【Hystrix技术指南】（7）故障切换的运作流程原理分析（含源码）

背景介绍

目前对于一些非核心操作，如增减库存后保存操作日志发送异步消息时（具体业务流程），一旦出现MQ服务异常时，会导致接口响应超时，因此可以考虑对非核心操作引入服务降级、服务隔离。

Hystrix说明

官方文档

Hystrix是Netflix开源的一个容灾框架，解决当外部依赖故障时拖垮业务系统、甚至引起雪崩的问题。

为什么需要Hystrix?

在大中型分布式系统中，通常系统很多依赖(HTTP,hession,Netty,Dubbo等)，在高并发访问下,这些依赖的稳定性与否对系统的影响非常大,但是依赖有很多不可控问题：如网络连接缓慢，资源繁忙，暂时不可用，服务脱机等。
*当依赖阻塞时,大多数服务器的线程池就出现阻塞(BLOCK),影响整个线上服务的稳定性，在复杂的分布式架构的应用程序有很多的依赖，都会不可避免地在某些时候失败。高并发的依赖失败时如果没有隔离措施，当前应用服务就有被拖垮的风险。

例如:一个依赖30个SOA服务的系统,每个服务99.9999.990.3换算成时间大约每月有2个小时服务不稳定.随着服务依赖数量的变多，服务不稳定的概率会成指数性提高.解决问题方案：对依赖做隔离。
复制代码

Hystrix设计理念

想要知道如何使用，必须先明白其核心设计理念，Hystrix基于命令模式，通过UML图先直观的认识一下这一设计模式。

可见，Command是在 Receiver和 Invoker之间添加的中间层， Command实现了对Receiver的封装。
API既可以是Invoker又可以是reciever，通过继承Hystrix核心类HystrixCommand来封装这些API（例如，远程接口调用，数据库查询之类可能会产生延时的操作）。
*就可以为API提供弹性保护了。

Hystrix如何解决依赖隔离

Hystrix使用命令模式HystrixCommand(Command)包装依赖调用逻辑，每个命令在单独线程中/信号授权下执行。
可配置依赖调用超时时间，超时时间一般设为比99.5%平均时间略高即可。当调用超时时，直接返回或执行fallback逻辑。
为每个依赖提供一个小的线程池（或信号），如果线程池已满调用将被立即拒绝，默认不采用排队，加速失败判定时间。
依赖调用结果分，成功，失败（抛出异常），超时，线程拒绝，短路。请求失败(异常，拒绝，超时，短路)时执行fallback(降级)逻辑。
提供熔断器组件，可以自动运行或手动调用，停止当前依赖一段时间(10秒)，熔断器默认错误率阈值为50%，超过将自动运行。
提供近实时依赖的统计和监控

Hystrix流程结构解析

、

流程说明:

每次调用构建HystrixCommand或者HystrixObservableCommand对象，把依赖调用封装在run()方法中.
结果是否有缓存如果没有执行execute()/queue做sync或async调用，对应真正的run()/construct()
判断熔断器(circuit-breaker)是否打开，如果打开跳到步骤8，进行降级策略，如果关闭进入步骤.
判断线程池/队列/信号量是否跑满，如果跑满进入降级步骤8，否则继续后续步骤.
使用HystrixObservableCommand.construct()还是HystrixCommand.run()，运行依赖逻辑
依赖逻辑调用超时，进入步骤8
判断逻辑是否调用成功

6a 返回成功调用结果
6b 调用出错，进入步骤8.

计算熔断器状态,所有的运行状态(成功, 失败, 拒绝,超时)上报给熔断器，用于统计从而判断熔断器状态.
getFallback()降级逻辑. a. 没有实现getFallback的Command将直接抛出异常 b. fallback降级逻辑调用成功直接返回 c. 降级逻辑调用失败抛出异常
返回执行成功结果

以下四种情况将触发getFallback调用：

run()方法抛出非HystrixBadRequestException异常。
run()方法调用超时
熔断器开启短路调用
线程池/队列/信号量是否跑满

熔断器:Circuit Breaker

每个熔断器默认维护10个bucket，每秒一个bucket，每个bucket记录成功，失败,超时,拒绝的状态，默认错误超过50%且10秒内超过20个请求进行中断短路。

Hystrix隔离分析

Hystrix隔离方式采用线程/信号的方式,通过隔离限制依赖的并发量和阻塞扩散.

