Redis（十一）：哨兵模式架构设计分析

业务最初的应用场景中，我们也许使用单机 redis 就可以应付业务要求，但并非一直可行。

比如单机的读写能力问题，单机的可用性问题，单机的数据安全性问题。这些都是许多互联网应用经常会遇到的问题，也基本上都有一套理论去解决它，只是百花齐放。

哨兵是Redis中解决高可用问题的解决方案之一，我们就一起来看看 Redis是如何实现的吧！不过此方案，仅提供思路供参考，不要以此为标准方案。

前面介绍的主从复制功能，可以说已经一定程度上解决了数据安全性问题问题，即有了备份数据，我们可以可以做读写分离了。只是，可用性问题还未解决，即当 master 宕机或出现其他故障时，整个写服务就不可用了。解决方法是，手动操作，要么重启master使其恢复服务，要么把master切换为其他slave机器。

如果服务的可用性需要人工介入的话，那就算不得高可用了，所以我们需要一个自动处理机制。这就是哨兵模式。

一、哨兵系统介绍

哨兵系统要解决的问题核心，自然是高可用问题。而如何解决，则是其设计问题。而最终呈现给用户的，应该一个个的功能单元，即其提供的能力。如下：

监控（Monitoring）： Sentinel 会不断地检查你的主服务器和从服务器是否运作正常。

提醒（Notification）： 当被监控的某个 Redis 服务器出现问题时， Sentinel 可以通过 API 向管理员或者其他应用程序发送通知。

自动故障迁移（Automatic failover）： 当一个主服务器不能正常工作时， Sentinel 会开始一次自动故障迁移操作， 它会将失效主服务器的其中一个从服务器升级为新的主服务器， 并让失效主服务器的其他从服务器改为复制新的主服务器；

配置提供者: Sentinel充当客户端服务发现的授权来源：客户端连接到Sentinels，以询问负责给定服务的当前Redis主服务器的地址。 如果发生故障转移，Sentinels将报告新地址。（这也是客户端接入入口）

哨兵系统的架构图如下：

（一）服务端架构

（二）请求处理流程图

二、哨兵系统搭建步骤

哨兵可以搭建在 redis服务所在机器，也可以在单独的机器实例上搭建。

1. 有多个在运行的 redis master/slave 实例；

主从服务的搭建，slaveof 设置，请参照主从配置篇。

2. 编写哨兵配置文件；

# Example sentinel.conf
# 定义sentinel 服务端口号
port 26379

# 针对 使用端口映射的方式的启动，指定ip:port
# sentinel announce-ip <ip>
# sentinel announce-port <port>
# 工作目录定义
dir /tmp

# 要监视的redis master 定义, 可配置多个 master-name 不同即可
# sentinel monitor <master-name> <ip> <redis-port> <quorum>
sentinel monitor mymaster 127.0.0.1 6379 2

# 定义master/slave 的密码，要求同一主从服务所有密码必须保持一致
# sentinel auth-pass <master-name> <password>

# 定义master 不可达持续多少毫秒后开始定义为节点下线，默认30s
sentinel down-after-milliseconds mymaster 30000

# sentinel parallel-syncs <master-name> <numslaves>
# 在故障转移期间同时与新的master同步的slave数量
sentinel parallel-syncs mymaster 1

# 定义进行故障转移的超时时间，默认3分钟
sentinel failover-timeout mymaster 180000

# 发生故障转移时调用的通知脚本,被调用时会传递两个参数: eventType, eventDescription
# sentinel notification-script mymaster /var/redis/notify.sh

# master 变更时调用脚本配置
# 调用时会传递如下参数
# <master-name> <role> <state> <from-ip> <from-port> <to-ip> <to-port>
# sentinel client-reconfig-script mymaster /var/redis/reconfig.sh

3. 启动哨兵节点；

    # 使用 redis-sentinel 程序启动, 这个程序不一定会有，需要自己编译
    redis-sentinel /path/to/sentinel.conf
    # 使用 redis-server 程序启动, 一定可用
    # 测试时可以加上   --protected-mode no, 在不设置密码情况下访问redis
    redis-server /path/to/sentinel.conf --sentinel

4. 验证哨兵运行情况

        通过redis-cli 连接到sentinel 服务内部：
        redis-cli -p 26379   # 连接到sentinel
        info sentinel         # 查看哨兵信息
        SENTINEL slaves mymaster    # 查看master下的slave服务器情况
        SENTINEL sentinels mymaster    # 查看master的哨兵服务器列表
        SENTINEL get-master-addr-by-name mymaster    # 获取master地址信息

