聊聊scalecube-cluster的FailureDetector

序

本文主要研究一下scalecube-cluster的FailureDetector

FailureDetector

scalecube-cluster-2.2.5/cluster/src/main/java/io/scalecube/cluster/fdetector/FailureDetector.java

/**
 * Failure Detector component responsible for monitoring availability of other members in the
 * cluster. This interface is supposed to be used internally as part cluster membership protocol. It
 * doesn't specify that particular node is failed, but just provide information that either it is
 * suspected or trusted at current moment of time. So it is up to cluster membership or other top
 * level component to define when suspected member is actually failed.
 */
public interface FailureDetector {

  /**
   * Starts running failure detection algorithm. After started it begins to receive and send ping
   * messages.
   */
  void start();

  /** Stops running failure detection algorithm and releases occupied resources. */
  void stop();

  /** Listens for results of ping checks (ALIVE/SUSPECT) done periodically by failure detector. */
  Flux<FailureDetectorEvent> listen();
}

• FailureDetector定義了start、stop、listen方法

FailureDetectorImpl

scalecube-cluster-2.2.5/cluster/src/main/java/io/scalecube/cluster/fdetector/FailureDetectorImpl.java

public final class FailureDetectorImpl implements FailureDetector {

  private static final Logger LOGGER = LoggerFactory.getLogger(FailureDetectorImpl.class);

  // Qualifiers

  public static final String PING = "sc/fdetector/ping";
  public static final String PING_REQ = "sc/fdetector/pingReq";
  public static final String PING_ACK = "sc/fdetector/pingAck";

  // Injected

  private final Member localMember;
  private final Transport transport;
  private final FailureDetectorConfig config;
  private final CorrelationIdGenerator cidGenerator;

  // State

  private long currentPeriod = 0;
  private List<Member> pingMembers = new ArrayList<>();
  private int pingMemberIndex = 0; // index for sequential ping member selection

  // Disposables

  private final Disposable.Composite actionsDisposables = Disposables.composite();

  // Subject
  private final FluxProcessor<FailureDetectorEvent, FailureDetectorEvent> subject =
      DirectProcessor.<FailureDetectorEvent>create().serialize();

  private final FluxSink<FailureDetectorEvent> sink = subject.sink();

  // Scheduled
  private final Scheduler scheduler;

  /**
   * Creates new instance of failure detector with given transport and settings.
   *
   * @param localMember local cluster member
   * @param transport cluster transport
   * @param membershipProcessor membership event processor
   * @param config failure detector settings
   * @param scheduler scheduler
   * @param cidGenerator correlationId generator
   */
  public FailureDetectorImpl(
      Member localMember,
      Transport transport,
      Flux<MembershipEvent> membershipProcessor,
      FailureDetectorConfig config,
      Scheduler scheduler,
      CorrelationIdGenerator cidGenerator) {

    this.localMember = Objects.requireNonNull(localMember);
    this.transport = Objects.requireNonNull(transport);
    this.config = Objects.requireNonNull(config);
    this.scheduler = Objects.requireNonNull(scheduler);
    this.cidGenerator = Objects.requireNonNull(cidGenerator);

    // Subscribe
    actionsDisposables.addAll(
        Arrays.asList(
            membershipProcessor //
                .publishOn(scheduler)
                .subscribe(this::onMemberEvent, this::onError),
            transport
                .listen() //
                .publishOn(scheduler)
                .subscribe(this::onMessage, this::onError)));
  }

  @Override
  public void start() {
    actionsDisposables.add(
        scheduler.schedulePeriodically(
            this::doPing,
            config.getPingInterval(),
            config.getPingInterval(),
            TimeUnit.MILLISECONDS));
  }

  @Override
  public void stop() {
    // Stop accepting requests and sending pings
    actionsDisposables.dispose();

    // Stop publishing events
    sink.complete();
  }

  @Override
  public Flux<FailureDetectorEvent> listen() {
    return subject.onBackpressureBuffer();
  }

  //......

