SpringCloud Eureka源码分析

Eureka核心功能点

服务注册(register)：Eureka Client会通过发送REST请求的方式向Eureka Server注册自己的服务，提供自身的元数 据，比如ip地址、端口、运行状况指标的url、主页地址等信息。Eureka Server接收到注册请求后，就会把这些元数 据信息存储在一个双层的Map中。

 

服务续约(renew)：在服务注册后，Eureka Client会维护一个心跳来持续通知Eureka Server，说明服务一直处于可 用状态，防止被剔除。Eureka Client在默认的情况下会每隔30秒。

(eureka.instance.leaseRenewallIntervalInSeconds)发送一次心跳来进行服务续约。

 

服务同步(replicate)：Eureka Server之间会互相进行注册，构建Eureka Server集群，不同Eureka Server之间会进 行服务同步，用来保证服务信息的一致性。

获取服务(get registry)：服务消费者（Eureka Client）在启动的时候，会发送一个REST请求给Eureka Server，获 取上面注册的服务清单，并且缓存在Eureka Client本地，默认缓存30秒 (eureka.client.registryFetchIntervalSeconds)。同时，为了性能考虑，Eureka Server也会维护一份只读的服务清 单缓存，该缓存每隔30秒更新一次。

服务调用：服务消费者在获取到服务清单后，就可以根据清单中的服务列表信息，查找到其他服务的地址，从而进行 远程调用。Eureka有Region和Zone的概念，一个Region可以包含多个Zone，在进行服务调用时，优先访问处于同 一个Zone中的服务提供者。

 

服务下线(cancel)：当Eureka Client需要关闭或重启时，就不希望在这个时间段内再有请求进来，所以，就需要提前 先发送REST请求给Eureka Server，告诉Eureka Server自己要下线了，Eureka Server在收到请求后，就会把该服务 状态置为下线（DOWN），并把该下线事件传播出去。

 

服务剔除(evict)：有时候，服务实例可能会因为网络故障等原因导致不能提供服务，而此时该实例也没有发送请求给 Eureka Server来进行服务下线，所以，还需要有服务剔除的机制。Eureka Server在启动的时候会创建一个定时任 务，每隔一段时间（默认60秒），从当前服务清单中把超时没有续约（默认90秒， eureka.instance.leaseExpirationDurationInSeconds）的服务剔除。

 

自我保护：既然Eureka Server会定时剔除超时没有续约的服务，那就有可能出现一种场景，网络一段时间内发生了 异常，所有的服务都没能够进行续约，Eureka Server就把所有的服务都剔除了，这样显然不太合理。所以，就有了 自我保护机制，当短时间内，统计续约失败的比例，如果达到一定阈值，则会触发自我保护的机制，在该机制下， Eureka Server不会剔除任何的微服务，等到正常后，再退出自我保护机制。自我保护开关(eureka.server.enableself-preservation: false)

 

Eureka Server端源码分析

@Configuration
2 @Import(EurekaServerInitializerConfiguration.class)
3 @ConditionalOnBean(EurekaServerMarkerConfiguration.Marker.class)
4 @EnableConfigurationProperties({ EurekaDashboardProperties.class,
5 InstanceRegistryProperties.class })
6 @PropertySource("classpath:/eureka/server.properties")
7 public class EurekaServerAutoConfiguration extends WebMvcConfigurerAdapter {
8
9 // 此处省略大部分代码，仅抽取一些关键的代码片段
10
11 // 加载EurekaController, spring‐cloud 提供了一些额外的接口，用来获取eurekaServer的信息
12 @Bean
13 @ConditionalOnProperty(prefix = "eureka.dashboard", name = "enabled", matchIfMissing = true)
14 public EurekaController eurekaController() {
15 return new EurekaController(this.applicationInfoManager);
16 }
17
18 //初始化集群注册表
19 @Bean
20 public PeerAwareInstanceRegistry peerAwareInstanceRegistry(
21 ServerCodecs serverCodecs) {
22 this.eurekaClient.getApplications(); // force initialization
23 return new InstanceRegistry(this.eurekaServerConfig, this.eurekaClientConfig,
24 serverCodecs, this.eurekaClient,
25 this.instanceRegistryProperties.getExpectedNumberOfRenewsPerMin(),
26 this.instanceRegistryProperties.getDefaultOpenForTrafficCount());
27 }
28
29 // 配置服务节点信息，这里的作用主要是为了配置Eureka的peer节点，也就是说当有收到有节点注册上来
30 //的时候，需要通知给那些服务节点， （互为一个集群）
31 @Bean
32 @ConditionalOnMissingBean
33 public PeerEurekaNodes peerEurekaNodes(PeerAwareInstanceRegistry registry,
34 ServerCodecs serverCodecs) {
35 return new PeerEurekaNodes(registry, this.eurekaServerConfig,
36 this.eurekaClientConfig, serverCodecs, this.applicationInfoManager);
37 }
38 // EurekaServer的上下文
39 @Bean
40 public EurekaServerContext eurekaServerContext(ServerCodecs serverCodecs,
41 PeerAwareInstanceRegistry registry, PeerEurekaNodes peerEurekaNodes) {
42 return new DefaultEurekaServerContext(this.eurekaServerConfig, serverCodecs,
43 registry, peerEurekaNodes, this.applicationInfoManager);
44 }
45 // 这个类的作用是spring‐cloud和原生eureka的胶水代码，通过这个类来启动EurekaSever
46 // 后面这个类会在EurekaServerInitializerConfiguration被调用，进行eureka启动
47 @Bean
48 public EurekaServerBootstrap eurekaServerBootstrap(PeerAwareInstanceRegistry registry,
49 EurekaServerContext serverContext) {
50 return new EurekaServerBootstrap(this.applicationInfoManager,
51 this.eurekaClientConfig, this.eurekaServerConfig, registry,
52 serverContext);
53 }
54 // 配置拦截器，ServletContainer里面实现了jersey框架，通过他来实现eurekaServer对外的restFull接口
55 @Bean
56 public FilterRegistrationBean jerseyFilterRegistration(
57 javax.ws.rs.core.Application eurekaJerseyApp) {
58 FilterRegistrationBean bean = new FilterRegistrationBean();
59 bean.setFilter(new ServletContainer(eurekaJerseyApp));
60 bean.setOrder(Ordered.LOWEST_PRECEDENCE);
61 bean.setUrlPatterns(
62 Collections.singletonList(EurekaConstants.DEFAULT_PREFIX + "/*"));
63
64 return bean;
65 }
66 }

EurekaServerAutoConfiguration会导入EurekaServerInitializerConfiguration

/**
2 * @author Dave Syer
3 */
4 @Configuration
5 @CommonsLog
6 public class EurekaServerInitializerConfiguration
7 implements ServletContextAware, SmartLifecycle, Ordered {
8
9 @Autowired
10 private EurekaServerConfig eurekaServerConfig;
11
12 private ServletContext servletContext;
13
14 @Autowired
15 private ApplicationContext applicationContext;
16
17 @Autowired
18 private EurekaServerBootstrap eurekaServerBootstrap;
19
20 private boolean running;
21
22 private int order = 1;
23
24 @Override
25 public void setServletContext(ServletContext servletContext) {
26 this.servletContext = servletContext;
27 }
28
29 @Override
30 public void start() {
31 // 启动一个线程
32 new Thread(new Runnable() {
33 @Override
34 public void run() {
35 try {
36 //初始化EurekaServer，同时启动Eureka Server
37 eurekaServerBootstrap.contextInitialized(EurekaServerInitializerConfiguration.this.servletConte
xt);
38 log.info("Started Eureka Server");
39 // 发布EurekaServer的注册事件
40 publish(new EurekaRegistryAvailableEvent(getEurekaServerConfig()));
41 // 设置启动的状态为true
42 EurekaServerInitializerConfiguration.this.running = true;
43 // 发送Eureka Start 事件 ， 其他还有各种事件，我们可以监听这种时间，然后做一些特定的业务需求
44 publish(new EurekaServerStartedEvent(getEurekaServerConfig()));
45 }
46 catch (Exception ex) {
47 // Help!
48 log.error("Could not initialize Eureka servlet context", ex);
49 }
50 }
51 }).start();
52 }
53
54 private EurekaServerConfig getEurekaServerConfig() {
55 return this.eurekaServerConfig;
56 }
57
58 private void publish(ApplicationEvent event) {
59 this.applicationContext.publishEvent(event);
60 }
61
62 @Override
63 public void stop() {
64 this.running = false;
65 eurekaServerBootstrap.contextDestroyed(this.servletContext);
66 }
67
68 @Override
69 public boolean isRunning() {
70 return this.running;
71 }
72
73 @Override
74 public int getPhase() {
75 return 0;
76 }
77
78 @Override
79 public boolean isAutoStartup() {
80 return true;
81 }
82
83 @Override
84 public void stop(Runnable callback) {
85 callback.run();
86 }
87
88 @Override
89 public int getOrder() {
90 return this.order;
91 }
92
93 }

 