线程隔离

执行依赖代码的线程与请求线程(如:jetty线程)分离，请求线程可以自由控制离开的时间(异步过程)。
通过线程池大小可以控制并发量，当线程池饱和时可以提前拒绝服务,防止依赖问题扩散。
*线上建议线程池不要设置过大，否则大量堵塞线程有可能会拖慢服务器。

实际案例：

Netflix公司内部认为线程隔离开销足够小，不会造成重大的成本或性能的影响。Netflix 内部API 每天100亿的HystrixCommand依赖请求使用线程隔，每个应用大约40多个线程池，每个线程池大约5-20个线程。

信号隔离

信号隔离也可以用于限制并发访问，防止阻塞扩散, 与线程隔离最大不同在于执行依赖代码的线程依然是请求线程（该线程需要通过信号申请）,如果客户端是可信的且可以快速返回，可以使用信号隔离替换线程隔离,降低开销。

信号量的大小可以动态调整, 线程池大小不可以。

线程隔离与信号隔离区别如下图:

fallback故障切换降级机制

有兴趣的小伙伴可以看看：官方参考文档

源码分析

hystrix-core-1.5.12-sources.jar!/com/netflix/hystrix/AbstractCommand.java

executeCommandAndObserve

private Observable executeCommandAndObserve(final AbstractCommand _cmd) {final Func1> handleFallback = new Func1>() {public Observable call(Throwable t) {circuitBreaker.markNonSuccess();Exception e = getExceptionFromThrowable(t);executionResult = executionResult.setExecutionException(e);if (e instanceof RejectedExecutionException) {return handleThreadPoolRejectionViaFallback(e);} else if (t instanceof HystrixTimeoutException) {return handleTimeoutViaFallback();} else if (t instanceof HystrixBadRequestException) {return handleBadRequestByEmittingError(e);} else {if (e instanceof HystrixBadRequestException) {eventNotifier.markEvent(HystrixEventType.BAD_REQUEST, commandKey);return Observable.error(e);}return handleFailureViaFallback(e);}}};Observable execution;if (properties.executionTimeoutEnabled().get()) {execution = executeCommandWithSpecifiedIsolation(_cmd).lift(new HystrixObservableTimeoutOperator(_cmd));} else {execution = executeCommandWithSpecifiedIsolation(_cmd);}return execution.doOnNext(markEmits).doOnCompleted(markOnCompleted).onErrorResumeNext(handleFallback).doOnEach(setRequestContext);}
复制代码

使用Observable的onErrorResumeNext，里头调用了handleFallback，handleFallback中区分不同的异常来调用不同的fallback。

RejectedExecutionException调用handleThreadPoolRejectionViaFallback
HystrixTimeoutException调用handleTimeoutViaFallback
*非HystrixBadRequestException的调用handleFailureViaFallback

applyHystrixSemantics

    private Observable applyHystrixSemantics(final AbstractCommand _cmd) {executionHook.onStart(_cmd);if (circuitBreaker.attemptExecution()) {final TryableSemaphore executionSemaphore = getExecutionSemaphore();final AtomicBoolean semaphoreHasBeenReleased = new AtomicBoolean(false);final Action0 singleSemaphoreRelease = new Action0() {public void call() {if (semaphoreHasBeenReleased.compareAndSet(false, true)) {executionSemaphore.release();}}};final Action1 markExceptionThrown = new Action1() {public void call(Throwable t) {eventNotifier.markEvent(HystrixEventType.EXCEPTION_THROWN, commandKey);}};if (executionSemaphore.tryAcquire()) {try {executionResult = executionResult.setInvocationStartTime(System.currentTimeMillis());return executeCommandAndObserve(_cmd).doOnError(markExceptionThrown).doOnTerminate(singleSemaphoreRelease).doOnUnsubscribe(singleSemaphoreRelease);} catch (RuntimeException e) {return Observable.error(e);}} else {return handleSemaphoreRejectionViaFallback();}} else {return handleShortCircuitViaFallback();}}
复制代码

applyHystrixSemantics方法针对executionSemaphore.tryAcquire()没通过的调用
handleSemaphoreRejectionViaFallback
applyHystrixSemantics方法针对circuitBreaker.attemptExecution()没通过的调用handleShortCircuitViaFallback()