5. 故障模拟

将master节点关闭后，等待一段时间，再获取master地址看看。master已经切换了。

   SENTINEL get-master-addr-by-name mymaster    # 获取master地址信息

三、客户端使用哨兵系统

哨兵系统搭建好之后，就可以提供服务了。那么，如何提供服务呢？从最前面的两张架构图中，我们可以看到，sentinel 差不多是作为一个配置中心或者存在的，它只会为客户端提供master/slave的相关信息，而并不会直接代替redis实例进行存取操作。所以，哨兵模式，需要客户端做更多的工作，原来的直接连接redis变为间接从sentinel获取信息，再连接，还要维护可能的信息变更。

当然，这种工作一般是要交给sdk做的，实现原理也差不多，我们就以 jedis 作为切入点，详解下客户端如何使用sentinel.

1. 引入pom依赖

        <dependency>
            <groupId>redis.clients</groupId>
            <artifactId>jedis</artifactId>
            <version>2.9.0</version>
        </dependency>

2. 单元测试

public class RedisSentinelTest {

    @Test
    public void testSentinel() throws Exception {
        String masterName = "mymaster";
        // 只需设置sentinel信息，真实的 redis实例信息由 sentinel 提供
        Set<String> sentinels = new HashSet<>();
        sentinels.add("127.0.0.1:26379");
        sentinels.add("127.0.0.1:26378");
        sentinels.add("127.0.0.1:26377");

        JedisSentinelPool pool = new JedisSentinelPool(masterName, sentinels);
        Jedis jedis = pool.getResource();
        String key = "key1";
        String value = "Value1";
        // set get 测试哨兵系统是否可用
        jedis.set(key, value);
        System.out.println("set a value to Redis over. " + key + "->" + value);
        value = jedis.get("key1");
        System.out.println("get a value from Redis over. " + key + "->" + value);
        pool.close();
    }
}

3. sentinel 处理过程解析

jedis的sdk中已经将哨兵封装得和普通的redis实例请求差不多了，所以，我们需要深入理解下其处理过程。

首先是在初始化 JedisSentinelPool 时，其会与sentinel列表中选择一个与其建立连接：

    // redis.clients.jedis.JedisSentinelPool#JedisSentinelPool
  public JedisSentinelPool(String masterName, Set<String> sentinels) {
    this(masterName, sentinels, new GenericObjectPoolConfig(), Protocol.DEFAULT_TIMEOUT, null,
        Protocol.DEFAULT_DATABASE);
  }
  public JedisSentinelPool(String masterName, Set<String> sentinels,
      final GenericObjectPoolConfig poolConfig, int timeout, final String password,
      final int database) {
    this(masterName, sentinels, poolConfig, timeout, timeout, password, database);
  }
  public JedisSentinelPool(String masterName, Set<String> sentinels,
      final GenericObjectPoolConfig poolConfig, final int timeout, final int soTimeout,
      final String password, final int database) {
    this(masterName, sentinels, poolConfig, timeout, soTimeout, password, database, null);
  }
  public JedisSentinelPool(String masterName, Set<String> sentinels,
      final GenericObjectPoolConfig poolConfig, final int connectionTimeout, final int soTimeout,
      final String password, final int database, final String clientName) {
    this.poolConfig = poolConfig;
    this.connectionTimeout = connectionTimeout;
    this.soTimeout = soTimeout;
    this.password = password;
    this.database = database;
    this.clientName = clientName;
    // 从sentinel中获取master信息，关键
    HostAndPort master = initSentinels(sentinels, masterName);
    // 初始化连接池，非本文重点
    initPool(master);
  }
  
  private HostAndPort initSentinels(Set<String> sentinels, final String masterName) {

    HostAndPort master = null;
    boolean sentinelAvailable = false;

    log.info("Trying to find master from available Sentinels...");
    // 依次遍历 sentinels, 直到找到一个可用的sentinel
    for (String sentinel : sentinels) {
      final HostAndPort hap = HostAndPort.parseString(sentinel);

      log.fine("Connecting to Sentinel " + hap);

      Jedis jedis = null;
      try {
        jedis = new Jedis(hap.getHost(), hap.getPort());
        // 向sentinel发送命令请求: SENTINEL get-master-addr-by-name mymaster, 获取master地址信息
        List<String> masterAddr = jedis.sentinelGetMasterAddrByName(masterName);