}

• FailureDetectorImpl實現了FailureDetector接口；它定義了3個message的qualifier分別是PING、PING_REQ、PING_ACK；同時也定義了pingMembers列表
• FailureDetectorImpl的構造器訂閱了membershipProcessor觸發onMemberEvent方法，訂閱了transport.listen()觸發onMessage方法
• start方法經過scheduler.schedulePeriodically註冊了doPing任務，每隔pingInterval執行，默認是5000ms；stop方法會執行actionsDisposables.dispose()及sink.complete()；listen()則返回subject.onBackpressureBuffer()

onMemberEvent

scalecube-cluster-2.2.5/cluster/src/main/java/io/scalecube/cluster/fdetector/FailureDetectorImpl.java

public final class FailureDetectorImpl implements FailureDetector {
  //......

  private void onMemberEvent(MembershipEvent event) {
    Member member = event.member();
    if (event.isRemoved()) {
      pingMembers.remove(member);
    }
    if (event.isAdded()) {
      // insert member into random positions
      int size = pingMembers.size();
      int index = size > 0 ? ThreadLocalRandom.current().nextInt(size) : 0;
      pingMembers.add(index, member);
    }
  }

  //......
}

• onMemberEvent會根據MembershipEvent來移除或使用隨機的index添加member到pingMembers中

onMessage

scalecube-cluster-2.2.5/cluster/src/main/java/io/scalecube/cluster/fdetector/FailureDetectorImpl.java

public final class FailureDetectorImpl implements FailureDetector {
  //......

  private void onMessage(Message message) {
    if (isPing(message)) {
      onPing(message);
    } else if (isPingReq(message)) {
      onPingReq(message);
    } else if (isTransitPingAck(message)) {
      onTransitPingAck(message);
    }
  }

  private boolean isPing(Message message) {
    return PING.equals(message.qualifier());
  }

  private boolean isPingReq(Message message) {
    return PING_REQ.equals(message.qualifier());
  }

  private boolean isTransitPingAck(Message message) {
    return PING_ACK.equals(message.qualifier())
        && message.<PingData>data().getOriginalIssuer() != null;
  }

  /** Listens to PING message and answers with ACK. */
  private void onPing(Message message) {
    long period = this.currentPeriod;
    LOGGER.trace("Received Ping[{}]", period);
    PingData data = message.data();
    if (!data.getTo().id().equals(localMember.id())) {
      LOGGER.warn(
          "Received Ping[{}] to {}, but local member is {}", period, data.getTo(), localMember);
      return;
    }
    String correlationId = message.correlationId();
    Message ackMessage =
        Message.withData(data)
            .qualifier(PING_ACK)
            .correlationId(correlationId)
            .sender(localMember.address())
            .build();
    Address address = data.getFrom().address();
    LOGGER.trace("Send PingAck[{}] to {}", period, address);
    transport
        .send(address, ackMessage)
        .subscribe(
            null,
            ex ->
                LOGGER.debug(
                    "Failed to send PingAck[{}] to {}, cause: {}", period, address, ex.toString()));
  }

  /** Listens to PING_REQ message and sends PING to requested cluster member. */
  private void onPingReq(Message message) {
    long period = this.currentPeriod;
    LOGGER.trace("Received PingReq[{}]", period);
    PingData data = message.data();
    Member target = data.getTo();
    Member originalIssuer = data.getFrom();
    String correlationId = message.correlationId();
    PingData pingReqData = new PingData(localMember, target, originalIssuer);
    Message pingMessage =
        Message.withData(pingReqData)
            .qualifier(PING)
            .correlationId(correlationId)
            .sender(localMember.address())
            .build();
    Address address = target.address();
    LOGGER.trace("Send transit Ping[{}] to {}", period, address);
    transport
        .send(address, pingMessage)
        .subscribe(
            null,
            ex ->
                LOGGER.debug(
                    "Failed to send transit Ping[{}] to {}, cause: {}",
                    period,
                    address,
                    ex.toString()));
  }