ViaFallback方法

    private Observable handleSemaphoreRejectionViaFallback() {Exception semaphoreRejectionException = new RuntimeException("could not acquire a semaphore for execution");executionResult = executionResult.setExecutionException(semaphoreRejectionException);eventNotifier.markEvent(HystrixEventType.SEMAPHORE_REJECTED, commandKey);logger.debug("HystrixCommand Execution Rejection by Semaphore.");return getFallbackOrThrowException(this, HystrixEventType.SEMAPHORE_REJECTED, FailureType.REJECTED_SEMAPHORE_EXECUTION,"could not acquire a semaphore for execution", semaphoreRejectionException);}private Observable handleShortCircuitViaFallback() {eventNotifier.markEvent(HystrixEventType.SHORT_CIRCUITED, commandKey);Exception shortCircuitException = new RuntimeException("Hystrix circuit short-circuited and is OPEN");executionResult = executionResult.setExecutionException(shortCircuitException);try {return getFallbackOrThrowException(this, HystrixEventType.SHORT_CIRCUITED, FailureType.SHORTCIRCUIT,"short-circuited", shortCircuitException);} catch (Exception e) {return Observable.error(e);}}private Observable handleThreadPoolRejectionViaFallback(Exception underlying) {eventNotifier.markEvent(HystrixEventType.THREAD_POOL_REJECTED, commandKey);threadPool.markThreadRejection();return getFallbackOrThrowException(this, HystrixEventType.THREAD_POOL_REJECTED, FailureType.REJECTED_THREAD_EXECUTION, "could not be queued for execution", underlying);}private Observable handleTimeoutViaFallback() {return getFallbackOrThrowException(this, HystrixEventType.TIMEOUT, FailureType.TIMEOUT, "timed-out", new TimeoutException());}private Observable handleFailureViaFallback(Exception underlying) {logger.debug("Error executing HystrixCommand.run(). Proceeding to fallback logic ...", underlying);eventNotifier.markEvent(HystrixEventType.FAILURE, commandKey);executionResult = executionResult.setException(underlying);return getFallbackOrThrowException(this, HystrixEventType.FAILURE, FailureType.COMMAND_EXCEPTION, "failed", underlying);}
复制代码

handleSemaphoreRejectionViaFallback、handleShortCircuitViaFallback、handleThreadPoolRejectionViaFallback、handleTimeoutViaFallback、handleFailureViaFallback这几个方法调用了getFallbackOrThrowException
其eventType分别是SEMAPHORE_REJECTED、SHORT_CIRCUITED、THREAD_POOL_REJECTED、TIMEOUT、FAILURE
AbstractCommand.getFallbackOrThrowException