        // connected to sentinel...
        sentinelAvailable = true;

        if (masterAddr == null || masterAddr.size() != 2) {
          log.warning("Can not get master addr, master name: " + masterName + ". Sentinel: " + hap
              + ".");
          continue;
        }

        master = toHostAndPort(masterAddr);
        log.fine("Found Redis master at " + master);
        break;
      } catch (JedisException e) {
        // resolves #1036, it should handle JedisException there's another chance
        // of raising JedisDataException
        log.warning("Cannot get master address from sentinel running @ " + hap + ". Reason: " + e
            + ". Trying next one.");
      } finally {
        if (jedis != null) {
          jedis.close();
        }
      }
    }

    if (master == null) {
      if (sentinelAvailable) {
        // can connect to sentinel, but master name seems to not
        // monitored
        throw new JedisException("Can connect to sentinel, but " + masterName
            + " seems to be not monitored...");
      } else {
        throw new JedisConnectionException("All sentinels down, cannot determine where is "
            + masterName + " master is running...");
      }
    }

    log.info("Redis master running at " + master + ", starting Sentinel listeners...");
    // 为每个 sentinel, 建立一个监听线程, 监听 sentinel 的 +switch-master 信息
    // 当master发生变化时，重新初始化连接池
    for (String sentinel : sentinels) {
      final HostAndPort hap = HostAndPort.parseString(sentinel);
      MasterListener masterListener = new MasterListener(masterName, hap.getHost(), hap.getPort());
      // whether MasterListener threads are alive or not, process can be stopped
      masterListener.setDaemon(true);
      masterListeners.add(masterListener);
      masterListener.start();
    }

    return master;
  }
    // 每个 sentinel 监听线程事务处理流程如下
    // redis.clients.jedis.JedisSentinelPool.MasterListener#run
    @Override
    public void run() {

      running.set(true);

      while (running.get()) {

        j = new Jedis(host, port);

        try {
          // double check that it is not being shutdown
          if (!running.get()) {
            break;
          }
            // SUBSCRIBE +switch-master
          j.subscribe(new JedisPubSub() {
            @Override
            public void onMessage(String channel, String message) {
              log.fine("Sentinel " + host + ":" + port + " published: " + message + ".");

              String[] switchMasterMsg = message.split(" ");
                // 格式为: masterName xx xx masterHost masterPort
              if (switchMasterMsg.length > 3) {

                if (masterName.equals(switchMasterMsg[0])) {
                  initPool(toHostAndPort(Arrays.asList(switchMasterMsg[3], switchMasterMsg[4])));
                } else {
                  log.fine("Ignoring message on +switch-master for master name "
                      + switchMasterMsg[0] + ", our master name is " + masterName);
                }

              } else {
                log.severe("Invalid message received on Sentinel " + host + ":" + port
                    + " on channel +switch-master: " + message);
              }
            }
          }, "+switch-master");

        } catch (JedisConnectionException e) {

          if (running.get()) {
            log.log(Level.SEVERE, "Lost connection to Sentinel at " + host + ":" + port
                + ". Sleeping 5000ms and retrying.", e);
            try {
              Thread.sleep(subscribeRetryWaitTimeMillis);
            } catch (InterruptedException e1) {
              log.log(Level.SEVERE, "Sleep interrupted: ", e1);
            }
          } else {
            log.fine("Unsubscribing from Sentinel at " + host + ":" + port);
          }
        } finally {
          j.close();
        }
      }
    }

从上面流程我们也就可以看出客户端是如何处理 sentinel 和 redis 的关系的了。简单来说就是通过 sentinel get-master-addr-by-name xxx, 获取master地址信息，然后连接过去就可以了。在master发生变化时，通过pub/sub订阅sentinel信息，从而进行连接池的重置。

这个连接池又是如何处理的呢？我们可以简单看一下：

    // redis.clients.jedis.JedisSentinelPool#initPool
  private void initPool(HostAndPort master) {
    if (!master.equals(currentHostMaster)) {
      currentHostMaster = master;
      if (factory == null) {
        factory = new JedisFactory(master.getHost(), master.getPort(), connectionTimeout,
            soTimeout, password, database, clientName, false, null, null, null);
        initPool(poolConfig, factory);
      } else {
        factory.setHostAndPort(currentHostMaster);
        // although we clear the pool, we still have to check the
        // returned object
        // in getResource, this call only clears idle instances, not
        // borrowed instances
        internalPool.clear();
      }

      log.info("Created JedisPool to master at " + master);
    }
  }
  // redis.clients.util.Pool#initPool
  public void initPool(final GenericObjectPoolConfig poolConfig, PooledObjectFactory<T> factory) {

    if (this.internalPool != null) {
      try {
        closeInternalPool();
      } catch (Exception e) {
      }
    }

    this.internalPool = new GenericObjectPool<T>(factory, poolConfig);
  }

当要向redis写入数据时，会先从连接池里获取一个连接实例，其池化框架使用的是 GenericObjectPool 的通用能力，调用 JedisFactory 的 makeObject() 方法进行创建 ：