EurekaServerBootstrap的contextInitialized初始化方法：

//初始化EurekaServer的运行环境和上下文
2 public void contextInitialized(ServletContext context) {
3 try {
4 initEurekaEnvironment();
5 initEurekaServerContext();
6
7 context.setAttribute(EurekaServerContext.class.getName(), this.serverContext);
8 }
9 catch (Throwable e) {
10 log.error("Cannot bootstrap eureka server :", e);
11 throw new RuntimeException("Cannot bootstrap eureka server :", e);
12 }
13 }
14
15 初始化EurekaServer的上下文
16 protected void initEurekaServerContext() throws Exception {
17 // For backward compatibility
18 JsonXStream.getInstance().registerConverter(new V1AwareInstanceInfoConverter(),
19 XStream.PRIORITY_VERY_HIGH);
20 XmlXStream.getInstance().registerConverter(new V1AwareInstanceInfoConverter(),
21 XStream.PRIORITY_VERY_HIGH);
22
23 if (isAws(this.applicationInfoManager.getInfo())) {
24 this.awsBinder = new AwsBinderDelegate(this.eurekaServerConfig,
25 this.eurekaClientConfig, this.registry, this.applicationInfoManager);
26 this.awsBinder.start();
27 }
28
29 //初始化eureka server上下文
30 EurekaServerContextHolder.initialize(this.serverContext);
31
32 log.info("Initialized server context");
33
34 // Copy registry from neighboring eureka node
35 // 从相邻的eureka节点复制注册表
36 int registryCount = this.registry.syncUp();
37 // 默认每30秒发送心跳，1分钟就是2次
38 // 修改eureka状态为up
39 // 同时，这里面会开启一个定时任务，用于清理60秒没有心跳的客户端。自动下线
40 this.registry.openForTraffic(this.applicationInfoManager, registryCount);
41
42 // Register all monitoring statistics.
43 EurekaMonitors.registerAllStats();
44 }
45
46 @Override
47 public int syncUp() {
48 // Copy entire entry from neighboring DS node
49 int count = 0;
50
51 for (int i = 0; ((i < serverConfig.getRegistrySyncRetries()) && (count == 0)); i++) {
52 if (i > 0) {
53 try {
54 Thread.sleep(serverConfig.getRegistrySyncRetryWaitMs());
55 } catch (InterruptedException e) {
56 logger.warn("Interrupted during registry transfer..");
57 break;
58 }
59 }
60 Applications apps = eurekaClient.getApplications();
61 for (Application app : apps.getRegisteredApplications()) {
62 for (InstanceInfo instance : app.getInstances()) {
63 try {
64 if (isRegisterable(instance)) {
65 //将其他节点的实例注册到本节点
66 register(instance, instance.getLeaseInfo().getDurationInSecs(), true);
67 count++;
68 }
69 } catch (Throwable t) {
70 logger.error("During DS init copy", t);
71 }
72 }
73 }
74 }
75 return count;
76 }
77
78 @Override
79 public void openForTraffic(ApplicationInfoManager applicationInfoManager, int count) {
80 // Renewals happen every 30 seconds and for a minute it should be a factor of 2.
81 // 计算每分钟最大续约数
82 this.expectedNumberOfRenewsPerMin = count * 2;
83 // 每分钟最小续约数
84 this.numberOfRenewsPerMinThreshold =
85 (int) (this.expectedNumberOfRenewsPerMin * serverConfig.getRenewalPercentThreshold());
86 logger.info("Got " + count + " instances from neighboring DS node");
87 logger.info("Renew threshold is: " + numberOfRenewsPerMinThreshold);
88 this.startupTime = System.currentTimeMillis();
89 if (count > 0) {
90 this.peerInstancesTransferEmptyOnStartup = false;
91 }
92 DataCenterInfo.Name selfName = applicationInfoManager.getInfo().getDataCenterInfo().getName();
93 boolean isAws = Name.Amazon == selfName;
94 if (isAws && serverConfig.shouldPrimeAwsReplicaConnections()) {
95 logger.info("Priming AWS connections for all replicas..");
96 primeAwsReplicas(applicationInfoManager);
97 }
98 logger.info("Changing status to UP");
99 // 设置实例的状态为UP
100 applicationInfoManager.setInstanceStatus(InstanceStatus.UP);
101 // 开启定时任务，默认60秒执行一次，用于清理60秒之内没有续约的实例
102 super.postInit();
103 }
104
105 protected void postInit() {
106 renewsLastMin.start();
107 if (evictionTaskRef.get() != null) {
108 evictionTaskRef.get().cancel();
109 }
110 evictionTaskRef.set(new EvictionTask());
111 //服务剔除任务
112 evictionTimer.schedule(evictionTaskRef.get(),
113 serverConfig.getEvictionIntervalTimerInMs(),
114 serverConfig.getEvictionIntervalTimerInMs());
115 }

 

从上面的EurekaServerAutoConfiguration类，我们可以看到有个初始化EurekaServerContext的方 法

@Bean
2 public EurekaServerContext eurekaServerContext(ServerCodecs serverCodecs,
3 PeerAwareInstanceRegistry registry, PeerEurekaNodes peerEurekaNodes) {
4 return new DefaultEurekaServerContext(this.eurekaServerConfig, serverCodecs,
5 registry, peerEurekaNodes, this.applicationInfoManager);
6 }

DefaultEurekaServerContext 这个类里面的的initialize()方法是被@PostConstruct 这个注解修饰的, 在应用加载的时候，会执行这个方法

1 public void initialize() throws Exception {
2 logger.info("Initializing ...");
3 // 启动一个线程，读取其他集群节点的信息，后面后续复制
4 peerEurekaNodes.start();
5 //
6 registry.init(peerEurekaNodes);
7 logger.info("Initialized");
8 }

 

peerEurekaNodes.start()主要是启动一个只拥有一个线程的线程池，第一次进去会更新一下集群其他节点信息 然后启动了一个定时线程，每60秒更新一次，也就是说后续可以根据配置动态的修改节点配置。（原生的spring cloud config支持）

1 public void start() {
2 taskExecutor = Executors.newSingleThreadScheduledExecutor(
3 new ThreadFactory() {
4 @Override
5 public Thread newThread(Runnable r) {
6 Thread thread = new Thread(r, "Eureka‐PeerNodesUpdater");
7 thread.setDaemon(true);
8 return thread;
9 }
10 }
11 );
12 try {
13 // 首次进来，更新集群节点信息
14 updatePeerEurekaNodes(resolvePeerUrls());
15 // 搞个线程
16 Runnable peersUpdateTask = new Runnable() {
17 @Override
18 public void run() {
19 try {
20 updatePeerEurekaNodes(resolvePeerUrls());
21 } catch (Throwable e) {
22 logger.error("Cannot update the replica Nodes", e);
23 }
24
25 }
26 };
27 taskExecutor.scheduleWithFixedDelay(
28 peersUpdateTask,
29 serverConfig.getPeerEurekaNodesUpdateIntervalMs(),
30 serverConfig.getPeerEurekaNodesUpdateIntervalMs(),
31 TimeUnit.MILLISECONDS
32 );
33 } catch (Exception e) {
34 throw new IllegalStateException(e);
35 }
36 for (PeerEurekaNode node : peerEurekaNodes) {
37 logger.info("Replica node URL: " + node.getServiceUrl());
38 }
39 }
40 // 根据URL 构建PeerEurekaNode信息
41 protected PeerEurekaNode createPeerEurekaNode(String peerEurekaNodeUrl) {
42 HttpReplicationClient replicationClient = JerseyReplicationClient.createReplicationClient(serve
rConfig, serverCodecs, peerEurekaNodeUrl);
43 String targetHost = hostFromUrl(peerEurekaNodeUrl);
44 if (targetHost == null) {
45 targetHost = "host";
46 }
47 return new PeerEurekaNode(registry, targetHost, peerEurekaNodeUrl, replicationClient, serverCon
fig);
48 }

 

