Ryan的砖篮 2014-12-30
接上篇(http://ningandjiao.iteye.com/blog/2171185),
执行方式
HystrixCommand提供了3种执行方式:
同步执行:即一旦开始执行该命令,当前线程就得阻塞着直到该命令返回结果,然后才能继续执行下面的逻辑。当调用命令的execute()方法即为同步执行,示例:
@Test public void synchronousExecute() throws Exception { ThreadEchoCommand command = new ThreadEchoCommand("xianlinbox"); String result = command.execute(); assertThat(result,equalTo("Echo: xianlinbox")); }
异步执行:命令开始执行会返回一个Future<T>的对象,不阻塞后面的逻辑,开发者自己根据需要去获取结果。当调用HystrixCommand的queue()方法即为异步执行
@Test public void asynchronousExecute() throws Exception { ThreadEchoCommand command = new ThreadEchoCommand("xianlinbox"); Future<String> result = command.queue(); while (!result.isDone()){ System.out.println("Do other things ..."); } assertThat(result.get(),equalTo("Echo: xianlinbox")); }
响应式执行:命令开始执行会返回一个Observable<T>对象,开发者可以给给Obeservable对象注册上Observer或者Action1对象,响应式地处理命令执行过程中的不同阶段。当调用HystrixCommand的observe()方法,或使用Observable的工厂方法(just(),from())即为响应式执行,这个功能的实现是基于Netflix的另一个开源项目RxJava(https://github.com/Netflix/RxJava)来的,更细节的用法可以参考:https://github.com/Netflix/Hystrix/wiki/How-To-Use#wiki-Reactive-Execution。示例:
@Test public void reactiveExecute1() throws Exception { ThreadEchoCommand command1 = new ThreadEchoCommand("xianlinbox"); Observable<String> result = command1.observe(); result.subscribe(new Action1<String>() { @Override public void call(String s) { logger.info("Command called. Result is:{}", s); }1 }); Thread.sleep(1000); } @Test public void reactiveExecute2() throws Exception { ThreadEchoCommand command = new ThreadEchoCommand("xianlinbox"); Observable<String> result = command.observe(); result.subscribe(new Observer<String>() { @Override public void onCompleted() { logger.info("Command Completed"); } @Override public void onError(Throwable e) { logger.error("Command failled", e); } @Override public void onNext(String args) { logger.info("Command finished,result is {}", args); } }); Thread.sleep(1000); }
隔离方式(ThreadPool和Semaphores)
Hystrix支持2种隔离方式:
ThreadPool:即根据配置把不同的命令分配到不同的线程池中,这是比较常用的隔离策略,该策略的优点是隔离性好,并且可以配置断路,某个依赖被设置断路之后,系统不会再尝试新起线程运行它,而是直接提示失败,或返回fallback值;缺点是新起线程执行命令,在执行的时候必然涉及到上下文的切换,这会造成一定的性能消耗,但是Netflix做过实验,这种消耗对比其带来的价值是完全可以接受的,具体的数据参见HystrixWiki(https://github.com/Netflix/Hystrix/wiki/How-it-Works#wiki-Isolation)。本文前面的例子都是使用的TheadPool隔离策略。
Semaphores:信号量,顾名思义就是使用一个信号量来做隔离,开发者可以限制系统对某一个依赖的最高并发数。这个基本上就是一个限流的策略。每次调用依赖时都会检查一下是否到达信号量的限制值,如达到,则拒绝。该隔离策略的优点不新起线程执行命令,减少上下文切换,缺点是无法配置断路,每次都一定会去尝试获取信号量。示例:
public class SemaphoreEchoCommand extends HystrixCommand<String> { private Logger logger = LoggerFactory.getLogger(ThreadEchoCommand.class); private String input; protected SemaphoreEchoCommand(String input) { super(Setter.withGroupKey(HystrixCommandGroupKey.Factory.asKey("Semaphore Echo")) .andCommandKey(HystrixCommandKey.Factory.asKey("Echo")) .andCommandPropertiesDefaults(HystrixCommandProperties.Setter() .withExecutionIsolationStrategy(HystrixCommandProperties.ExecutionIsolationStrategy.SEMAPHORE) .withExecutionIsolationSemaphoreMaxConcurrentRequests(2))); this.input = input; } @Override protected String run() throws Exception { logger.info("Run