  // redis.clients.jedis.JedisSentinelPool#getResource
  @Override
  public Jedis getResource() {
    while (true) {
      // 调用父类方法获取实例
      Jedis jedis = super.getResource();
      jedis.setDataSource(this);

      // get a reference because it can change concurrently
      final HostAndPort master = currentHostMaster;
      final HostAndPort connection = new HostAndPort(jedis.getClient().getHost(), jedis.getClient()
          .getPort());
      // host:port 比对，如果master未变化，说明获取到了正确的连接，返回
      if (master.equals(connection)) {
        // connected to the correct master
        return jedis;
      } 
      // 如果master 发生了切换，则将当前连接释放，继续尝试获取master连接
      else {
        returnBrokenResource(jedis);
      }
    }
  }
  // redis.clients.util.Pool#getResource
  public T getResource() {
    try {
      return internalPool.borrowObject();
    } catch (NoSuchElementException nse) {
      throw new JedisException("Could not get a resource from the pool", nse);
    } catch (Exception e) {
      throw new JedisConnectionException("Could not get a resource from the pool", e);
    }
  }
    // org.apache.commons.pool2.impl.GenericObjectPool#borrowObject()
    @Override
    public T borrowObject() throws Exception {
        return borrowObject(getMaxWaitMillis());
    }
    // org.apache.commons.pool2.impl.GenericObjectPool#borrowObject(long)
    public T borrowObject(final long borrowMaxWaitMillis) throws Exception {
        assertOpen();

        final AbandonedConfig ac = this.abandonedConfig;
        if (ac != null && ac.getRemoveAbandonedOnBorrow() &&
                (getNumIdle() < 2) &&
                (getNumActive() > getMaxTotal() - 3) ) {
            removeAbandoned(ac);
        }

        PooledObject<T> p = null;

        // Get local copy of current config so it is consistent for entire
        // method execution
        final boolean blockWhenExhausted = getBlockWhenExhausted();

        boolean create;
        final long waitTime = System.currentTimeMillis();

        while (p == null) {
            create = false;
            p = idleObjects.pollFirst();
            if (p == null) {
                // 没有获取到连接时，主动创建一个
                p = create();
                if (p != null) {
                    create = true;
                }
            }
            if (blockWhenExhausted) {
                if (p == null) {
                    if (borrowMaxWaitMillis < 0) {
                        p = idleObjects.takeFirst();
                    } else {
                        p = idleObjects.pollFirst(borrowMaxWaitMillis,
                                TimeUnit.MILLISECONDS);
                    }
                }
                if (p == null) {
                    throw new NoSuchElementException(
                            "Timeout waiting for idle object");
                }
            } else {
                if (p == null) {
                    throw new NoSuchElementException("Pool exhausted");
                }
            }
            if (!p.allocate()) {
                p = null;
            }

            if (p != null) {
                try {
                    // 确保激活当前数据库
                    factory.activateObject(p);
                } catch (final Exception e) {
                    try {
                        destroy(p);
                    } catch (final Exception e1) {
                        // Ignore - activation failure is more important
                    }
                    p = null;
                    if (create) {
                        final NoSuchElementException nsee = new NoSuchElementException(
                                "Unable to activate object");
                        nsee.initCause(e);
                        throw nsee;
                    }
                }
                if (p != null && (getTestOnBorrow() || create && getTestOnCreate())) {
                    boolean validate = false;
                    Throwable validationThrowable = null;
                    try {
                        validate = factory.validateObject(p);
                    } catch (final Throwable t) {
                        PoolUtils.checkRethrow(t);
                        validationThrowable = t;
                    }
                    if (!validate) {
                        try {
                            destroy(p);
                            destroyedByBorrowValidationCount.incrementAndGet();
                        } catch (final Exception e) {
                            // Ignore - validation failure is more important
                        }
                        p = null;
                        if (create) {
                            final NoSuchElementException nsee = new NoSuchElementException(
                                    "Unable to validate object");
                            nsee.initCause(validationThrowable);
                            throw nsee;
                        }
                    }
                }
            }
        }

        updateStatsBorrow(p, System.currentTimeMillis() - waitTime);

        return p.getObject();
    }

    /**
     * Attempts to create a new wrapped pooled object.
     * <p>
     * If there are {@link #getMaxTotal()} objects already in circulation
     * or in process of being created, this method returns null.
     *
     * @return The new wrapped pooled object
     *
     * @throws Exception if the object factory's {@code makeObject} fails
     */
    private PooledObject<T> create() throws Exception {
        int localMaxTotal = getMaxTotal();
        // This simplifies the code later in this method
        if (localMaxTotal < 0) {
            localMaxTotal = Integer.MAX_VALUE;
        }