Eureka Client端源码分析

client初始化

1 @Inject
2 DiscoveryClient(ApplicationInfoManager applicationInfoManager, EurekaClientConfig config, Abstra
ctDiscoveryClientOptionalArgs args,
3 Provider<BackupRegistry> backupRegistryProvider) {
4 //省略非关键代码。。。
5
6 logger.info("Initializing Eureka in region {}", clientConfig.getRegion());
7
8 //省略非关键代码。。。
9
10 try {
11 // default size of 2 ‐ 1 each for heartbeat and cacheRefresh
12 scheduler = Executors.newScheduledThreadPool(2,
13 new ThreadFactoryBuilder()
14 .setNameFormat("DiscoveryClient‐%d")
15 .setDaemon(true)
16 .build());
17
18 heartbeatExecutor = new ThreadPoolExecutor(
19 1, clientConfig.getHeartbeatExecutorThreadPoolSize(), 0, TimeUnit.SECONDS,
20 new SynchronousQueue<Runnable>(),
21 new ThreadFactoryBuilder()
22 .setNameFormat("DiscoveryClient‐HeartbeatExecutor‐%d")
23 .setDaemon(true)
24 .build()
25 ); // use direct handoff
26
27 cacheRefreshExecutor = new ThreadPoolExecutor(
28 1, clientConfig.getCacheRefreshExecutorThreadPoolSize(), 0, TimeUnit.SECONDS,
29 new SynchronousQueue<Runnable>(),
30 new ThreadFactoryBuilder()
31 .setNameFormat("DiscoveryClient‐CacheRefreshExecutor‐%d")
32 .setDaemon(true)
33 .build()
34 ); // use direct handoff
35
36 eurekaTransport = new EurekaTransport();
37 scheduleServerEndpointTask(eurekaTransport, args);
38
39 AzToRegionMapper azToRegionMapper;
40 if (clientConfig.shouldUseDnsForFetchingServiceUrls()) {
41 azToRegionMapper = new DNSBasedAzToRegionMapper(clientConfig);
42 } else {
43 azToRegionMapper = new PropertyBasedAzToRegionMapper(clientConfig);
44 }
45 if (null != remoteRegionsToFetch.get()) {
46 azToRegionMapper.setRegionsToFetch(remoteRegionsToFetch.get().split(","));
47 }
48 instanceRegionChecker = new InstanceRegionChecker(azToRegionMapper, clientConfig.getRegion());
49 } catch (Throwable e) {
50 throw new RuntimeException("Failed to initialize DiscoveryClient!", e);
51 }
52
53 if (clientConfig.shouldFetchRegistry() && !fetchRegistry(false)) {
54 fetchRegistryFromBackup();
55 }
56
57 // call and execute the pre registration handler before all background tasks (inc registration)
is started
58 if (this.preRegistrationHandler != null) {
59 this.preRegistrationHandler.beforeRegistration();
60 }
61
62 if (clientConfig.shouldRegisterWithEureka() && clientConfig.shouldEnforceRegistrationAtInit()) {
63 try {
64 if (!register() ) {
65 throw new IllegalStateException("Registration error at startup. Invalid server response.");
66 }
67 } catch (Throwable th) {
68 logger.error("Registration error at startup: {}", th.getMessage());
69 throw new IllegalStateException(th);
70 }
71 }
72
73 //最核心代码
74 // finally, init the schedule tasks (e.g. cluster resolvers, heartbeat, instanceInfo replicator,
fetch
75 initScheduledTasks();
76
77 try {
78 Monitors.registerObject(this);
79 } catch (Throwable e) {
80 logger.warn("Cannot register timers", e);
81 }
82
83 // This is a bit of hack to allow for existing code using DiscoveryManager.getInstance()
84 // to work with DI'd DiscoveryClient
85 DiscoveryManager.getInstance().setDiscoveryClient(this);
86 DiscoveryManager.getInstance().setEurekaClientConfig(config);
87
88 initTimestampMs = System.currentTimeMillis();
89 logger.info("Discovery Client initialized at timestamp {} with initial instances count: {}",
90 initTimestampMs, this.getApplications().size());
91 }

 

初始化时启动核心功能定时任务

1 private void initScheduledTasks() {
2 //获取服务注册列表信息
3 if (clientConfig.shouldFetchRegistry()) {
4 //服务注册列表更新的周期时间
5 int registryFetchIntervalSeconds = clientConfig.getRegistryFetchIntervalSeconds();
6 int expBackOffBound = clientConfig.getCacheRefreshExecutorExponentialBackOffBound();
7 //定时更新服务注册列表
8 scheduler.schedule(
9 new TimedSupervisorTask(
10 "cacheRefresh",
11 scheduler,
12 cacheRefreshExecutor,
13 registryFetchIntervalSeconds,
14 TimeUnit.SECONDS,
15 expBackOffBound,
16 new CacheRefreshThread() //该线程执行更新的具体逻辑
17 ),
18 registryFetchIntervalSeconds, TimeUnit.SECONDS);
19 }
20
21 if (clientConfig.shouldRegisterWithEureka()) {
22 //服务续约的周期时间
23 int renewalIntervalInSecs = instanceInfo.getLeaseInfo().getRenewalIntervalInSecs();
24 int expBackOffBound = clientConfig.getHeartbeatExecutorExponentialBackOffBound();
25 //应用启动可见此日志，内容是：Starting heartbeat executor: renew interval is: 30
26 logger.info("Starting heartbeat executor: " + "renew interval is: " + renewalIntervalInSecs);
27 // 服务定时续约
28 scheduler.schedule(
29 new TimedSupervisorTask(
30 "heartbeat",
31 scheduler,
32 heartbeatExecutor,
33 renewalIntervalInSecs,
34 TimeUnit.SECONDS,
35 expBackOffBound,
36 new HeartbeatThread() //该线程执行续约的具体逻辑
37 ),
38 renewalIntervalInSecs, TimeUnit.SECONDS);
39
40 //这个Runable中含有服务注册的逻辑
41 instanceInfoReplicator = new InstanceInfoReplicator(
42 this,
43 instanceInfo,
44 clientConfig.getInstanceInfoReplicationIntervalSeconds(),
45 2); // burstSize
46
47 statusChangeListener = new ApplicationInfoManager.StatusChangeListener() {
48 @Override
49 public String getId() {
50 return "statusChangeListener";
51 }
52
53 @Override
54 public void notify(StatusChangeEvent statusChangeEvent) {
55 if (InstanceStatus.DOWN == statusChangeEvent.getStatus() ||
56 InstanceStatus.DOWN == statusChangeEvent.getPreviousStatus()) {
57 // log at warn level if DOWN was involved
58 logger.warn("Saw local status change event {}", statusChangeEvent);
59 } else {
60 logger.info("Saw local status change event {}", statusChangeEvent);
61 }
62 instanceInfoReplicator.onDemandUpdate();
63 }
64 };
65
66 if (clientConfig.shouldOnDemandUpdateStatusChange()) {
67 applicationInfoManager.registerStatusChangeListener(statusChangeListener);
68 }
69 //服务注册
70 instanceInfoReplicator.start(clientConfig.getInitialInstanceInfoReplicationIntervalSeconds());
71 } else {
72 logger.info("Not registering with Eureka server per configuration");
73 }
74 }

 

TimedSupervisorTask是一个Runnable接口实现，看下它的run方法

1 @Override
2 public void run() {
3 Future<?> future = null;
4 try {
5 future = executor.submit(task);
6 threadPoolLevelGauge.set((long) executor.getActiveCount());
7 //指定等待子线程的最长时间
8 future.get(timeoutMillis, TimeUnit.MILLISECONDS); // block until done or timeout
9 //delay是个关键变量，后面会用到，这里记得每次执行任务成功都会将delay重置
10 delay.set(timeoutMillis);
11 threadPoolLevelGauge.set((long) executor.getActiveCount());
12 } catch (TimeoutException e) {
13 logger.warn("task supervisor timed out", e);
14 timeoutCounter.increment();
15
16 long currentDelay = delay.get();
17 //任务线程超时的时候，就把delay变量翻倍，但不会超过外部调用时设定的最大延时时间
18 long newDelay = Math.min(maxDelay, currentDelay * 2);
19 //设置为最新的值，考虑到多线程，所以用了CAS
20 delay.compareAndSet(currentDelay, newDelay);
21 } catch (RejectedExecutionException e) {
22 //一旦线程池的阻塞队列中放满了待处理任务，触发了拒绝策略，就会将调度器停掉
23 if (executor.isShutdown() || scheduler.isShutdown()) {
24 logger.warn("task supervisor shutting down, reject the task", e);
25 } else {
26 logger.warn("task supervisor rejected the task", e);
27 }
28
29 rejectedCounter.increment();
30 } catch (Throwable e) {
31 if (executor.isShutdown() || scheduler.isShutdown()) {
32 logger.warn("task supervisor shutting down, can't accept the task");
33 } else {
34 logger.warn("task supervisor threw an exception", e);
35 }
36
37 throwableCounter.increment();
38 } finally {
39 //这里任务要么执行完毕，要么发生异常，都用cancel方法来清理任务；
40 if (future != null) {
41 future.cancel(true);
42 }
43 //只要调度器没有停止，就再指定等待时间之后在执行一次同样的任务
44 if (!scheduler.isShutdown()) {
45 //假设外部调用时传入的超时时间为30秒（构造方法的入参timeout），最大间隔时间为50秒(构造方法的入参expBac
kOffBound)
46 //如果最近一次任务没有超时，那么就在30秒后开始新任务，
47 //如果最近一次任务超时了，那么就在50秒后开始新任务（异常处理中有个乘以二的操作，乘以二后的60秒超过了最大
间隔50秒）
48 scheduler.schedule(this, delay.get(), TimeUnit.MILLISECONDS);
49 }
50 }
51 }

 

scheduler.schedule(this, delay.get(), TimeUnit.MILLISECONDS)，从代码注释上可以看出这个方法是一次性调用 方法，但是实际上这个方法执行的任务会反复执行，秘密就在this对应的这个类TimedSupervisorTask的run方法 里，run方法任务执行完最后，会再次调用schedule方法，在指定的时间之后执行一次相同的任务，这个间隔时间和 最近一次任务是否超时有关，如果超时了则下一次执行任务的间隔时间就会变大；

 

源码精髓：

从整体上看，TimedSupervisorTask是固定间隔的周期性任务，一旦遇到超时就会将下一个周期的间隔时间调大，如 果连续超时，那么每次间隔时间都会增大一倍，一直到达外部参数设定的上限为止，一旦新任务不再超时，间隔时间 又会自动恢复为初始值，另外还有CAS来控制多线程同步，这些是我们看源码需要学习到的设计技巧