command with input: {}", input); Thread.currentThread().sleep(100); return "Echo: " + input; } } @Test public void semaphoresCommandExecute() throws Exception { SemaphoreEchoCommand command = new SemaphoreEchoCommand("xianlinbox"); assertThat(command.execute(), equalTo("Echo: xianlinbox")); } @Test public void semaphoresCommandMultiExecute() throws Exception { for (int i = 0; i < 5; i++) { final SemaphoreEchoCommand command = new SemaphoreEchoCommand("xianlinbox-" + i); Thread thread = new Thread(new Runnable() { @Override public void run() { command.queue(); } }); thread.start(); } Thread.sleep(1000); }
第一个测试的运行日志如下:
23:10:34.996[main]INFOc.n.c.DynamicPropertyFactory-DynamicPropertyFactoryisinitializedwithconfigurationsources:com.netflix.config.ConcurrentCompositeConfiguration@2224df87
23:10:35.045[main]INFOd.ThreadEchoCommand-Runcommandwithinput:xianlinbox
从运行日志可以看到,HystrixCommand一样是在主线程中执行。
第二个测试的运行日志如下:
14:56:22.285[Thread-5]INFOd.ThreadEchoCommand-Runcommandwithinput:xianlinbox-4
14:56:22.285[Thread-1]INFOd.ThreadEchoCommand-Runcommandwithinput:xianlinbox-0
Exceptioninthread"Thread-2"Exceptioninthread"Thread-4"com.netflix.hystrix.exception.HystrixRuntimeException:Echocouldnotacquireasemaphoreforexecutionandnofallbackavailable.
示例中,设置的信号量最大值为2,因此可以看到有2个线程可以成功运行命令,第三个则会得到一个无法获取信号量的HystrixRuntimeException。
优雅降级
在调用第三方服务时,总是无可避免会出现一些错误(fail,timeout等),再加上上面提到的线程池大小,信号量的限制等等,在执行HystrixComamnd的过程中,总难免会抛出一些异常。而Hystrix为执行过程中的异常情况提供了优雅的降级方案,只需要在自己的HystrixCommand中实现getFallback()方法,当异常出现时,就会自动调用getFallback()方法的值.示例:为第一小节中的AddressHystrixCommand和ContactHystrixCommand添加getFallback()方法,当有异常发生的时候,直接返回null:
@Override protected Contact getFallback() { logger.info("Met error, using fallback value: {}", customerId); return null; }
然后,停掉Stub的Contact和Address服务,再次调用GetCustomer服务(http://localhost:8080/HystrixDemo/customers/1),得到结果如下:
{"id":"1","name":"xianlinbox","contact":null,"address":null}
运行日志:
15:22:08.847[hystrix-Contact-1]INFOc.x.h.d.ContactHystrixCommand-Getcontactforcustomer1
15:22:09.098[hystrix-Contact-1]INFOc.x.h.d.ContactHystrixCommand-Meterror,usingfallbackvalue:1
15:22:09.101[hystrix-Address-1]INFOc.x.h.d.AddressHystrixCommand-Getaddressforcustomer1
15:22:09.103[hystrix-Address-1]INFOc.x.h.d.AddressHystrixCommand-Meterror,usingfallbackvalue:1
请求作用域特性
作用域设置:要想要使用请求作用域特性,首先必须把HystrixCommand置于HystrixRequestContext的生命周期管理中,其典型用法是在Web应用中增加一个ServletFilter,把每个用户Request用HystrixRequestContext包起来。示例:
public class HystrixRequestContextServletFilter implements Filter { ... public void doFilter(ServletRequest request, ServletResponse response, FilterChain chain) throws IOException, ServletException { //启动HystrixRequestContext HystrixRequestContext context = HystrixRequestContext.initializeContext(); try { chain.doFilter(request, response); } finally { //关闭HystrixRequestContext context.shutdown(); } } }然后把该SevletFilter配置到web.xml中:
<filter> <display-name>HystrixRequestContextServletFilter</display-name> <filter-name>HystrixRequestContextServletFilter</filter-name> <filter-class>com.xianlinbox.hystrix.filter.HystrixRequestContextServletFilter</filter-class> </filter> <filter-mapping> <filter-name>HystrixRequestContextServletFilter</filter-name> <url-pattern>/*</url-pattern> </filter-mapping>