        // Flag that indicates if create should:
        // - TRUE:  call the factory to create an object
        // - FALSE: return null
        // - null:  loop and re-test the condition that determines whether to
        //          call the factory
        Boolean create = null;
        while (create == null) {
            synchronized (makeObjectCountLock) {
                final long newCreateCount = createCount.incrementAndGet();
                if (newCreateCount > localMaxTotal) {
                    // The pool is currently at capacity or in the process of
                    // making enough new objects to take it to capacity.
                    createCount.decrementAndGet();
                    if (makeObjectCount == 0) {
                        // There are no makeObject() calls in progress so the
                        // pool is at capacity. Do not attempt to create a new
                        // object. Return and wait for an object to be returned
                        create = Boolean.FALSE;
                    } else {
                        // There are makeObject() calls in progress that might
                        // bring the pool to capacity. Those calls might also
                        // fail so wait until they complete and then re-test if
                        // the pool is at capacity or not.
                        makeObjectCountLock.wait();
                    }
                } else {
                    // The pool is not at capacity. Create a new object.
                    makeObjectCount++;
                    create = Boolean.TRUE;
                }
            }
        }

        if (!create.booleanValue()) {
            return null;
        }

        final PooledObject<T> p;
        try {
            // 调用指定factory的 makeObject() 方法
            p = factory.makeObject();
        } catch (final Exception e) {
            createCount.decrementAndGet();
            throw e;
        } finally {
            synchronized (makeObjectCountLock) {
                makeObjectCount--;
                makeObjectCountLock.notifyAll();
            }
        }

        final AbandonedConfig ac = this.abandonedConfig;
        if (ac != null && ac.getLogAbandoned()) {
            p.setLogAbandoned(true);
        }

        createdCount.incrementAndGet();
        allObjects.put(new IdentityWrapper<T>(p.getObject()), p);
        return p;
    }
  // 使用 JedisFactory 创建一个连接到 master
  // redis.clients.jedis.JedisFactory#makeObject
  @Override
  public PooledObject<Jedis> makeObject() throws Exception {
    final HostAndPort hostAndPort = this.hostAndPort.get();
    final Jedis jedis = new Jedis(hostAndPort.getHost(), hostAndPort.getPort(), connectionTimeout,
        soTimeout, ssl, sslSocketFactory, sslParameters, hostnameVerifier);

    try {
      jedis.connect();
      // 如果存在密码设置，则进行 auth xxx 操作
      // redis 配置: requirepass xxx
      if (null != this.password) {
        jedis.auth(this.password);
      }
      if (database != 0) {
        jedis.select(database);
      }
      if (clientName != null) {
        jedis.clientSetname(clientName);
      }
    } catch (JedisException je) {
      jedis.close();
      throw je;
    }

    return new DefaultPooledObject<Jedis>(jedis);

  }
  // redis.clients.jedis.JedisFactory#activateObject
  @Override
  public void activateObject(PooledObject<Jedis> pooledJedis) throws Exception {
    final BinaryJedis jedis = pooledJedis.getObject();
    if (jedis.getDB() != database) {
      jedis.select(database);
    }

  }

获取到client连接后，主可以任意地通过网络io与真实redis进行交互了。哨兵也不会成为性能问题了。

四、思考

哨兵模式的出现，仅为了解决单机的高可用问题，而并不会解决单机容量问题（集群模式会处理这个问题）。在当前的互联网环境中，应用面也许没有那么广。但思路是值得借鉴的。

Sentinel 在配置时只需配置master地址即可，其slave信息，sentinel信息，都是通过master来推断的。所以，一定要确保在启动时master是可用的，否则系统本身必须无法启动。

如果redis中设置了密码，则要求必须保持全部一致，这在一定程度上会有些误会。

redis Sentinel 本身是一个对等集群系统，连接任意节点结果都是一样的，节点间保持通过pub/sub两两通信。