 

定时更新服务注册列表线程CacheRefreshThread

1 /**
2 * The task that fetches the registry information at specified intervals.
3 *
4 */
5 class CacheRefreshThread implements Runnable {
6 public void run() {
7 refreshRegistry();
8 }
9 }
10
11 @VisibleForTesting
12 void refreshRegistry() {
13 try {
14 boolean isFetchingRemoteRegionRegistries = isFetchingRemoteRegionRegistries();
15
16 boolean remoteRegionsModified = false;
17 // This makes sure that a dynamic change to remote regions to fetch is honored.
18 String latestRemoteRegions = clientConfig.fetchRegistryForRemoteRegions();
19 //不做aws环境的配置这个if逻辑不会执行
20 if (null != latestRemoteRegions) {
21 String currentRemoteRegions = remoteRegionsToFetch.get();
22 if (!latestRemoteRegions.equals(currentRemoteRegions)) {
23 // Both remoteRegionsToFetch and AzToRegionMapper.regionsToFetch need to be in sync
24 synchronized (instanceRegionChecker.getAzToRegionMapper()) {
25 if (remoteRegionsToFetch.compareAndSet(currentRemoteRegions, latestRemoteRegions)) {
26 String[] remoteRegions = latestRemoteRegions.split(",");
27 remoteRegionsRef.set(remoteRegions);
28 instanceRegionChecker.getAzToRegionMapper().setRegionsToFetch(remoteRegions);
29 remoteRegionsModified = true;
30 } else {
31 logger.info("Remote regions to fetch modified concurrently," +
32 " ignoring change from {} to {}", currentRemoteRegions, latestRemoteRegions);
33 }
34 }
35 } else {
36 // Just refresh mapping to reflect any DNS/Property change
37 instanceRegionChecker.getAzToRegionMapper().refreshMapping();
38 }
39 }
40
41 //获取注册信息方法
42 boolean success = fetchRegistry(remoteRegionsModified);
43 if (success) {
44 registrySize = localRegionApps.get().size();
45 lastSuccessfulRegistryFetchTimestamp = System.currentTimeMillis();
46 }
47
48 //省略非关键代码。。。
49 } catch (Throwable e) {
50 logger.error("Cannot fetch registry from server", e);
51 }
52 }
53
54 private boolean fetchRegistry(boolean forceFullRegistryFetch) {
55 Stopwatch tracer = FETCH_REGISTRY_TIMER.start();
56
57 try {
58 // If the delta is disabled or if it is the first time, get all
59 // applications
60 // 取出本地缓存之前获取的服务列表信息
61 Applications applications = getApplications();
62
63 //判断多个条件，确定是否触发全量更新，如下任一个满足都会全量更新：
64 //1. 是否禁用增量更新；
65 //2. 是否对某个region特别关注；
66 //3. 外部调用时是否通过入参指定全量更新；
67 //4. 本地还未缓存有效的服务列表信息；
68 if (clientConfig.shouldDisableDelta()
69 || (!Strings.isNullOrEmpty(clientConfig.getRegistryRefreshSingleVipAddress()))
70 || forceFullRegistryFetch
71 || (applications == null)
72 || (applications.getRegisteredApplications().size() == 0)
73 || (applications.getVersion() == ‐1)) //Client application does not have latest library support
ing delta
74 {
75 logger.info("Disable delta property : {}", clientConfig.shouldDisableDelta());
76 logger.info("Single vip registry refresh property : {}", clientConfig.getRegistryRefreshSingleV
ipAddress());
77 logger.info("Force full registry fetch : {}", forceFullRegistryFetch);
78 logger.info("Application is null : {}", (applications == null));
79 logger.info("Registered Applications size is zero : {}",
80 (applications.getRegisteredApplications().size() == 0));
81 logger.info("Application version is ‐1: {}", (applications.getVersion() == ‐1));
82 //全量更新
83 getAndStoreFullRegistry();
84 } else {
85 //增量更新
86 getAndUpdateDelta(applications);
87 }
88 //重新计算和设置一致性hash码
89 applications.setAppsHashCode(applications.getReconcileHashCode());
90 logTotalInstances();
91 } catch (Throwable e) {
92 logger.error(PREFIX + "{} ‐ was unable to refresh its cache! status = {}", appPathIdentifier,
e.getMessage(), e);
93 return false;
94 } finally {
95 if (tracer != null) {
96 tracer.stop();
97 }
98 }
99
100 // Notify about cache refresh before updating the instance remote status
101 //将本地缓存更新的事件广播给所有已注册的监听器，注意该方法已被CloudEurekaClient类重写
102 onCacheRefreshed();
103
104 // Update remote status based on refreshed data held in the cache
105 //检查刚刚更新的缓存中，有来自Eureka server的服务列表，其中包含了当前应用的状态，
106 //当前实例的成员变量lastRemoteInstanceStatus，记录的是最后一次更新的当前应用状态，
107 //上述两种状态在updateInstanceRemoteStatus方法中作比较 ，如果不一致，就更新lastRemoteInstanceStatu
s，并且广播对应的事件
108 updateInstanceRemoteStatus();
109
110 // registry was fetched successfully, so return true
111 return true;
112 }

全量更新getAndStoreFullRegistry

1 private void getAndStoreFullRegistry() throws Throwable {
2 long currentUpdateGeneration = fetchRegistryGeneration.get();
3
4 logger.info("Getting all instance registry info from the eureka server");
5
6 Applications apps = null;
7 //由于并没有配置特别关注的region信息，因此会调用eurekaTransport.queryClient.getApplications方法从服
务端获取服务列表
8 EurekaHttpResponse<Applications> httpResponse =
clientConfig.getRegistryRefreshSingleVipAddress() == null
9 ? eurekaTransport.queryClient.getApplications(remoteRegionsRef.get())
10 : eurekaTransport.queryClient.getVip(clientConfig.getRegistryRefreshSingleVipAddress(), remoteR
egionsRef.get());
11 if (httpResponse.getStatusCode() == Status.OK.getStatusCode()) {
12 //返回对象就是服务列表
13 apps = httpResponse.getEntity();
14 }
15 logger.info("The response status is {}", httpResponse.getStatusCode());
16
17 if (apps == null) {
18 logger.error("The application is null for some reason. Not storing this information");
19 }
20 //考虑到多线程同步，只有CAS成功的线程，才会把自己从Eureka server获取的数据来替换本地缓存
21 else if (fetchRegistryGeneration.compareAndSet(currentUpdateGeneration, currentUpdateGeneration
+ 1)) {
22 //localRegionApps就是本地缓存，是个AtomicReference实例
23 localRegionApps.set(this.filterAndShuffle(apps));
24 logger.debug("Got full registry with apps hashcode {}", apps.getAppsHashCode());
25 } else {
26 logger.warn("Not updating applications as another thread is updating it already");
27 }
28 }

其中最重要的一段代码eurekaTransport.queryClient.getApplications(remoteRegionsRef.get())，和Eureka server交互的逻辑都在这里面，方法getApplications的具体实现是在EurekaHttpClientDecorator类

1 @Override
2 public EurekaHttpResponse<Applications> getApplications(final String... regions) {
3 return execute(new RequestExecutor<Applications>() {
4 @Override
5 public EurekaHttpResponse<Applications> execute(EurekaHttpClient delegate) {
6 return delegate.getApplications(regions);
7 }
8
9 @Override
10 public RequestType getRequestType() {
11 //本次向Eureka server请求的类型：获取服务列表
12 return RequestType.GetApplications;
13 }
14 });
15 }

 

debug进去delegate.getApplications(regions)方法会发现delegate实际用的是 AbstractJerseyEurekaHttpClient，里面都是具体的jersey实现的网络接口请求

1 @Override
2 public EurekaHttpResponse<Applications> getApplications(String... regions) {
3 //取全量数据的path是"apps"
4 return getApplicationsInternal("apps/", regions);
5 }
6
7 @Override
8 public EurekaHttpResponse<Applications> getDelta(String... regions) {
9 //取增量数据的path是"apps/delta"
10 return getApplicationsInternal("apps/delta", regions);
11 }
12
13 //具体的请求响应处理都在此方法中
14 private EurekaHttpResponse<Applications> getApplicationsInternal(String urlPath, String[] region
s) {
15 ClientResponse response = null;
16 String regionsParamValue = null;
17 try {
18 //jersey、resource这些关键词都预示着这是个restful请求
19 WebResource webResource = jerseyClient.resource(serviceUrl).path(urlPath);
20 if (regions != null && regions.length > 0) {
21 regionsParamValue = StringUtil.join(regions);
22 webResource = webResource.queryParam("regions", regionsParamValue);
23 }
24 Builder requestBuilder = webResource.getRequestBuilder();
25 addExtraHeaders(requestBuilder);
26 //发起网络请求，将响应封装成ClientResponse实例
27 response = requestBuilder.accept(MediaType.APPLICATION_JSON_TYPE).get(ClientResponse.class);
28
29 Applications applications = null;
30 if (response.getStatus() == Status.OK.getStatusCode() && response.hasEntity()) {
31 //取得全部应用信息
32 applications = response.getEntity(Applications.class);
33 }
34 return anEurekaHttpResponse(response.getStatus(), Applications.class)
35 .headers(headersOf(response))
36 .entity(applications)
37 .build();
38 } finally {
39 if (logger.isDebugEnabled()) {
40 logger.debug("Jersey HTTP GET {}/{}?{}; statusCode={}",
41 serviceUrl, urlPath,
42 regionsParamValue == null ? "" : "regions=" + regionsParamValue,
43 response == null ? "N/A" : response.getStatus()
44 );
45 }
46 if (response != null) {
47 response.close();
48 }
49 }
50 }