设置了请求作用域之后,接下来看看,我们从中可以得到哪些好处:
请求缓存(RequestCache):即当用户调用HystrixCommand时,HystrixCommand直接从缓存中取而不需要调用外部服务。HystrixCommand从缓存中取需要3个条件:
1.该HystrixCommand被包裹一个HystrixRequestContext中
2.该HystrixCommand实现了getCacheKey()方法
3.在HystrixRequestContext中已有相同CacheKey值的缓存
示例:
public void requestCache() throws Exception { HystrixRequestContext context = HystrixRequestContext.initializeContext(); try { ThreadEchoCommand command1 = new ThreadEchoCommand("xianlinbox"); ThreadEchoCommand command2 = new ThreadEchoCommand("xianlinbox"); assertThat(command1.execute(),equalTo("Echo: xianlinbox")); assertThat(command1.isResponseFromCache(),equalTo(false)); assertThat(command2.execute(),equalTo("Echo: xianlinbox")); assertThat(command2.isResponseFromCache(),equalTo(true)); } finally { context.shutdown(); } context = HystrixRequestContext.initializeContext(); try { ThreadEchoCommand command3 = new ThreadEchoCommand("xianlinbox"); assertThat(command3.execute(),equalTo("Echo: xianlinbox")); assertThat(command3.isResponseFromCache(),equalTo(false)); } finally { context.shutdown(); } }从上面的例子看以得到,一旦重新初始化了RequestContext,Cache也都全部失效了。另外,从Cache中获取值不会去执行HystrixCommand的run()方法。
除了重新初始化RequestContext,Hystrix还提供了另外一种方式来刷新Cache,该方式需要使用HystrixRequestCache的clear()方法,示例:在ThreadEchoCommand中实现一个静态方法flushCache(),该方法会调用HystrixRequestCache的clear方法清理Cache
public static void flushCache(String cacheKey) { HystrixRequestCache.getInstance(HystrixCommandKey.Factory.asKey("Echo"), HystrixConcurrencyStrategyDefault.getInstance()).clear(cacheKey); } @Test public void flushCacheTest() throws Exception { HystrixRequestContext context = HystrixRequestContext.initializeContext(); try { ThreadEchoCommand command1 = new ThreadEchoCommand("xianlinbox"); ThreadEchoCommand command2 = new ThreadEchoCommand("xianlinbox"); assertThat(command1.execute(), equalTo("Echo: xianlinbox")); assertThat(command1.isResponseFromCache(), equalTo(false)); assertThat(command2.execute(), equalTo("Echo: xianlinbox")); assertThat(command2.isResponseFromCache(), equalTo(true)); ThreadEchoCommand.flushCache("xianlinbox"); ThreadEchoCommand command3 = new ThreadEchoCommand("xianlinbox"); assertThat(command3.execute(), equalTo("Echo: xianlinbox")); assertThat(command3.isResponseFromCache(), equalTo(false)); } finally { context.shutdown(); } }
通过这个机制,开发者可以实现Get-Set-Get的Cache验证机制,防止因为Cache导致的不一致状况。
批量执行请求(RequestCollapsing):即用户可以把多个命令封装到一个HystrixCommand中执行以提升效率,这多个命令会在一个线程中依次执行(注:经笔者测试,在JDK6下线程数固定,但是在JDK7下的运行线程数不固定)。要使用该特性需要把依赖调用封装到一个HystrixCollapser<BatchReturnType,ResponseType,RequestArgumentType>中,该抽象类的主要作用有3个:
1.把所有的依赖调用封装到一个CollapseRequest的集合中
2.以第一步得到的CollapseRequest集合为参数创建一个HystrixCommand
3.把第二步得到的结果集一一对应的设置到对应的CollapseRequest中
为了支持上面的功能,该抽象类提供了3个泛型参数:
BatchReturnType:即BatchCommand的返回值,通常为ResponseType的集合。
ResponseType:依赖调用的返回值。
RequestArgumentType:依赖调用的参数,如果有多个参数,需封装为一个对象或使用集合。
示例:
