使用Spring Integration和Hazelcast進行叢集領導者選舉

banq發表於2019-01-10
SpringAST

最近在檢查Spring Integration區域時，我注意到與Hazelcastdatagrid 的非常好的整合。在以下位置檢視：

https://github.com/spring-projects/spring-integration-extensions/tree/master/spring-integration-hazelcast

如果您使用Hazelcast，那麼您可以從Hazelcast支援的各種分散式資料結構中提供Spring Integration通道基礎架構，如：

  • com.hazelcast.core.IMap，
  • com.hazelcast.core.MultiMap，
  • com.hazelcast.core.IList，
  • com.hazelcast.core.ISet，
  • com.hazelcast.core.IQueue，
  • com.hazelcast.core.ITopic，
  • com.hazelcast.core.ReplicatedMap

真正令我印象深刻的是Spring Integration團隊透過Hazelcast實施叢集領導者選舉。讓我們來看看並測試它。但首先要做的事情是：
演示任務：我們有兩個Spring Boot微服務，每10秒產生一個隨機值放入分散式IMap。現在只允許一個微服務在某一時刻消費使用IMap的資料。為了使它變得有點辣，微服務應該在訊息傳遞給其他節點之後放棄其領導。

透過Spring Integration Cluster Leadership解決方案：
首先，您需要將Spring Boot MicroService新增到領導遊戲中：

@Bean
    public Candidate nodeService1Candidate() {
        final NodeCandidate candidate = new NodeCandidate("service1", HazelcastConfiguration.ROLE_JOB_MAP);
        return candidate;
    }


    @Bean
    public LeaderInitiator initiator() {
        final LeaderInitiator leaderInitiator = new LeaderInitiator(hazelcastConfiguration.hazelcastInstance(), nodeService1Candidate());
        return leaderInitiator;
    }


但這不是全部。我們的目標是在領導權被授予後從IMap開始資料消費，另一方面在領導被撤銷後停止資料消費。
為此，我們需要監聽org.springframework.integration.leader.DefaultCandidate子類NodeCandidate中的onGranted和onRevoked事件。第一個建構函式引數是節點id，第二個是角色名稱。閱讀Spring Integration角色，但我不會使用它們。我將手動啟動IMap更改生成者。
為了瞭解資料更改，SI Hazelcast整合提供了HazelcastEventDrivenMessageProducer，它可以監聽分散式IMap更改並將適當的資料更改事件委派給Spring Integration通道基礎結構。

@Configuration
public class HazelcastConfiguration {
    .
    .
    @Bean
    public IMap<String, String> getDistributedMapForJobInput() {
        return hazelcastInstance().getMap(INPUT_JOB_MAP);
    }

    @Bean
    public MessageChannel inputJobChannel() {
        return new DirectChannel();
    }

    @Bean
    public HazelcastEventDrivenMessageProducer hazelcastEventDrivenMessageProducer() {
        final HazelcastEventDrivenMessageProducer producer =
                new HazelcastEventDrivenMessageProducer(
                        getDistributedMapForJobInput()
                );
        producer.setOutputChannel(inputJobChannel());
        producer.setCacheEventTypes("ADDED,REMOVED,UPDATED,CLEAR_ALL");
        producer.setCacheListeningPolicy(CacheListeningPolicyType.SINGLE);
        producer.setAutoStartup(false);

        return producer;
    }
}


setAutostartup（false）的注意事項。我們希望讓這位生產者在獲得領導時能夠開始啟動：

/**
 * Created by tomask79 on 24.08.17.
 */
public class NodeCandidate extends DefaultCandidate {

    @Autowired
    private HazelcastConfiguration hazelcastConfiguration;

    public NodeCandidate(String nodeId, String role) {
        super(nodeId, role);
    }

    @Override
    public void onGranted(Context ctx) {
        super.onGranted(ctx);
        System.out.println("Leader granted to: "+ctx.toString());
        hazelcastConfiguration.hazelcastEventDrivenMessageProducer().start();
    }

    @Override
    public void onRevoked(Context ctx) {
        super.onRevoked(ctx);
        System.out.println("Leader revoked to: "+ctx.toString());
        hazelcastConfiguration.hazelcastEventDrivenMessageProducer().stop();
    }
}

最後一項任務是消費來自分散式IMap的訊息並放棄領導，以便其他節點可以接受工作並享受一些樂趣。因此，讓我們宣告ServiceActivator監聽來自jobInputChannel DirectChannel 的資料：

@Bean 
    @ServiceActivator（inputChannel =“inputJobChannel”）
    public MessageHandler logger（）{ 
        return new LogAndGiveInitiatorHandler（）; 
    }        


將訊息記錄到標準輸出：

/**
 * Created by tomask79 on 24.08.17.
 */
public class LogAndGiveInitiatorHandler implements MessageHandler{

    @Autowired
    private JobServices jobServices;

    @Override
    public void handleMessage(Message<?> message) throws MessagingException {
        System.out.println(message.toString());
        System.out.println("Waiting for another node to take the work...!");
        jobServices.giveUp();
        try {
            Thread.sleep(5000);
        } catch (InterruptedException e) {
            e.printStackTrace();
        }
        System.out.println("........");
    }
}