hystrix-core-1.5.12-sources.jar!/com/netflix/hystrix/AbstractCommand.java

private Observable getFallbackOrThrowException(final AbstractCommand _cmd, final HystrixEventType eventType, final FailureType failureType, final String message, final Exception originalException) {final HystrixRequestContext requestContext = HystrixRequestContext.getContextForCurrentThread();long latency = System.currentTimeMillis() - executionResult.getStartTimestamp();executionResult = executionResult.addEvent((int) latency, eventType);if (isUnrecoverable(originalException)) {logger.error("Unrecoverable Error for HystrixCommand so will throw HystrixRuntimeException and not apply fallback. ", originalException);Exception e = wrapWithOnErrorHook(failureType, originalException);return Observable.error(new HystrixRuntimeException(failureType, this.getClass(), getLogMessagePrefix() + " " + message + " and encountered unrecoverable error.", e, null));} else {if (isRecoverableError(originalException)) {logger.warn("Recovered from java.lang.Error by serving Hystrix fallback", originalException);}if (properties.fallbackEnabled().get()) {final Action1super R>> setRequestContext = new Action1super R>>() {public void call(Notificationsuper R> rNotification) {setRequestContextIfNeeded(requestContext);}};final Action1 markFallbackEmit = new Action1() {public void call(R r) {if (shouldOutputOnNextEvents()) {executionResult = executionResult.addEvent(HystrixEventType.FALLBACK_EMIT);eventNotifier.markEvent(HystrixEventType.FALLBACK_EMIT, commandKey);}}};final Action0 markFallbackCompleted = new Action0() {public void call() {long latency = System.currentTimeMillis() - executionResult.getStartTimestamp();eventNotifier.markEvent(HystrixEventType.FALLBACK_SUCCESS, commandKey);executionResult = executionResult.addEvent((int) latency,HystrixEventType.FALLBACK_SUCCESS);}};final Func1> handleFallbackError = new Func1>() {public Observable call(Throwable t) {Exception e = wrapWithOnErrorHook(failureType, originalException);Exception fe = getExceptionFromThrowable(t);long latency = System.currentTimeMillis() - executionResult.getStartTimestamp();Exception toEmit;if (fe instanceof UnsupportedOperationException) {logger.debug("No fallback for HystrixCommand. ", fe);eventNotifier.markEvent(HystrixEventType.FALLBACK_MISSING, commandKey);executionResult = executionResult.addEvent((int) latency, HystrixEventType.FALLBACK_MISSING);toEmit = new HystrixRuntimeException(failureType, _cmd.getClass(), getLogMessagePrefix() + " " + message + " and no fallback available.", e, fe);} else {logger.debug("HystrixCommand execution " + failureType.name() + " and fallback failed.", fe);eventNotifier.markEvent(HystrixEventType.FALLBACK_FAILURE, commandKey);executionResult = executionResult.addEvent((int) latency, HystrixEventType.FALLBACK_FAILURE);toEmit = new HystrixRuntimeException(failureType, _cmd.getClass(), getLogMessagePrefix() + " " + message + " and fallback failed.", e, fe);}if (shouldNotBeWrapped(originalException)) {return Observable.error(e);}return Observable.error(toEmit);}};final TryableSemaphore fallbackSemaphore = getFallbackSemaphore();final AtomicBoolean semaphoreHasBeenReleased = new AtomicBoolean(false);final Action0 singleSemaphoreRelease = new Action0() {public void call() {if (semaphoreHasBeenReleased.compareAndSet(false, true)) {fallbackSemaphore.release();}}};Observable fallbackExecutionChain;if (fallbackSemaphore.tryAcquire()) {try {if (isFallbackUserDefined()) {executionHook.onFallbackStart(this);fallbackExecutionChain = getFallbackObservable();} else {fallbackExecutionChain = getFallbackObservable();}} catch (Throwable ex) {fallbackExecutionChain = Observable.error(ex);}return fallbackExecutionChain.doOnEach(setRequestContext).lift(new FallbackHookApplication(_cmd)).lift(new DeprecatedOnFallbackHookApplication(_cmd)).doOnNext(markFallbackEmit).doOnCompleted(markFallbackCompleted).onErrorResumeNext(handleFallbackError).doOnTerminate(singleSemaphoreRelease).doOnUnsubscribe(singleSemaphoreRelease);} else {return handleFallbackRejectionByEmittingError();}} else {return handleFallbackDisabledByEmittingError(originalException, failureType, message);}}}
复制代码

fallbackExecutionChain的onErrorResumeNext，调用了handleFallbackError
fallbackExecutionChain的doOnCompleted，调用了markFallbackCompleted
AbstractCommand.getFallbackSemaphore

hystrix-core-1.5.12-sources.jar!/com/netflix/hystrix/AbstractCommand.java

protected TryableSemaphore getFallbackSemaphore() {if (fallbackSemaphoreOverride == null) {TryableSemaphore _s = fallbackSemaphorePerCircuit.get(commandKey.name());if (_s == null) {fallbackSemaphorePerCircuit.putIfAbsent(commandKey.name(), new TryableSemaphoreActual(properties.fallbackIsolationSemaphoreMaxConcurrentRequests()));return fallbackSemaphorePerCircuit.get(commandKey.name());} else {return _s;}} else {return fallbackSemaphoreOverride;}}
复制代码