概要
Spring实现了一套重试机制,功能简单实用。Spring Retry是从Spring Batch独立出来的一个功能,已经广泛应用于Spring Batch,Spring Integration, Spring for Apache Hadoop等Spring项目。本文将讲述如何使用Spring Retry及其实现原理。
背景
重试,其实我们其实很多时候都需要的,为了保证容错性,可用性,一致性等。一般用来应对外部系统的一些不可预料的返回、异常等,特别是网络延迟,中断等情况。还有在现在流行的微服务治理框架中,通常都有自己的重试与超时配置,比如dubbo可以设置retries=1,timeout=500调用失败只重试1次,超过500ms调用仍未返回则调用失败。
如果我们要做重试,要为特定的某个操作做重试功能,则要硬编码,大概逻辑基本都是写个循环,根据返回或异常,计数失败次数,然后设定退出条件。这样做,且不说每个操作都要写这种类似的代码,而且重试逻辑和业务逻辑混在一起,给维护和扩展带来了麻烦。从面向对象的角度来看,我们应该把重试的代码独立出来。
使用介绍
基本使用
先举个例子:
@Configuration @EnableRetry public class Application { @Bean public RetryService retryService () { return new RetryService(); } public static void main (String[] args) throws Exception { ApplicationContext applicationContext = new AnnotationConfigApplicationContext("springretry" ); RetryService service1 = applicationContext.getBean("service" , RetryService.class ) ; service1.service(); } } @Service ("service" ) public class RetryService { @Retryable (value = IllegalAccessException.class , maxAttempts = 5 , backoff= @Backoff (value = 1500 , maxDelay = 100000 , multiplier = 1.2 )) public void service () throws IllegalAccessException { System.out.println("service method..." ); throw new IllegalAccessException("manual exception" ); } @Recover public void recover (IllegalAccessException e) { System.out.println("service retry after Recover => " + e.getMessage()); } }
@EnableRetry - 表示开启重试机制 @Retryable - 表示这个方法需要重试,它有很丰富的参数,可以满足你对重试的需求 @Backoff - 表示重试中的退避策略 @Recover - 兜底方法,即多次重试后还是失败就会执行这个方法
Spring-Retry 的功能丰富在于其重试策略和退避策略,还有兜底,监听器等操作。
然后每个注解里面的参数,都是很简单的,大家看一下就知道是什么意思,怎么用了,我就不多讲了。
重试策略
看一下Spring Retry自带的一些重试策略,主要是用来判断当方法调用异常时是否需要重试。(下文原理部分会深入分析实现)
SimpleRetryPolicy 默认最多重试3次
TimeoutRetryPolicy 默认在1秒内失败都会重试
ExpressionRetryPolicy 符合表达式就会重试
CircuitBreakerRetryPolicy 增加了熔断的机制,如果不在熔断状态,则允许重试
CompositeRetryPolicy 可以组合多个重试策略
NeverRetryPolicy 从不重试(也是一种重试策略哈)
….等等
退避策略
看一下退避策略,退避是指怎么去做下一次的重试,在这里其实就是等待多长时间。(下文原理部分会深入分析实现)
FixedBackOffPolicy 默认固定延迟1秒后执行下一次重试
ExponentialBackOffPolicy 指数递增延迟执行重试,默认初始0.1秒,系数是2,那么下次延迟0.2秒,再下次就是延迟0.4秒,如此类推,最大30秒。
ExponentialRandomBackOffPolicy 在上面那个策略上增加随机性
UniformRandomBackOffPolicy 这个跟上面的区别就是,上面的延迟会不停递增,这个只会在固定的区间随机
StatelessBackOffPolicy 这个说明是无状态的,所谓无状态就是对上次的退避无感知,从它下面的子类也能看出来
原理
原理部分我想分开两部分来讲,一是重试机制的切入点,即它是如何使得你的代码实现重试功能的;二是重试机制的详细,包括重试的逻辑以及重试策略和退避策略的实现。
切入点
@EnableRetry
@Target (ElementType.TYPE) @Retention (RetentionPolicy.RUNTIME) @EnableAspectJAutoProxy (proxyTargetClass = false ) @Import (RetryConfiguration.class ) @Documented public @interface EnableRetry { /** * Indicate whether subclass-based (CGLIB) proxies are to be created as opposed * to standard Java interface-based proxies. The default is {@code false}. * * @return whether to proxy or not to proxy the class */
boolean proxyTargetClass () default false ; }
我们可以看到
@EnableAspectJAutoProxy(proxyTargetClass = false)
这个并不陌生,就是打开Spring AOP功能。重点看看
@Import(RetryConfiguration.class)
@Import相当于注册这个Bean
我们看看这个
RetryConfiguration
是个什么东西
它是一个AbstractPointcutAdvisor,它有一个pointcut和一个advice。我们知道,在IOC过程中会根据PointcutAdvisor类来对Bean进行Pointcut的过滤,然后生成对应的AOP代理类,用advice来加强处理。看看RetryConfiguration的初始化:
@PostConstruct public void init () { Set> retryableAnnotationTypes = new LinkedHashSet>(1 ); retryableAnnotationTypes.add(Retryable.class ) ; //创建pointcut this .pointcut = buildPointcut(retryableAnnotationTypes); //创建advice this .advice = buildAdvice(); if (this .advice instanceof BeanFactoryAware) { ((BeanFactoryAware) this .advice).setBeanFactory(beanFactory); } }
~
