JedisCluster模式尝试进行批量操作
搭建完redis集群后,可以通过jedis的JedisCluster来访问Redis集群,这里列出使用jedisCluster的spring bean配置方式:
<bean id="jedisPoolConfig" class="redis.clients.jedis.JedisPoolConfig"> <property name="maxIdle" value="5" /> <property name="maxTotal" value="1024" /> <property name="maxWaitMillis" value="5000" /> <property name="testOnBorrow" value="true" /> </bean> <bean id="redisCluster" class="redis.clients.jedis.JedisCluster" <constructor-arg name=“nodes"> <set> <ref bean="hostport1" /> <ref bean="hostport2" /> <ref bean="hostport3" /> <ref bean="hostport4" /> <ref bean="hostport5" /> <ref bean="hostport6" /> </set> </constructor-arg> <constructor-arg name="timeout" value="6000" /> <constructor-arg name="poolConfig" <ref bean="jedisPoolConfig" /> </constructor-arg> </bean>
//此处省略hostport1....6的配置
本质上,JedisCluster中的JedisPool同样也是实现了apache common pool2的对象池,其中的getResource可以拿到对应的Jedis连接。
JedisCluster是如何构建整个连接池
正如类图中所呈现的,JedisClusterConnectionHandler中使用了JedisClusterInfoCache作为缓存初始化容器,将Set<HostAndPort>作为JedisClusterConnectionHandler的构造函数参数传递过去之后,进行初始化slotsCache操作,
public JedisClusterConnectionHandler(Set<HostAndPort> nodes, final GenericObjectPoolConfig poolConfig, int connectionTimeout, int soTimeout) { this.cache = new JedisClusterInfoCache(poolConfig, connectionTimeout, soTimeout); initializeSlotsCache(nodes, poolConfig); } private void initializeSlotsCache(Set<HostAndPort> startNodes, GenericObjectPoolConfig poolConfig) { for (HostAndPort hostAndPort : startNodes) { Jedis jedis = new Jedis(hostAndPort.getHost(), hostAndPort.getPort()); try { cache.discoverClusterNodesAndSlots(jedis); break; } catch (JedisConnectionException e) { // try next nodes } finally { if (jedis != null) { jedis.close(); } } } for (HostAndPort node : startNodes) { cache.setNodeIfNotExist(node); } }
在cache.discoverClusterNodesAndSlots中,用到了Jedis.clusterNodes,它可以通过该Redis连接找到其他连接的相关配置,例如可以发现整个集群的配置,其中三个master,三个slave,并且能够识别出自身连接,可参考文档:http://redis.io/commands/cluster-nodes:
./redis-cli -c -h xxx -p 63xx cluster nodes e54b82fd2b5ab238906cff7fc6250a7bc66c6fec 192.168.1.1xx:6389 master - 0 1469151811362 31 connected 0-5460 166baa38c8ab56339c11f0446257c7a6059a219b 192.168.1.1xx:6389 slave 1609b090dfaaac702449b72d30b2330521ce2506 0 1469151812364 29 connected 1609b090dfaaac702449b72d30b2330521ce2506 192.168.1.1xx:6390 master - 0 1469151811362 29 connected 10923-16383 539627a393aa43e82ca8c16d1e935611fec4e709 192.168.1.1xx:6388 myself,master - 0 0 28 connected 5461-10922 d9b3738ff16e99075242b865a0b6cc137c20d502 192.168.1.1xx:6390 slave 539627a393aa43e82ca8c16d1e935611fec4e709 0 1469151810859 28 connected 101227d3cb13f08a47ad2afe1b348d0efc3cb3b0 192.168.1.1xx:6388 slave e54b82fd2b5ab238906cff7fc6250a7bc66c6fec 0 1469151810357 31 connected
命令返回的结果中,可以看到有3个master,3个slave,而且每个slave都有对应的masterid,每个master都有对应的slot范围。
在ClusterNodeInformationParser中,去解析每一行并将对应的slot填充进去,因为只有master上有slot,因此不会填充slave的slot:
public void discoverClusterSlots(Jedis jedis) { w.lock(); try { this.slots.clear(); List<Object> slots = jedis.clusterSlots(); for (Object slotInfoObj : slots) { List<Object> slotInfo = (List<Object>) slotInfoObj; if (slotInfo.size() <= 2) { continue; } List<Integer> slotNums = getAssignedSlotArray(slotInfo); // hostInfos List<Object> hostInfos = (List<Object>) slotInfo.get(2); if (hostInfos.size() <= 0) { continue; } // at this time, we just use master, discard slave information HostAndPort targetNode = generateHostAndPort(hostInfos); setNodeIfNotExist(targetNode); assignSlotsToNode(slotNums, targetNode); } } finally { w.unlock(); } }
因此,当我们正常地通过访问JedisCluster的get/set时,通过计算key的slot来获取对应的Jedis Connection,根本不会使用到slave,只会访问master节点。只有一种情况,在tryRandomMode开启时(此时,正常通过slot无法获取有效连接时,可能考虑重新排序)。