 

获取全量数据，是通过jersey-client库的API向Eureka server发起restful请求 http://localhost:8761/eureka/apps实现的，并将响应的服务列表数据放在一个成员变量中作为本地 缓存

1 <applications>
2 <versions__delta>1</versions__delta>
3 <apps__hashcode>UP_1_</apps__hashcode>
4 <application>
5 <name>MICROSERVICE‐PROVIDER‐USER</name>
6 <instance>
7 <instanceId>localhost:microservice‐provider‐user:8002</instanceId>
8 <hostName>192.168.101.1</hostName>
9 <app>MICROSERVICE‐PROVIDER‐USER</app>
10 <ipAddr>192.168.101.1</ipAddr>
11 <status>UP</status>
12 <overriddenstatus>UNKNOWN</overriddenstatus>
13 <port enabled="true">8002</port>
14 <securePort enabled="false">443</securePort>
15 <countryId>1</countryId>
16 <dataCenterInfo class="com.netflix.appinfo.InstanceInfo$DefaultDataCenterInfo">
17 <name>MyOwn</name>
18 </dataCenterInfo>
19 <leaseInfo>
20 <renewalIntervalInSecs>30</renewalIntervalInSecs>
21 <durationInSecs>90</durationInSecs>
22 <registrationTimestamp>1554360812763</registrationTimestamp>
23 <lastRenewalTimestamp>1554360812763</lastRenewalTimestamp>
24 <evictionTimestamp>0</evictionTimestamp>
25 <serviceUpTimestamp>1554360812763</serviceUpTimestamp>
26 </leaseInfo>
27 <metadata>
28 <management.port>8002</management.port>
29 <jmx.port>61822</jmx.port>
30 </metadata>
31 <homePageUrl>http://192.168.101.1:8002/</homePageUrl>
32 <statusPageUrl>http://192.168.101.1:8002/actuator/info</statusPageUrl>
33 <healthCheckUrl>http://192.168.101.1:8002/actuator/health</healthCheckUrl>
34 <vipAddress>microservice‐provider‐user</vipAddress>
35 <secureVipAddress>microservice‐provider‐user</secureVipAddress>
36 <isCoordinatingDiscoveryServer>false</isCoordinatingDiscoveryServer>
37 <lastUpdatedTimestamp>1554360812764</lastUpdatedTimestamp>
38 <lastDirtyTimestamp>1554360812649</lastDirtyTimestamp>
39 <actionType>ADDED</actionType>
40 </instance>
41 </application>
42 </applications>

获取服务列表信息的增量更新getAndUpdateDelta

1 private void getAndUpdateDelta(Applications applications) throws Throwable {
2 long currentUpdateGeneration = fetchRegistryGeneration.get();
3
4 Applications delta = null;
5 //增量信息是通过eurekaTransport.queryClient.getDelta方法完成的
6 EurekaHttpResponse<Applications> httpResponse = eurekaTransport.queryClient.getDelta(remoteRegi
onsRef.get());
7 if (httpResponse.getStatusCode() == Status.OK.getStatusCode()) {
8 //delta中保存了Eureka server返回的增量更新
9 delta = httpResponse.getEntity();
10 }
11
12 if (delta == null) {
13 logger.warn("The server does not allow the delta revision to be applied because it is not safe.
"
14 + "Hence got the full registry.");
15 //如果增量信息为空，就直接发起一次全量更新
16 getAndStoreFullRegistry();
17 }
18 //考虑到多线程同步问题，这里通过CAS来确保请求发起到现在是线程安全的，
19 //如果这期间fetchRegistryGeneration变了，就表示其他线程也做了类似操作，因此放弃本次响应的数据
20 else if (fetchRegistryGeneration.compareAndSet(currentUpdateGeneration, currentUpdateGeneration
+ 1)) {
21 logger.debug("Got delta update with apps hashcode {}", delta.getAppsHashCode());
22 String reconcileHashCode = "";
23 if (fetchRegistryUpdateLock.tryLock()) {
24 try {
25 //用Eureka返回的增量数据和本地数据做合并操作，这个方法稍后会细说
26 updateDelta(delta);
27 //用合并了增量数据之后的本地数据来生成一致性哈希码
28 reconcileHashCode = getReconcileHashCode(applications);
29 } finally {
30 fetchRegistryUpdateLock.unlock();
31 }
32 } else {
33 logger.warn("Cannot acquire update lock, aborting getAndUpdateDelta");
34 }
35 //Eureka server在返回增量更新数据时，也会返回服务端的一致性哈希码，
36 //理论上每次本地缓存数据经历了多次增量更新后，计算出的一致性哈希码应该是和服务端一致的，
37 //如果发现不一致，就证明本地缓存的服务列表信息和Eureka server不一致了，需要做一次全量更新
38 if (!reconcileHashCode.equals(delta.getAppsHashCode()) || clientConfig.shouldLogDeltaDiff()) {
39 //一致性哈希码不同，就在reconcileAndLogDifference方法中做全量更新
40 reconcileAndLogDifference(delta, reconcileHashCode); // this makes a remoteCall
41 }
42 } else {
43 logger.warn("Not updating application delta as another thread is updating it already");
44 logger.debug("Ignoring delta update with apps hashcode {}, as another thread is updating it alr
eady", delta.getAppsHashCode());
45 }
46 }

 

 

updateDelta方法将增量更新数据和本地数据做合并

1 private void updateDelta(Applications delta) {
2 int deltaCount = 0;
3 //遍历所有服务
4 for (Application app : delta.getRegisteredApplications()) {
5 //遍历当前服务的所有实例
6 for (InstanceInfo instance : app.getInstances()) {
7 //取出缓存的所有服务列表，用于合并
8 Applications applications = getApplications();
9 String instanceRegion = instanceRegionChecker.getInstanceRegion(instance);
10 //判断正在处理的实例和当前应用是否在同一个region
11 if (!instanceRegionChecker.isLocalRegion(instanceRegion)) {
12 //如果不是同一个region，接下来合并的数据就换成专门为其他region准备的缓存
13 Applications remoteApps = remoteRegionVsApps.get(instanceRegion);
14 if (null == remoteApps) {
15 remoteApps = new Applications();
16 remoteRegionVsApps.put(instanceRegion, remoteApps);
17 }
18 applications = remoteApps;
19 }
20
21 ++deltaCount;
22
23 if (ActionType.ADDED.equals(instance.getActionType())) { //对新增的实例的处理
24 Application existingApp = applications.getRegisteredApplications(instance.getAppName());
25 if (existingApp == null) {
26 applications.addApplication(app);
27 }
28 logger.debug("Added instance {} to the existing apps in region {}", instance.getId(), instanceR
egion);
29 applications.getRegisteredApplications(instance.getAppName()).addInstance(instance);
30 } else if (ActionType.MODIFIED.equals(instance.getActionType())) { //对修改实例的处理
31 Application existingApp = applications.getRegisteredApplications(instance.getAppName());
32 if (existingApp == null) {
33 applications.addApplication(app);
34 }
35 logger.debug("Modified instance {} to the existing apps ", instance.getId());
36
37 applications.getRegisteredApplications(instance.getAppName()).addInstance(instance);
38
39 } else if (ActionType.DELETED.equals(instance.getActionType())) { //对删除实例的处理
40 Application existingApp = applications.getRegisteredApplications(instance.getAppName());
41 if (existingApp == null) {
42 applications.addApplication(app);
43 }
44 logger.debug("Deleted instance {} to the existing apps ", instance.getId());
45 applications.getRegisteredApplications(instance.getAppName()).removeInstance(instance);
46 }
47 }
48 }
49 logger.debug("The total number of instances fetched by the delta processor : {}", deltaCount);
50
51 getApplications().setVersion(delta.getVersion());
52 //整理数据，使得后续使用过程中，这些应用的实例总是以相同顺序返回
53 getApplications().shuffleInstances(clientConfig.shouldFilterOnlyUpInstances());
54
55 //和当前应用不在同一个region的应用，其实例数据也要整理
56 for (Applications applications : remoteRegionVsApps.values()) {
57 applications.setVersion(delta.getVersion());
58 applications.shuffleInstances(clientConfig.shouldFilterOnlyUpInstances());
59 }
60 }

 

 