  /**
   * Listens to ACK with message containing ORIGINAL_ISSUER then converts message to plain ACK and
   * sends it to ORIGINAL_ISSUER.
   */
  private void onTransitPingAck(Message message) {
    long period = this.currentPeriod;
    LOGGER.trace("Received transit PingAck[{}]", period);
    PingData data = message.data();
    Member target = data.getOriginalIssuer();
    String correlationId = message.correlationId();
    PingData originalAckData = new PingData(target, data.getTo());
    Message originalAckMessage =
        Message.withData(originalAckData)
            .qualifier(PING_ACK)
            .correlationId(correlationId)
            .sender(localMember.address())
            .build();
    Address address = target.address();
    LOGGER.trace("Resend transit PingAck[{}] to {}", period, address);
    transport
        .send(address, originalAckMessage)
        .subscribe(
            null,
            ex ->
                LOGGER.debug(
                    "Failed to resend transit PingAck[{}] to {}, cause: {}",
                    period,
                    address,
                    ex.toString()));
  }

  //......

}

• onMessage方法則根據消息的不一樣qualifier及originalIssuer信息來判斷執行onPing或onPingReq或onTransitPingAck方法；onPing方法會返回PING_ACK消息給sender；onPingReq方法則會發送PING給pingReq指定的member；onTransitPingAck方法則會將pingReq請求返回的ack再轉發回給originalIssuer

doPing

scalecube-cluster-2.2.5/cluster/src/main/java/io/scalecube/cluster/fdetector/FailureDetectorImpl.java

public final class FailureDetectorImpl implements FailureDetector {
  //......

  private void doPing() {
    // Increment period counter
    long period = currentPeriod++;

    // Select ping member
    Member pingMember = selectPingMember();
    if (pingMember == null) {
      return;
    }

    // Send ping
    String cid = cidGenerator.nextCid();
    PingData pingData = new PingData(localMember, pingMember);
    Message pingMsg =
        Message.withData(pingData)
            .qualifier(PING)
            .correlationId(cid)
            .sender(localMember.address())
            .build();

    LOGGER.trace("Send Ping[{}] to {}", period, pingMember);
    Address address = pingMember.address();
    transport
        .requestResponse(pingMsg, address)
        .timeout(Duration.ofMillis(config.getPingTimeout()), scheduler)
        .publishOn(scheduler)
        .subscribe(
            message -> {
              LOGGER.trace("Received PingAck[{}] from {}", period, pingMember);
              publishPingResult(period, pingMember, MemberStatus.ALIVE);
            },
            ex -> {
              LOGGER.debug(
                  "Failed to get PingAck[{}] from {} within {} ms",
                  period,
                  pingMember,
                  config.getPingTimeout());

              final int timeLeft = config.getPingInterval() - config.getPingTimeout();
              final List<Member> pingReqMembers = selectPingReqMembers(pingMember);

              if (timeLeft <= 0 || pingReqMembers.isEmpty()) {
                LOGGER.trace("No PingReq[{}] occurred", period);
                publishPingResult(period, pingMember, MemberStatus.SUSPECT);
              } else {
                doPingReq(currentPeriod, pingMember, pingReqMembers, cid);
              }
            });
  }

  private Member selectPingMember() {
    if (pingMembers.isEmpty()) {
      return null;
    }
    if (pingMemberIndex >= pingMembers.size()) {
      pingMemberIndex = 0;
      Collections.shuffle(pingMembers);
    }
    return pingMembers.get(pingMemberIndex++);
  }

  private List<Member> selectPingReqMembers(Member pingMember) {
    if (config.getPingReqMembers() <= 0) {
      return Collections.emptyList();
    }
    List<Member> candidates = new ArrayList<>(pingMembers);
    candidates.remove(pingMember);
    if (candidates.isEmpty()) {
      return Collections.emptyList();
    }
    Collections.shuffle(candidates);
    boolean selectAll = candidates.size() < config.getPingReqMembers();
    return selectAll ? candidates : candidates.subList(0, config.getPingReqMembers());
  }

  private void doPingReq(
      long period, final Member pingMember, final List<Member> pingReqMembers, String cid) {
    Message pingReqMsg =
        Message.withData(new PingData(localMember, pingMember))
            .qualifier(PING_REQ)
            .correlationId(cid)
            .sender(localMember.address())
            .build();
    LOGGER.trace("Send PingReq[{}] to {} for {}", period, pingReqMembers, pingMember);