@Override public Jedis getConnection() { // In antirez's redis-rb-cluster implementation, // getRandomConnection always return valid connection (able to // ping-pong) // or exception if all connections are invalid List<JedisPool> pools = getShuffledNodesPool(); for (JedisPool pool : pools) { Jedis jedis = null; try { jedis = pool.getResource(); if (jedis == null) { continue; } String result = jedis.ping(); if (result.equalsIgnoreCase("pong")) return jedis; pool.returnBrokenResource(jedis); } catch (JedisConnectionException ex) { if (jedis != null) { pool.returnBrokenResource(jedis); } } } throw new JedisConnectionException("no reachable node in cluster"); }
但此时拿到一个slave节点的可用Connection是非常危险的,加入当前的操作为写操作,将某个字段写入Redis时,由于master不会从slave节点进行复制,会导致该数据操作没有被持久化至master上。
开发基于redis key统一批量处理的中间层
根据redis cluster nodes命令来进行,该命令可以识别出当前集群其余节点的所有状态,master/slave,以及检测的slot位置。
cluster nodes e54b82fd2b5ab238906cff7fc6250a7bc66c6fec 192.168.1.163:6389 master - 0 1469600305090 31 connected 0-5460 166baa38c8ab56339c11f0446257c7a6059a219b 192.168.1.165:6389 slave 1609b090dfaaac702449b72d30b2330521ce2506 0 1469600304588 29 connected 1609b090dfaaac702449b72d30b2330521ce2506 192.168.1.163:6390 master - 0 1469600305592 29 connected 10923-16383 539627a393aa43e82ca8c16d1e935611fec4e709 192.168.1.163:6388 myself,master - 0 0 28 connected 5461-10922 d9b3738ff16e99075242b865a0b6cc137c20d502 192.168.1.165:6390 slave 539627a393aa43e82ca8c16d1e935611fec4e709 0 1469600305090 28 connected 101227d3cb13f08a47ad2afe1b348d0efc3cb3b0 192.168.1.165:6388 slave e54b82fd2b5ab238906cff7fc6250a7bc66c6fec 0 1469600304088 31 connected
cluster nodes 命令的输出有点儿复杂, 它的每一行都是由以下信息组成的:
- 节点 ID :例如 3fc783611028b1707fd65345e763befb36454d73 。
- ip:port :节点的 IP 地址和端口号, 例如 127.0.0.1:7000 , 其中 :0 表示的是客户端当前连接的 IP 地址和端口号。
- flags :节点的角色(例如 master 、 slave 、 myself )以及状态(例如 fail ,等等)。
- 如果节点是一个从节点的话, 那么跟在 flags 之后的将是主节点的节点 ID : 例如 127.0.0.1:7002 的主节点的节点 ID 就是 3c3a0c74aae0b56170ccb03a76b60cfe7dc1912e 。
- 集群最近一次向节点发送 PING 命令之后, 过去了多长时间还没接到回复。
- 节点最近一次返回 PONG 回复的时间。
- 节点的配置纪元(configuration epoch):详细信息请参考 Redis 集群规范 。
- 本节点的网络连接情况:例如 connected 。
- 节点目前包含的槽:例如 127.0.0.1:7001 目前包含号码为 5960 至 10921 的哈希槽。
可以看出redis cluster的slot范围:0-16383,可以采用二分查找的方式,以上面为例,可以分成3个部分的范围slot,以其开头为标识,通过Collections.binarySearch来进行二分查找搜索:
0——5460,5461——10922,10923——16383;
通过JedisPool/JedisConnection初始化客户端连接,并建立slotStarts,其中的ClusterNodeObject作为描述cluster nodes命令返回的行对象:
public void initCluster() { if (jedis instanceof BinaryJedisCluster) { BinaryJedisCluster jedisCluster = (BinaryJedisCluster) jedis; Map<String, JedisPool> clusterNodes = jedisCluster.getClusterNodes(); Map<String, ClusterNodeObject> hpToNodeObjectMap = new HashMap<>(clusterNodes.size()); for (Map.Entry<String, JedisPool> entry : clusterNodes.entrySet()) { JedisPool jedisPool = entry.getValue(); Jedis jedis = jedisPool.getResource(); String clusterNodesCommand = jedis.clusterNodes(); String[] allNodes = clusterNodesCommand.split("\n"); for (String allNode : allNodes) { String[] splits = allNode.split(" "); String hostAndPort = splits[1]; ClusterNodeObject clusterNodeObject = new ClusterNodeObject(splits[0], splits[1], splits[2].contains("master"), splits[3], Long.parseLong(splits[4]), Long.parseLong(splits[5]), splits[6], splits[7].equalsIgnoreCase("connected"), splits.length == 9 ? splits[8] : null); hpToNodeObjectMap.put(hostAndPort, clusterNodeObject); } } List<Integer> slotStarts = new ArrayList<>(); for (ClusterNodeObject clusterNodeObject : hpToNodeObjectMap.values()) { if (clusterNodeObject.isConnected() && clusterNodeObject.isMaster()) { String slot = clusterNodeObject.getSlot(); String[] slotSplits = slot.split("-"); int slotStart = Integer.parseInt(slotSplits[0]); // int slotEnd = Integer.parseInt(slotSplits[1]); slotStarts.add(slotStart); } } Collections.sort(slotStarts); this.slotStarts = slotStarts; } }
