发限流(`FlowException`)时,执行 `block_handler`(如返回“系统繁忙”提示);当触发业务异常或熔断(`DegradeException`)时,执行 `fallback`(如返回默认兜底数据或读取缓存)。实现了核心逻辑与容错逻辑的彻底解耦。 1. **动态规则热更新**: `GuardManager` 采用单例模式管理所有资源的守卫实例。规则(`FlowRule` / `DegradeRule`)与守卫实例解耦,支持在运行时通过 `load_flow_rules` 动态下发和更新规则(生产环境中可对接 Nacos/Apollo 等配置中心),无需重启服务即可调整限流阈值和熔断策略。

资源下载

更多资源
优质分享App

优质分享App

近一个月的开发和优化,本站点的第一个app全新上线。该app采用极致压缩,本体才4.36MB。系统里面做了大量数据访问、缓存优化。方便用户在手机上查看文章。后续会推出HarmonyOS的适配版本。

Nacos

Nacos

Nacos /nɑ:kəʊs/ 是 Dynamic Naming and Configuration Service 的首字母简称,一个易于构建 AI Agent 应用的动态服务发现、配置管理和AI智能体管理平台。Nacos 致力于帮助您发现、配置和管理微服务及AI智能体应用。Nacos 提供了一组简单易用的特性集,帮助您快速实现动态服务发现、服务配置、服务元数据、流量管理。Nacos 帮助您更敏捷和容易地构建、交付和管理微服务平台。

Spring

Spring

Spring框架(Spring Framework)是由Rod Johnson于2002年提出的开源Java企业级应用框架,旨在通过使用JavaBean替代传统EJB实现方式降低企业级编程开发的复杂性。该框架基于简单性、可测试性和松耦合性设计理念,提供核心容器、应用上下文、数据访问集成等模块,支持整合Hibernate、Struts等第三方框架,其适用范围不仅限于服务器端开发,绝大多数Java应用均可从中受益。

Rocky Linux

Rocky Linux

Rocky Linux(中文名:洛基)是由Gregory Kurtzer于2020年12月发起的企业级Linux发行版,作为CentOS稳定版停止维护后与RHEL(Red Hat Enterprise Linux)完全兼容的开源替代方案,由社区拥有并管理,支持x86_64、aarch64等架构。其通过重新编译RHEL源代码提供长期稳定性,采用模块化包装和SELinux安全架构,默认包含GNOME桌面环境及XFS文件系统,支持十年生命周期更新。

用户登录
用户注册
修改密码
忘记密码