服务续约

1 // 服务定时续约
2 scheduler.schedule(
3 new TimedSupervisorTask(
4 "heartbeat",
5 scheduler,
6 heartbeatExecutor,
7 renewalIntervalInSecs,
8 TimeUnit.SECONDS,
9 expBackOffBound,
10 new HeartbeatThread() //该线程执行续约的具体逻辑，会调用下面的renew()方法
11 ),
12 renewalIntervalInSecs, TimeUnit.SECONDS);
13
14 private class HeartbeatThread implements Runnable {
15 public void run() {
16 if (renew()) {
17 lastSuccessfulHeartbeatTimestamp = System.currentTimeMillis();
18 }
19 }
20 }
21
22 boolean renew() {
23 EurekaHttpResponse<InstanceInfo> httpResponse;
24 try {
25 httpResponse = eurekaTransport.registrationClient.sendHeartBeat(instanceInfo.getAppName(), inst
anceInfo.getId(), instanceInfo, null);
26 logger.debug(PREFIX + "{} ‐ Heartbeat status: {}", appPathIdentifier, httpResponse.getStatusCod
e());
27 if (httpResponse.getStatusCode() == 404) {
28 REREGISTER_COUNTER.increment();
29 logger.info(PREFIX + "{} ‐ Re‐registering apps/{}", appPathIdentifier,
instanceInfo.getAppName());
30 long timestamp = instanceInfo.setIsDirtyWithTime();
31 boolean success = register();
32 if (success) {
33 instanceInfo.unsetIsDirty(timestamp);
34 }
35 return success;
36 }
37 return httpResponse.getStatusCode() == 200;
38 } catch (Throwable e) {
39 logger.error(PREFIX + "{} ‐ was unable to send heartbeat!", appPathIdentifier, e);
40 return false;
41 }
42 }

 

服务注册

1 //服务注册
2 instanceInfoReplicator.start(clientConfig.getInitialInstanceInfoReplicationIntervalSeconds());
3
4 public void start(int initialDelayMs) {
5 if (started.compareAndSet(false, true)) {
6 instanceInfo.setIsDirty(); // for initial register
7 Future next = scheduler.schedule(this, initialDelayMs, TimeUnit.SECONDS);
8 scheduledPeriodicRef.set(next);
9 }
10 }
11
12 public void run() {
13 try {
14 discoveryClient.refreshInstanceInfo();
15
16 Long dirtyTimestamp = instanceInfo.isDirtyWithTime();
17 if (dirtyTimestamp != null) {
18 discoveryClient.register();
19 instanceInfo.unsetIsDirty(dirtyTimestamp);
20 }
21 } catch (Throwable t) {
22 logger.warn("There was a problem with the instance info replicator", t);
23 } finally {
24 Future next = scheduler.schedule(this, replicationIntervalSeconds, TimeUnit.SECONDS);
25 scheduledPeriodicRef.set(next);
26 }
27 }

 

Eureka Server服务端Jersey接口源码分析

服务端Jersey接口处理类ApplicationResource

其中有一个addInstance方法就是用来接收客户端的注册请求接口

1 //ApplicationResource.java
2 @POST
3 @Consumes({"application/json", "application/xml"})
4 public Response addInstance(InstanceInfo info,
5 @HeaderParam(PeerEurekaNode.HEADER_REPLICATION) String isReplication) {
6 logger.debug("Registering instance {} (replication={})", info.getId(), isReplication);
7 // validate that the instanceinfo contains all the necessary required fields
8 // 参数校验，不符合验证规则的，返回400状态码，此处不做详解
9 if (isBlank(info.getId())) {
10 return Response.status(400).entity("Missing instanceId").build();
11 } else if (isBlank(info.getHostName())) {
12 return Response.status(400).entity("Missing hostname").build();
13 } else if (isBlank(info.getAppName())) {
14 return Response.status(400).entity("Missing appName").build();
15 } else if (!appName.equals(info.getAppName())) {
16 return Response.status(400).entity("Mismatched appName, expecting " + appName + " but was " + i
nfo.getAppName()).build();
17 } else if (info.getDataCenterInfo() == null) {
18 return Response.status(400).entity("Missing dataCenterInfo").build();
19 } else if (info.getDataCenterInfo().getName() == null) {
20 return Response.status(400).entity("Missing dataCenterInfo Name").build();
21 }
22
23 // handle cases where clients may be registering with bad DataCenterInfo with missing data
24 DataCenterInfo dataCenterInfo = info.getDataCenterInfo();
25 if (dataCenterInfo instanceof UniqueIdentifier) {
26 String dataCenterInfoId = ((UniqueIdentifier) dataCenterInfo).getId();
27 if (isBlank(dataCenterInfoId)) {
28 boolean experimental = "true".equalsIgnoreCase(serverConfig.getExperimental("registration.valid
ation.dataCenterInfoId"));
29 if (experimental) {
30 String entity = "DataCenterInfo of type " + dataCenterInfo.getClass() + " must contain a valid
id";
31 return Response.status(400).entity(entity).build();
32 } else if (dataCenterInfo instanceof AmazonInfo) {
33 AmazonInfo amazonInfo = (AmazonInfo) dataCenterInfo;
34 String effectiveId = amazonInfo.get(AmazonInfo.MetaDataKey.instanceId);
35 if (effectiveId == null) {
36 amazonInfo.getMetadata().put(AmazonInfo.MetaDataKey.instanceId.getName(), info.getId());
37 }
38 } else {
39 logger.warn("Registering DataCenterInfo of type {} without an appropriate id", dataCenterInfo.g
etClass());
40 }
41 }
42 }
43 // 重点在这里
44 registry.register(info, "true".equals(isReplication));
45 return Response.status(204).build(); // 204 to be backwards compatible
46 }

 

AbstractInstanceRegistry的注册方法

1 public void register(InstanceInfo registrant, int leaseDuration, boolean isReplication) {
2 try {
3 // 上只读锁
4 read.lock();
5 // 从本地MAP里面获取当前实例的信息。
6 Map<String, Lease<InstanceInfo>> gMap = registry.get(registrant.getAppName());
7 // 增加注册次数到监控信息里面去。
8 REGISTER.increment(isReplication);
9 if (gMap == null) {
10 // 如果第一次进来，那么gMap为空，则创建一个ConcurrentHashMap放入到registry里面去
11 final ConcurrentHashMap<String, Lease<InstanceInfo>> gNewMap = new ConcurrentHashMap<String, Le
ase<InstanceInfo>>();
12 // putIfAbsent方法主要是在向ConcurrentHashMap中添加键—值对的时候，它会先判断该键值对是否已经存在。
13 // 如果不存在（新的entry），那么会向map中添加该键值对，并返回null。
14 // 如果已经存在，那么不会覆盖已有的值，直接返回已经存在的值。
15 gMap = registry.putIfAbsent(registrant.getAppName(), gNewMap);
16 if (gMap == null) {
17 // 表明map中确实不存在，则设置gMap为最新创建的那个
18 gMap = gNewMap;
19 }
20 }
21 // 从MAP中查询已经存在的Lease信息 （比如第二次来）
22 Lease<InstanceInfo> existingLease = gMap.get(registrant.getId());
23 // 当Lease的对象不为空时。
24 if (existingLease != null && (existingLease.getHolder() != null)) {
25 // 当instance已经存在是，和客户端的instance的信息做比较，时间最新的那个，为有效instance信息
26 Long existingLastDirtyTimestamp = existingLease.getHolder().getLastDirtyTimestamp(); // server
27 Long registrationLastDirtyTimestamp = registrant.getLastDirtyTimestamp(); // client
28 logger.debug("Existing lease found (existing={}, provided={}", existingLastDirtyTimestamp, regi
strationLastDirtyTimestamp);
29 if (existingLastDirtyTimestamp > registrationLastDirtyTimestamp) {
30 logger.warn("There is an existing lease and the existing lease's dirty timestamp {} is greater"
+
31 " than the one that is being registered {}", existingLastDirtyTimestamp, registrationLastDirtyT
imestamp);
32 logger.warn("Using the existing instanceInfo instead of the new instanceInfo as the
registrant");
33 registrant = existingLease.getHolder();
34 }
35 } else {
36 // 这里只有当existinglease不存在时，才会进来。 像那种恢复心跳，信息过期的，都不会进入这里。
37 // Eureka‐Server的自我保护机制做的操作，为每分钟最大续约数+2 ，同时重新计算每分钟最小续约数
38 synchronized (lock) {
39 if (this.expectedNumberOfRenewsPerMin > 0) {
40 // Since the client wants to cancel it, reduce the threshold
41 // (1 for 30 seconds, 2 for a minute)
42 this.expectedNumberOfRenewsPerMin = this.expectedNumberOfRenewsPerMin + 2;
43 this.numberOfRenewsPerMinThreshold =
44 (int) (this.expectedNumberOfRenewsPerMin * serverConfig.getRenewalPercentThreshold());
45 }
46 }
47 logger.debug("No previous lease information found; it is new registration");
48 }
49 // 构建一个最新的Lease信息
50 Lease<InstanceInfo> lease = new Lease<InstanceInfo>(registrant, leaseDuration);
51 if (existingLease != null) {
52 // 当原来存在Lease的信息时，设置他的serviceUpTimestamp, 保证服务开启的时间一直是第一次的那个
53 lease.setServiceUpTimestamp(existingLease.getServiceUpTimestamp());
54 }
55 // 放入本地Map中
56 gMap.put(registrant.getId(), lease);
57 // 添加到最近的注册队列里面去，以时间戳作为Key， 名称作为value，主要是为了运维界面的统计数据。
58 synchronized (recentRegisteredQueue) {
59 recentRegisteredQueue.add(new Pair<Long, String>(
60 System.currentTimeMillis(),
61 registrant.getAppName() + "(" + registrant.getId() + ")"));
62 }
63 // This is where the initial state transfer of overridden status happens
64 // 分析instanceStatus
65 if (!InstanceStatus.UNKNOWN.equals(registrant.getOverriddenStatus())) {
66 logger.debug("Found overridden status {} for instance {}. Checking to see if needs to be add to
the "
67 + "overrides", registrant.getOverriddenStatus(), registrant.getId());
68 if (!overriddenInstanceStatusMap.containsKey(registrant.getId())) {
69 logger.info("Not found overridden id {} and hence adding it", registrant.getId());
70 overriddenInstanceStatusMap.put(registrant.getId(), registrant.getOverriddenStatus());
71 }
72 }
73 InstanceStatus overriddenStatusFromMap = overriddenInstanceStatusMap.get(registrant.getId());
74 if (overriddenStatusFromMap != null) {
75 logger.info("Storing overridden status {} from map", overriddenStatusFromMap);
76 registrant.setOverriddenStatus(overriddenStatusFromMap);
77 }
78
79 // Set the status based on the overridden status rules
80 InstanceStatus overriddenInstanceStatus = getOverriddenInstanceStatus(registrant,
existingLease, isReplication);
81 registrant.setStatusWithoutDirty(overriddenInstanceStatus);
82
83 // If the lease is registered with UP status, set lease service up timestamp
84 // 得到instanceStatus，判断是否是UP状态，
85 if (InstanceStatus.UP.equals(registrant.getStatus())) {
86 lease.serviceUp();
87 }
88 // 设置注册类型为添加
89 registrant.setActionType(ActionType.ADDED);
90 // 租约变更记录队列，记录了实例的每次变化， 用于注册信息的增量获取、
91 recentlyChangedQueue.add(new RecentlyChangedItem(lease));
92 registrant.setLastUpdatedTimestamp();
93 // 清理缓存 ，传入的参数为key
94 invalidateCache(registrant.getAppName(), registrant.getVIPAddress(), registrant.getSecureVipAdd
ress());
95 logger.info("Registered instance {}/{} with status {} (replication={})",
96 registrant.getAppName(), registrant.getId(), registrant.getStatus(), isReplication);
97 } finally {
98 read.unlock();
99 }
100 }