並命令微服務放棄其領導地位：

/**
 * Created by tomask79 on 10.08.17.
 */
@Service
public class JobServices {    
    @Autowired
    private LeaderInitiator initiator;
    .
    .
    public void giveUp() {
        if (initiator.getContext().isLeader()) {
            System.out.println("Giving up on leadership: "+initiator.getContext().toString());
            initiator.getContext().yield();
        }
    }
}


就是這樣！讓我們測試整個包。
  • git clone https://bitbucket.org/tomask79/spring-leader-hazelcast.git
  • mvn clean all (in the directory with top pom.xml to build all three projects)


輸出：

[INFO] Reactor Summary:
<p class="indent">[INFO] 
<p class="indent">[INFO] spring-cloud-cluster-demo .......................... SUCCESS [  0.412 s]
<p class="indent">[INFO] spring-microservice-hazelcast ...................... SUCCESS [  2.380 s]
<p class="indent">[INFO] spring-microservice-service1 ....................... SUCCESS [  3.685 s]
<p class="indent">[INFO] spring-microservice-service2 ....................... SUCCESS [  2.745 s]
<p class="indent">[INFO] ------------------------------------------------------------------------
<p class="indent">[INFO] BUILD SUCCESS
<p class="indent">[INFO] ------------------------------------------------------------------------
<p class="indent">[INFO] Total time: 10.047 s
<p class="indent">[INFO] Finished at: 2017-08-28T19:57:53+02:00
<p class="indent">[INFO] Final Memory: 40M/532M
<p class="indent">[INFO] ------------------------------------------------------------------------


現在開啟兩個終端並執行：
  • java -jar spring-microservice-service1 / target / service1-0.0.1-SNAPSHOT.war（在第一個終端）
  • java -jar spring-microservice-service2 / target / service2-0.0.1-SNAPSHOT.war（在第二個終端）

要驗證兩個微服務是否形成有效的Hazelcast叢集，您應該看到類似的內容：

Members [2] {
    Member [192.168.1.112]:5702
    Member [192.168.1.112]:5701 this
}


在形成Hazelcast群集設定後，您應該看到以下輸出

第一終端（獲取領導權並放棄給服務2）：

[st-leadership-0] com.example.hazelcast.NodeCandidate      : DefaultCandidate{role=leader, id=service1} has been granted leadership; context: HazelcastContext{role=leader, id=service1, isLeader=true}
Leader granted to: HazelcastContext{role=leader, id=service1, isLeader=true}
<p class="indent">[st-leadership-0] .h.i.HazelcastEventDrivenMessageProducer : started hazelcastEventDrivenMessageProducer
GenericMessage [payload=EntryEventMessagePayload [key=service18eff005d-6da8-4fb8-b747-f977ad8e1544, value=a61b5f9a-1b96-493d-b240-61ccb549ba17, oldValue=null], headers={hazelcast_cacheName=randomInputDataMap, hazelcast_member=/192.168.1.112:5702, id=f9c5455b-b42d-3ab7-ec49-9bd33db9ec5f, hazelcast_eventType=ADDED, timestamp=1503945864993}]
Waiting for another node to take the work...!
Giving up on leadership: HazelcastContext{role=leader, id=service1, isLeader=true}


第二終端（獲取領導權並放棄給服務1）

Leader granted to: HazelcastContext{role=leader, id=service2, isLeader=true}
2017-08-28 20:47:08.001  INFO 1357 --- [st-leadership-0] .h.i.HazelcastEventDrivenMessageProducer : started hazelcastEventDrivenMessageProducer
2017-08-28 20:47:08.019  INFO 1357 --- [           main] s.b.c.e.t.TomcatEmbeddedServletContainer : Tomcat started on port(s): 8082 (http)
2017-08-28 20:47:08.029  INFO 1357 --- [           main] c.e.SpringMicroserviceServiceComponent   : Started SpringMicroserviceServiceComponent in 12.807 seconds (JVM running for 13.507)
........
GenericMessage [payload=EntryEventMessagePayload [key=service249cde108-5045-4b77-84f7-cdc9f524df04, value=c4fe775f-e44d-4f10-ac43-0fe7157c0e67, oldValue=null], headers={hazelcast_cacheName=randomInputDataMap, hazelcast_member=/192.168.1.112:5701, id=df474178-1ff1-35e5-e1e1-d3f6f25d6d68, hazelcast_eventType=ADDED, timestamp=1503946037904}]
Waiting for another node to take the work...!
Giving up on leadership: HazelcastContext{role=leader, id=service2, isLeader=true}


總結
只是一些想法。如果在生產過程中流經系統的訊息數量每天只有幾千（我們在Embedit的生產系統中的速率），那麼Hazelcast肯定是一種有過度殺傷力工作。建議始終使用JMS / AMPQ以迴圈方式將資料分發到您的節點。 但是當處理儲存在記憶體中的大資料時。你不應該錯過由Hazelcast支援的Spring Integration Election演算法。

相關文章