    Duration timeout = Duration.ofMillis(config.getPingInterval() - config.getPingTimeout());
    pingReqMembers.forEach(
        member ->
            transport
                .requestResponse(pingReqMsg, member.address())
                .timeout(timeout, scheduler)
                .publishOn(scheduler)
                .subscribe(
                    message -> {
                      LOGGER.trace(
                          "Received transit PingAck[{}] from {} to {}",
                          period,
                          message.sender(),
                          pingMember);
                      publishPingResult(period, pingMember, MemberStatus.ALIVE);
                    },
                    throwable -> {
                      LOGGER.trace(
                          "Timeout getting transit PingAck[{}] from {} to {} within {} ms",
                          period,
                          pingReqMembers,
                          pingMember,
                          timeout);
                      publishPingResult(period, pingMember, MemberStatus.SUSPECT);
                    }));
  }

  private void publishPingResult(long period, Member member, MemberStatus status) {
    LOGGER.debug("Member {} detected as {} at [{}]", member, status, period);
    sink.next(new FailureDetectorEvent(member, status));
  }

  //......

}

• doPing方法首先遞增currentPeriod，而後經過selectPingMember隨機選擇pingMember，以後構造pingData，而後經過transport.requestResponse發送請求，請求成功時執行publishPingResult，異常狀況下則經過selectPingReqMembers隨機選擇pingReqMembers，在config.getPingInterval() - config.getPingTimeout()小於等於0或者pingReqMembers爲空時則執行publishPingResult(period, pingMember, MemberStatus.SUSPECT)，不然進行doPingReq
• selectPingMember方法在pingMemberIndex大於等於pingMembers.size()的時候會重置該index爲0，並執行Collections.shuffle(pingMembers)，以後遞增pingMemberIndex；selectPingReqMembers方法則基於pingMembers建立新的list而後移除pingMember獲得candidates，以後進行Collections.shuffle(candidates)，而後根據config.getPingReqMembers()來subList該candidates列表獲得pingReqMembers
• onPingReq方法則遍歷pingReqMembers對其發送pingReqMsg，當接收到transit PingAck時則執行publishPingResult(period, pingMember, MemberStatus.ALIVE)，出現異常時執行publishPingResult(period, pingMember, MemberStatus.SUSPECT)；publishPingResult方法往sink裏頭放入FailureDetectorEvent事件

小結

• FailureDetector定義了start、stop、listen方法；FailureDetectorImpl實現了FailureDetector接口；它定義了3個message的qualifier分別是PING、PING_REQ、PING_ACK；同時也定義了pingMembers列表；FailureDetectorImpl的構造器訂閱了membershipProcessor觸發onMemberEvent方法，訂閱了transport.listen()觸發onMessage方法；start方法經過scheduler.schedulePeriodically註冊了doPing任務，每隔pingInterval執行，默認是5000ms；stop方法會執行actionsDisposables.dispose()及sink.complete()；listen()則返回subject.onBackpressureBuffer()
• onMemberEvent會根據MembershipEvent來移除或使用隨機的index添加member到pingMembers中；onMessage方法則根據消息的不一樣qualifier及originalIssuer信息來判斷執行onPing或onPingReq或onTransitPingAck方法；onPing方法會返回PING_ACK消息給sender；onPingReq方法則會發送PING給pingReq指定的member；onTransitPingAck方法則會將pingReq請求返回的ack再轉發回給originalIssuer
• doPing方法首先遞增currentPeriod，而後經過selectPingMember隨機選擇pingMember，以後構造pingData，而後經過transport.requestResponse發送請求，請求成功時執行publishPingResult，異常狀況下則經過selectPingReqMembers隨機選擇pingReqMembers，在config.getPingInterval() - config.getPingTimeout()小於等於0或者pingReqMembers爲空時則執行publishPingResult(period, pingMember, MemberStatus.SUSPECT)，不然進行doPingReq

doc

• FailureDetector
• FailureDetectorImpl
• FailureDetectorEvent