 

理解上面的register还需要先了解下注册实例信息存放的的map，这是个两层的ConcurrentHashMap>>，外层map的key是appName，也就是服务名，内层map的key是 instanceId，也就是实例名 注册表map数据示例如下：

{ MICROSERVICE - PROVIDER - USER = { DESKTOP - 1 SLJLB7: microservice - provider - user: 8002 = com.netflix.eureka.lease.Lease @2cd36af6, DESKTOP - 1 SLJLB7: microservice - provider - user: 8001 = com.netflix.eureka.lease.Lease @600b7073 } }

 

内层map的value对应的类Lease需要重点理解下

1 public class Lease<T> {
2
3 enum Action {
4 Register, Cancel, Renew
5 };
6
7 public static final int DEFAULT_DURATION_IN_SECS = 90;
8
9 private T holder;
10 private long evictionTimestamp;
11 private long registrationTimestamp;
12 private long serviceUpTimestamp;
13 // Make it volatile so that the expiration task would see this quicker
14 private volatile long lastUpdateTimestamp;
15 private long duration;
16
17 public Lease(T r, int durationInSecs) {
18 holder = r;
19 registrationTimestamp = System.currentTimeMillis();
20 lastUpdateTimestamp = registrationTimestamp;
21 duration = (durationInSecs * 1000);
22
23 }
24
25 /**
26 * Renew the lease, use renewal duration if it was specified by the
27 * associated {@link T} during registration, otherwise default duration is
28 * {@link #DEFAULT_DURATION_IN_SECS}.
29 */
30 public void renew() {
31 lastUpdateTimestamp = System.currentTimeMillis() + duration; //有个小bug，不应该加duration
32
33 }
34
35 /**
36 * Cancels the lease by updating the eviction time.
37 */
38 public void cancel() {
39 if (evictionTimestamp <= 0) {
40 evictionTimestamp = System.currentTimeMillis();
41 }
42 }
43
44 /**
45 * Mark the service as up. This will only take affect the first time called,
46 * subsequent calls will be ignored.
47 */
48 public void serviceUp() {
49 if (serviceUpTimestamp == 0) {
50 serviceUpTimestamp = System.currentTimeMillis();
51 }
52 }
53
54 /**
55 * Set the leases service UP timestamp.
56 */
57 public void setServiceUpTimestamp(long serviceUpTimestamp) {
58 this.serviceUpTimestamp = serviceUpTimestamp;
59 }
60
61 /**
62 * Checks if the lease of a given {@link com.netflix.appinfo.InstanceInfo} has expired or not.
63 */
64 public boolean isExpired() {
65 return isExpired(0l);
66 }
67
68 /**
69 * Checks if the lease of a given {@link com.netflix.appinfo.InstanceInfo} has expired or not.
70 *
71 * Note that due to renew() doing the 'wrong" thing and setting lastUpdateTimestamp to +duration
more than
72 * what it should be, the expiry will actually be 2 * duration. This is a minor bug and should o
nly affect
73 * instances that ungracefully shutdown. Due to possible wide ranging impact to existing usage,
this will
74 * not be fixed.
75 *
76 * @param additionalLeaseMs any additional lease time to add to the lease evaluation in ms.
77 */
78 public boolean isExpired(long additionalLeaseMs) {
79 return (evictionTimestamp > 0 || System.currentTimeMillis() > (lastUpdateTimestamp + duration +
additionalLeaseMs));
80 }
81
82 /**
83 * Gets the milliseconds since epoch when the lease was registered.
84 *
85 * @return the milliseconds since epoch when the lease was registered.
86 */
87 public long getRegistrationTimestamp() {
88 return registrationTimestamp;
89 }
90
91 /**
92 * Gets the milliseconds since epoch when the lease was last renewed.
93 * Note that the value returned here is actually not the last lease renewal time but the renewal
+ duration.
94 *
95 * @return the milliseconds since epoch when the lease was last renewed.
96 */
97 public long getLastRenewalTimestamp() {
98 return lastUpdateTimestamp;
99 }
100
101 /**
102 * Gets the milliseconds since epoch when the lease was evicted.
103 *
104 * @return the milliseconds since epoch when the lease was evicted.
105 */
106 public long getEvictionTimestamp() {
107 return evictionTimestamp;
108 }
109
110 /**
111 * Gets the milliseconds since epoch when the service for the lease was marked as up.
112 *
113 * @return the milliseconds since epoch when the service for the lease was marked as up.
114 */
115 public long getServiceUpTimestamp() {
116 return serviceUpTimestamp;
117 }
118
119 /**
120 * Returns the holder of the lease.
121 */
122 public T getHolder() {
123 return holder;
124 }
125
126 }

 

DEFAULT_DURATION_IN_SECS : 租约过期的时间常量，默认未90秒，也就说90秒没有心跳过来，那么这边将会自 动剔除该节点 holder ：这个租约是属于谁的， 目前占用这个属性的是 instanceInfo，也就是客户端实例信息。 evictionTimestamp ： 租约是啥时候过期的，当服务下线的时候，会过来更新这个时间戳registrationTimestamp ： 租约的注册时间 serviceUpTimestamp ：服务启动时间 ，当客户端在注册的时候，instanceInfo的status 为UP的时候，则更新这个 时间戳 lastUpdateTimestamp ：最后更新时间，每次续约的时候，都会更新这个时间戳，在判断实例 是否过期时，需要用到这个属性。 duration：过期时间，毫秒单位

 