protected Pointcut buildPointcut (Set> retryAnnotationTypes) { ComposablePointcut result = null ; for (Class extends Annotation> retryAnnotationType : retryAnnotationTypes) { Pointcut filter = new AnnotationClassOrMethodPointcut(retryAnnotationType); if (result == null ) { result = new ComposablePointcut(filter); } else { result.union(filter); } } return result; }
上面代码用到了AnnotationClassOrMethodPointcut,其实它最终还是用到了AnnotationMethodMatcher来根据注解进行切入点的过滤。这里就是@Retryable注解了。
//创建advice对象,即拦截器 protected Advice buildAdvice () { //下面关注这个对象 AnnotationAwareRetryOperationsInterceptor interceptor = new AnnotationAwareRetryOperationsInterceptor(); if (retryContextCache != null ) { interceptor.setRetryContextCache(retryContextCache); } if (retryListeners != null ) { interceptor.setListeners(retryListeners); } if (methodArgumentsKeyGenerator != null ) { interceptor.setKeyGenerator(methodArgumentsKeyGenerator); } if (newMethodArgumentsIdentifier != null ) { interceptor.setNewItemIdentifier(newMethodArgumentsIdentifier); } if (sleeper != null ) { interceptor.setSleeper(sleeper); } return interceptor; }
AnnotationAwareRetryOperationsInterceptor
可以看出AnnotationAwareRetryOperationsInterceptor是一个MethodInterceptor,在创建AOP代理过程中如果目标方法符合pointcut的规则,它就会加到interceptor列表中,然后做增强,我们看看invoke方法做了什么增强。
@Override public Object invoke (MethodInvocation invocation) throws Throwable { MethodInterceptor delegate = getDelegate(invocation.getThis(), invocation.getMethod()); if (delegate != null ) { return delegate.invoke(invocation); } else { return invocation.proceed(); } }
这里用到了委托,主要是需要根据配置委托给具体“有状态”的interceptor还是“无状态”的interceptor。
private MethodInterceptor getDelegate (Object target, Method method) { if (!this .delegates.containsKey(target) || !this .delegates.get(target).containsKey(method)) { synchronized (this .delegates) { if (!this .delegates.containsKey(target)) { this .delegates.put(target, new HashMap()); } Map delegatesForTarget = this .delegates.get(target); if (!delegatesForTarget.containsKey(method)) { Retryable retryable = AnnotationUtils.findAnnotation(method, Retryable.class ) ; if (retryable == null ) { retryable = AnnotationUtils.findAnnotation(method.getDeclaringClass(), Retryable.class ) ; } if (retryable == null ) { retryable = findAnnotationOnTarget(target, method); } if (retryable == null ) { return delegatesForTarget.put(method, null ); } MethodInterceptor delegate; //支持自定义MethodInterceptor,而且优先级最高 if (StringUtils.hasText(retryable.interceptor())) { delegate = this .beanFactory.getBean(retryable.interceptor(), MethodInterceptor.class ) ; } else if (retryable.stateful()) { //得到“有状态”的interceptor delegate = getStatefulInterceptor(target, method, retryable); } else { //得到“无状态”的interceptor delegate = getStatelessInterceptor(target, method, retryable); } delegatesForTarget.put(method, delegate); } } } return this .delegates.get(target).get(method); }
getStatefulInterceptor和getStatelessInterceptor都是差不多,我们先看看比较简单的getStatelessInterceptor。
private MethodInterceptor getStatelessInterceptor (Object target, Method method, Retryable retryable) { //生成一个RetryTemplate RetryTemplate template = createTemplate(retryable.listeners()); //生成retryPolicy template.setRetryPolicy(getRetryPolicy(retryable)); //生成backoffPolicy template.setBackOffPolicy(getBackoffPolicy(retryable.backoff())); return RetryInterceptorBuilder.stateless() .retryOperations(template) .label(retryable.label()) .recoverer(getRecoverer(target, method)) .build(); }
具体生成retryPolicy和backoffPolicy的规则,我们等下再回头来看。RetryInterceptorBuilder其实就是为了生成