public class CollapseEchoHystrixCommand extends HystrixCollapser<List<String>, String, String> { private Logger logger = LoggerFactory.getLogger(CollapseEchoHystrixCommand.class); private String input; public CollapseEchoHystrixCommand(String input) { super(HystrixCollapser.Setter .withCollapserKey(HystrixCollapserKey.Factory.asKey("Echo Collapse"))); this.input = input; } @Override public String getRequestArgument() { return input; } @Override protected HystrixCommand<List<String>> createCommand(Collection<CollapsedRequest<String, String>> collapsedRequests) { return new BatchCommand(collapsedRequests); } @Override protected void mapResponseToRequests(List<String> batchResponse, Collection<CollapsedRequest<String, String>> collapsedRequests) { logger.info("Mapping response to Request"); int count = 0; for (CollapsedRequest<String, String> request : collapsedRequests) { request.setResponse(batchResponse.get(count++)); } } private class BatchCommand extends HystrixCommand<List<String>> { private Collection<CollapsedRequest<String, String>> requests; public BatchCommand(Collection<CollapsedRequest<String, String>> requests) { super(HystrixCommandGroupKey.Factory.asKey("Batch")); this.requests = requests; } @Override protected List<String> run() throws Exception { logger.info("Run batch command"); List<String> responses = new ArrayList<String>(); for (CollapsedRequest<String, String> request : requests) { logger.info("Run request: {}", request.getArgument()); responses.add("Echo: " + request.getArgument()); } return responses; } } } @Test public void collapseCommandTest() throws Exception { HystrixRequestContext context = HystrixRequestContext.initializeContext(); try { Future<String> result1 = new CollapseEchoHystrixCommand("xianlinbox-1").queue(); Future<String> result2 = new CollapseEchoHystrixCommand("xianlinbox-2").queue(); Future<String> result3 = new CollapseEchoHystrixCommand("xianlinbox-3").queue(); assertThat(result1.get(),equalTo("Echo: xianlinbox-1")); assertThat(result2.get(),equalTo("Echo: xianlinbox-2")); assertThat(result3.get(),equalTo("Echo: xianlinbox-3")); assertEquals(1, HystrixRequestLog.getCurrentRequest().getExecutedCommands().size()); } finally { context.shutdown(); } }
运行日志:
03:10:58.584[main]INFOd.CollapseEchoHystrixCommand-Getargument
03:10:58.597[main]INFOd.CollapseEchoHystrixCommand-Getargument
03:10:58.597[main]INFOd.CollapseEchoHystrixCommand-Getargument
03:10:58.598[HystrixTimer-1]INFOd.CollapseEchoHystrixCommand-Createbatchcommand
03:10:58.637[hystrix-Batch-1]INFOd.CollapseEchoHystrixCommand-Runbatchcommand
03:10:58.637[hystrix-Batch-1]INFOd.CollapseEchoHystrixCommand-Runrequest:xianlinbox-1
03:10:58.639[hystrix-Batch-1]INFOd.CollapseEchoHystrixCommand-Runrequest:xianlinbox-2
03:10:58.639[hystrix-Batch-1]INFOd.CollapseEchoHystrixCommand-Runrequest:xianlinbox-3
03:10:58.644[RxComputationThreadPool-1]INFOd.CollapseEchoHystrixCommand-MappingresponsetoRequest
从运行日志可以看到,整个Collapser的运行过程:
1.获取调用参数,封装到CollapseRequest中
2.以封装后的List<CollapseRequest>为参数创建BatchHystrixComand
3.BatchHystrixCommand运行所有的请求,把所有的返回放到List<Response>中
4.把Response设置到对应的CollapseRequest中,返回给调用者。