服务端Jersey接口处理类ApplicationsResource

其中有一个getContainers方法就是用来获取所有注册实例信息的接口

1 @GET
2 public Response getContainers(@PathParam("version") String version,
3 @HeaderParam(HEADER_ACCEPT) String acceptHeader,
4 @HeaderParam(HEADER_ACCEPT_ENCODING) String acceptEncoding,
5 @HeaderParam(EurekaAccept.HTTP_X_EUREKA_ACCEPT) String eurekaAccept,
6 @Context UriInfo uriInfo,
7 @Nullable @QueryParam("regions") String regionsStr) {
8
9 boolean isRemoteRegionRequested = null != regionsStr && !regionsStr.isEmpty();
10 String[] regions = null;
11 if (!isRemoteRegionRequested) {
12 EurekaMonitors.GET_ALL.increment();
13 } else {
14 regions = regionsStr.toLowerCase().split(",");
15 Arrays.sort(regions); // So we don't have different caches for same regions queried in differen
t order.
16 EurekaMonitors.GET_ALL_WITH_REMOTE_REGIONS.increment();
17 }
18
19 // Check if the server allows the access to the registry. The server can
20 // restrict access if it is not
21 // ready to serve traffic depending on various reasons.
22 if (!registry.shouldAllowAccess(isRemoteRegionRequested)) {
23 return Response.status(Status.FORBIDDEN).build();
24 }
25 CurrentRequestVersion.set(Version.toEnum(version));
26 KeyType keyType = Key.KeyType.JSON;
27 String returnMediaType = MediaType.APPLICATION_JSON;
28 if (acceptHeader == null || !acceptHeader.contains(HEADER_JSON_VALUE)) {
29 keyType = Key.KeyType.XML;
30 returnMediaType = MediaType.APPLICATION_XML;
31 }
32
33 //获取服务实例对应的缓存key
34 Key cacheKey = new Key(Key.EntityType.Application,
35 ResponseCacheImpl.ALL_APPS,
36 keyType, CurrentRequestVersion.get(), EurekaAccept.fromString(eurekaAccept), regions
37 );
38
39 Response response;
40 if (acceptEncoding != null && acceptEncoding.contains(HEADER_GZIP_VALUE)) {
41 response = Response.ok(responseCache.getGZIP(cacheKey))
42 .header(HEADER_CONTENT_ENCODING, HEADER_GZIP_VALUE)
43 .header(HEADER_CONTENT_TYPE, returnMediaType)
44 .build();
45 } else {
46 //从缓存里获取服务实例注册信息
47 response = Response.ok(responseCache.get(cacheKey))
48 .build();
49 }
50 return response;
51 }
52
53 responseCache.get(cacheKey)对应的源码如下：
54 @VisibleForTesting
55 String get(final Key key, boolean useReadOnlyCache) {
56 //从多级缓存里获取注册实例信息
57 Value payload = getValue(key, useReadOnlyCache);
58 if (payload == null || payload.getPayload().equals(EMPTY_PAYLOAD)) {
59 return null;
60 } else {
61 return payload.getPayload();
62 }
63 }
64
65 @VisibleForTesting
66 Value getValue(final Key key, boolean useReadOnlyCache) {
67 Value payload = null;
68 try {
69 if (useReadOnlyCache) {
70 final Value currentPayload = readOnlyCacheMap.get(key);
71 if (currentPayload != null) {
72 payload = currentPayload;
73 } else {
74 payload = readWriteCacheMap.get(key);
75 readOnlyCacheMap.put(key, payload);
76 }
77 } else {
78 payload = readWriteCacheMap.get(key);
79 }
80 } catch (Throwable t) {
81 logger.error("Cannot get value for key : {}", key, t);
82 }
83 return payload;
84 }
85
86
87 ResponseCacheImpl(EurekaServerConfig serverConfig, ServerCodecs serverCodecs, AbstractInstanceRe
gistry registry) {
88 this.serverConfig = serverConfig;
89 this.serverCodecs = serverCodecs;
90 this.shouldUseReadOnlyResponseCache = serverConfig.shouldUseReadOnlyResponseCache();
91 this.registry = registry;
92
93 long responseCacheUpdateIntervalMs = serverConfig.getResponseCacheUpdateIntervalMs();
94 this.readWriteCacheMap =
95 CacheBuilder.newBuilder().initialCapacity(1000)
96 //读写缓存默认180秒会自动定时过期
97 .expireAfterWrite(serverConfig.getResponseCacheAutoExpirationInSeconds(), TimeUnit.SECONDS)
98 .removalListener(new RemovalListener<Key, Value>() {
99 @Override
100 public void onRemoval(RemovalNotification<Key, Value> notification) {
101 Key removedKey = notification.getKey();
102 if (removedKey.hasRegions()) {
103 Key cloneWithNoRegions = removedKey.cloneWithoutRegions();
104 regionSpecificKeys.remove(cloneWithNoRegions, removedKey);
105 }
106 }
107 })
108 .build(new CacheLoader<Key, Value>() {
109 @Override
110 public Value load(Key key) throws Exception {
111 if (key.hasRegions()) {
112 Key cloneWithNoRegions = key.cloneWithoutRegions();
113 regionSpecificKeys.put(cloneWithNoRegions, key);
114 }
115 Value value = generatePayload(key);
116 return value;
117 }
118 });
119
120 if (shouldUseReadOnlyResponseCache) {
121 //默认30秒用读写缓存的数据更新只读缓存的数据
122 timer.schedule(getCacheUpdateTask(),
123 new Date(((System.currentTimeMillis() / responseCacheUpdateIntervalMs) * responseCacheUpdateIn
tervalMs)
124 + responseCacheUpdateIntervalMs),
125 responseCacheUpdateIntervalMs);
126 }
127
128 try {
129 Monitors.registerObject(this);
130 } catch (Throwable e) {
131 logger.warn("Cannot register the JMX monitor for the InstanceRegistry", e);
132 }
133 }
134
135 //初始化直接从注册表registry里那数据放入readWriteCacheMap
136 private Value generatePayload(Key key) {
137 Stopwatch tracer = null;
138 try {
139 String payload;
140 switch (key.getEntityType()) {
141 case Application:
142 boolean isRemoteRegionRequested = key.hasRegions();
143
144 if (ALL_APPS.equals(key.getName())) {
145 if (isRemoteRegionRequested) {
146 tracer = serializeAllAppsWithRemoteRegionTimer.start();
147 payload = getPayLoad(key, registry.getApplicationsFromMultipleRegions(key.getRegions()));
148 } else {
149 tracer = serializeAllAppsTimer.start();
150 payload = getPayLoad(key, registry.getApplications());
151 }
152 } else if (ALL_APPS_DELTA.equals(key.getName())) {
153 if (isRemoteRegionRequested) {
154 tracer = serializeDeltaAppsWithRemoteRegionTimer.start();
155 versionDeltaWithRegions.incrementAndGet();
156 versionDeltaWithRegionsLegacy.incrementAndGet();
157 payload = getPayLoad(key,
158 registry.getApplicationDeltasFromMultipleRegions(key.getRegions()));
159 } else {
160 tracer = serializeDeltaAppsTimer.start();
161 versionDelta.incrementAndGet();
162 versionDeltaLegacy.incrementAndGet();
163 payload = getPayLoad(key, registry.getApplicationDeltas());
164 }
165 } else {
166 tracer = serializeOneApptimer.start();
167 payload = getPayLoad(key, registry.getApplication(key.getName()));
168 }
169 break;
170 case VIP:
171 case SVIP:
172 tracer = serializeViptimer.start();
173 payload = getPayLoad(key, getApplicationsForVip(key, registry));
174 break;
175 default:
176 logger.error("Unidentified entity type: {} found in the cache key.", key.getEntityType());
177 payload = "";
178 break;
179 }
180 return new Value(payload);
181 } finally {
182 if (tracer != null) {
183 tracer.stop();
184 }
185 }
186 }
187
188 //用读写缓存的数据更新只读缓存的数据
189 private TimerTask getCacheUpdateTask() {
190 return new TimerTask() {
191 @Override
192 public void run() {
193 logger.debug("Updating the client cache from response cache");
194 for (Key key : readOnlyCacheMap.keySet()) {
195 if (logger.isDebugEnabled()) {
196 logger.debug("Updating the client cache from response cache for key : {} {} {} {}",
197 key.getEntityType(), key.getName(), key.getVersion(), key.getType());
198 }
199 try {
200 CurrentRequestVersion.set(key.getVersion());
201 Value cacheValue = readWriteCacheMap.get(key);
202 Value currentCacheValue = readOnlyCacheMap.get(key);
203 if (cacheValue != currentCacheValue) {
204 readOnlyCacheMap.put(key, cacheValue);
205 }
206 } catch (Throwable th) {
207 logger.error("Error while updating the client cache from response cache for key {}", key.toStr
ingCompact(), th);
208 }
209 }
210 }
211 };
212 }

 

源码精髓：多级缓存设计思想

在拉取注册表的时候： 首先从ReadOnlyCacheMap里查缓存的注册表。 若没有，就找ReadWriteCacheMap里缓存的注册表。 如果还没有，就从内存中获取实际的注册表数据。

 

在注册表发生变更的时候： 会在内存中更新变更的注册表数据，同时过期掉ReadWriteCacheMap。 此过程不会影响ReadOnlyCacheMap提供人家查询注册表。 默认每30秒Eureka Server会将ReadWriteCacheMap更新到 ReadOnlyCacheMap里 默认每180秒Eureka Server会将ReadWriteCacheMap里是数据失效 下次有服务拉取注册表，又会从内存中获取最新的数据了，同时填充 各级缓存。

 

多级缓存机制的优点： 尽可能保证了内存注册表数据不会出现频繁的读写冲突问题。 并且进一步保证对Eureka Server的大量请求，都是快速从纯内存走，性能极高（可以稍微估计下对于一 线互联网公司，内部上千个eureka client实例，每分钟对eureka上千次的访问，一天就是上千万次的访问）

 

看源码彻底搞懂一些诡异的问题： 看完多级缓存这块源码我们可以搞清楚一个常见的问题，就是当我们eureka服务实例有注册或下线或有实例发生故 障，内存注册表虽然会及时更新数据，但是客户端不一定能及时感知到，可能会过30秒才能感知到，因为客户端拉 取注册表实例这里面有一个多级缓存机制 还有服务剔除的不是默认90秒没心跳的实例，剔除的是180秒没心跳的实例(eureka的bug导致)