RetryOperationsInterceptor
。RetryOperationsInterceptor也是一个MethodInterceptor,我们来看看它的
invoke
方法。
public Object invoke (final MethodInvocation invocation) throws Throwable { String name; if (StringUtils.hasText(label)) { name = label; } else { name = invocation.getMethod().toGenericString(); } final String label = name; //定义了一个RetryCallback,其实看它的doWithRetry方法,调用了invocation的proceed()方法,是不是有点眼熟,这就是AOP的拦截链调用,如果没有拦截链,那就是对原来方法的调用。 RetryCallback retryCallback = new RetryCallback() { public Object doWithRetry (RetryContext context) throws Exception { context.setAttribute(RetryContext.NAME, label); /* * If we don't copy the invocation carefully it won't keep a reference to * the other interceptors in the chain. We don't have a choice here but to * specialise to ReflectiveMethodInvocation (but how often would another * implementation come along?). */ if (invocation instanceof ProxyMethodInvocation) { try { return ((ProxyMethodInvocation) invocation).invocableClone().proceed(); } catch (Exception e) { throw e; } catch (Error e) { throw e; } catch (Throwable e) { throw new IllegalStateException(e); } } else { throw new IllegalStateException( "MethodInvocation of the wrong type detected - this should not happen with Spring AOP, " + "so please raise an issue if you see this exception" ); } } }; if (recoverer != null ) { ItemRecovererCallback recoveryCallback = new ItemRecovererCallback( invocation.getArguments(), recoverer); return this .retryOperations.execute(retryCallback, recoveryCallback); } //最终还是进入到retryOperations的execute方法,这个retryOperations就是在之前的builder set进来的RetryTemplate。 return this .retryOperations.execute(retryCallback); }
无论是
RetryOperationsInterceptor
还是
StatefulRetryOperationsInterceptor
,最终的拦截处理逻辑还是调用到RetryTemplate的execute方法,从名字也看出来,RetryTemplate作为一个模板类,里面包含了重试统一逻辑。不过,我看这个RetryTemplate并不是很“模板”,因为它没有很多可以扩展的地方。
搜索Java知音公众号,回复“后端面试”,送你一份Java面试题宝典
.pdf
重试逻辑及策略实现
上面介绍了Spring Retry利用了AOP代理使重试机制对业务代码进行“入侵”。下面我们继续看看重试的逻辑做了什么。RetryTemplate的doExecute方法。
protected T doExecute (RetryCallback retryCallback, RecoveryCallback recoveryCallback, RetryState state) throws E, ExhaustedRetryException { RetryPolicy retryPolicy = this .retryPolicy; BackOffPolicy backOffPolicy = this .backOffPolicy; //新建一个RetryContext来保存本轮重试的上下文 RetryContext context = open(retryPolicy, state); if (this .logger.isTraceEnabled()) { this .logger.trace("RetryContext retrieved: " + context); } // Make sure the context is available globally for clients who need // it... RetrySynchronizationManager.register(context); Throwable lastException = null ; boolean exhausted = false ; try { //如果有注册RetryListener,则会调用它的open方法,给调用者一个通知。 boolean running = doOpenInterceptors(retryCallback, context); if (!running) { throw new TerminatedRetryException( "Retry terminated abnormally by interceptor before first attempt" ); } // Get or Start the backoff context... BackOffContext backOffContext = null ; Object resource = context.getAttribute("backOffContext" ); if (resource instanceof BackOffContext) { backOffContext = (BackOffContext) resource; } if (backOffContext == null ) { backOffContext = backOffPolicy.start(context); if (backOffContext != null ) { context.setAttribute("backOffContext" , backOffContext); } } //判断能否重试,就是调用RetryPolicy的canRetry方法来判断。 //这个循环会直到原方法不抛出异常,或不需要再重试 while (canRetry(retryPolicy, context) && !context.isExhaustedOnly()) { try { if (this .logger.isDebugEnabled()) { this .logger.debug("Retry: count=" + context.getRetryCount()); } //清除上次记录的异常 lastException = null ; //doWithRetry方法,一般来说就是原方法 return retryCallback.doWithRetry(context); } catch (Throwable e) { //原方法抛出了异常 lastException = e; try { //记录异常信息 registerThrowable(retryPolicy, state, context, e); } catch (Exception ex) { throw new TerminatedRetryException("Could not register throwable" , ex); } finally { //调用RetryListener的onError方法 doOnErrorInterceptors(retryCallback, context, e); } //再次判断能否重试 if (canRetry(retryPolicy, context) && !context.isExhaustedOnly()) { try { //如果可以重试则走退避策略 backOffPolicy.backOff(backOffContext); } catch (BackOffInterruptedException ex) { lastException = e; // back off was prevented by another thread - fail the retry if (this .logger.isDebugEnabled()) { this .logger .debug("Abort retry because interrupted: count=" + context.getRetryCount()); } throw ex; } } if (this .logger.isDebugEnabled()) { this .logger.debug( "Checking for rethrow: count=" + context.getRetryCount()); } if (shouldRethrow(retryPolicy, context, state)) { if (this .logger.isDebugEnabled()) { this .logger.debug("Rethrow in retry for policy: count=" + context.getRetryCount()); } throw RetryTemplate.wrapIfNecessary(e); } } /* * A stateful attempt that can retry may rethrow the exception before now, * but if we get this far in a stateful retry there's a reason for it, * like a circuit breaker or a rollback classifier. */ if (state != null && context.hasAttribute(GLOBAL_STATE)) { break ; } } if (state == null && this .logger.isDebugEnabled()) { this .logger.debug( "Retry failed last attempt: count=" + context.getRetryCount()); } exhausted = true ; //重试结束后如果有兜底Recovery方法则执行,否则抛异常 return handleRetryExhausted(recoveryCallback, context, state); } catch (Throwable e) { throw RetryTemplate.wrapIfNecessary(e); } finally { //处理一些关闭逻辑 close(retryPolicy, context, state, lastException == null || exhausted); //调用RetryListener的close方法 doCloseInterceptors(retryCallback, context, lastException); RetrySynchronizationManager.clear(); } }
主要核心重试逻辑就是上面的代码了,看上去还是挺简单的。在上面,我们漏掉了RetryPolicy的canRetry方法和BackOffPolicy的backOff方法,以及这两个Policy是怎么来的。我们回头看看
getStatelessInterceptor
方法中的
getRetryPolicy
和
getRetryPolicy
方法。
private RetryPolicy getRetryPolicy (Annotation retryable) { Map attrs = AnnotationUtils.getAnnotationAttributes(retryable); @SuppressWarnings ("unchecked" ) Class extends Throwable>[] includes = (Class extends Throwable>[]) attrs.get("value" ); String exceptionExpression = (String) attrs.get("exceptionExpression" ); boolean hasExpression = StringUtils.hasText(exceptionExpression); if (includes.length == 0 ) { @SuppressWarnings ("unchecked" ) Class extends Throwable>[] value = (Class extends Throwable>[]) attrs.get("include" ); includes = value; } @SuppressWarnings ("unchecked" ) Class extends Throwable>[] excludes = (Class extends Throwable>[]) attrs.get("exclude" ); Integer maxAttempts = (Integer) attrs.get("maxAttempts" ); String maxAttemptsExpression = (String) attrs.get("maxAttemptsExpression" ); if (StringUtils.hasText(maxAttemptsExpression)) { maxAttempts = PARSER.parseExpression(resolve(maxAttemptsExpression), PARSER_CONTEXT) .getValue(this .evaluationContext, Integer.class ) ; } if (includes.length == 0 && excludes.length == 0 ) { SimpleRetryPolicy simple = hasExpression ? new ExpressionRetryPolicy(resolve(exceptionExpression)) .withBeanFactory(this .beanFactory) : new SimpleRetryPolicy(); simple.setMaxAttempts(maxAttempts); return simple; } Map, Boolean> policyMap = new HashMap, Boolean>(); for (Class extends Throwable> type : includes) { policyMap.put(type, true