在拿到各个redis key后,通过getSlotByKey方法,获得对应的node编号:
private int getSlotByKey(String key) { int slot = JedisClusterCRC16.getSlot(key); int slotInsertion = Collections.binarySearch(slotStarts, slot); if (slotInsertion < 0) { slotInsertion = Math.abs(slotInsertion + 2); } return slotInsertion; }
最后,当批量查询的keys数组>2时,再进行批量出,否则,只进行单独查询。
if (keys.length > 2 && jedis instanceof JedisCluster) { //如果批量请求key长度大于2,启动批量查询方式 Map<Integer, List<String>> keySlotsMapList = new HashMap<>(); for (String key : keys) { int slotByKey = getSlotByKey(key); if (!keySlotsMapList.containsKey(slotByKey)) { keySlotsMapList.put(slotByKey, new ArrayList<String>()); } keySlotsMapList.get(slotByKey).add(key); } for (Map.Entry<Integer, List<String>> entry : keySlotsMapList.entrySet()) { List<String> slotSameKeys = entry.getValue(); List<String> mgetValues = ((ZhenJedisCluster) jedis) .mget(slotSameKeys.toArray(new String[slotSameKeys.size()])); for (int i = 0; i < slotSameKeys.size(); i++) { result.set(keyList.indexOf(slotSameKeys.get(i)), mgetValues.get(i)); } } } else { for (String key : keys) { result.add(jedis.get(key)); } }
但不能跳过Jedis客户端的slot key检查,其中的批量操作依赖slot是否相同:
public T run(int keyCount, String... keys) { if (keys == null || keys.length == 0) { throw new JedisClusterException("No way to dispatch this command to Redis Cluster."); } // For multiple keys, only execute if they all share the // same connection slot. if (keys.length > 1) { int slot = JedisClusterCRC16.getSlot(keys[0]); for (int i = 1; i < keyCount; i++) { int nextSlot = JedisClusterCRC16.getSlot(keys[i]); if (slot != nextSlot) { throw new JedisClusterException("No way to dispatch this command to Redis Cluster " + "because keys have different slots."); } } } return runWithRetries(SafeEncoder.encode(keys[0]), this.redirections, false, false); }
虽然可以通过重写JedisCluster以及JedisClusterCommand类型(由于有一些依赖包访问权限的类型,需要将这些重写的类型同样放到redis.clients.jedis包中),跳过了JedisCluster的校验,仍然出现连接错误:
Exception in thread "main" redis.clients.jedis.exceptions.JedisDataException: CROSSSLOT Keys in request don't hash to the same slot at redis.clients.jedis.Protocol.processError(Protocol.java:117) at redis.clients.jedis.Protocol.process(Protocol.java:151) at redis.clients.jedis.Protocol.read(Protocol.java:205) at redis.clients.jedis.Connection.readProtocolWithCheckingBroken(Connection.java:297) at redis.clients.jedis.Connection.getBinaryMultiBulkReply(Connection.java:233) at redis.clients.jedis.Connection.getMultiBulkReply(Connection.java:226) at redis.clients.jedis.Jedis.mget(Jedis.java:355) at redis.clients.jedis.ZhenJedisCluster$129.execute(ZhenJedisCluster.java:1382) at redis.clients.jedis.ZhenJedisCluster$129.execute(ZhenJedisCluster.java:1) at redis.clients.jedis.ZhenJedisClusterCommand.runWithRetries(ZhenJedisClusterCommand.java:119) at redis.clients.jedis.ZhenJedisClusterCommand.run(ZhenJedisClusterCommand.java:51) at redis.clients.jedis.ZhenJedisCluster.mget(ZhenJedisCluster.java:1384) at com.api.pub.cache.JedisClient.batchGet(JedisClient.java:525) at com.zhen.commons.redis.test.RedisTest.main(RedisTest.java:46)
可以看出,尽管两个slot在同一个连接上能够get到值,但是在cluster模式下,是通过slot判断而非节点node判断是否可以进行mget操作,不能靠跳过jedis客户端的方案来完成类似分组操作。
我们可以通过命令行操作,同样来验证这一点,注意redis-cli连接时保证在cluster模式下运行,加入-c参数:
redis-cli -c -h 192.168.1.138 -p 6388 192.168.1.138:6388> set key1 "key1" -> Redirected to slot [9189] located at 192.168.1.137:6390 OK 192.168.1.137:6390> set key2 "key2" -> Redirected to slot [4998] located at 192.168.1.137:6389 OK 192.168.1.137:6389> set key3 "key3" OK 192.168.1.137:6389> mget key2 key3 (error) CROSSSLOT Keys in request don't hash to the same slot 192.168.1.137:6389> get key2 "key2" 192.168.1.137:6389> get key3 "key3" 192.168.1.137:6389> get key1 -> Redirected to slot [9189] located at 192.168.1.137:6390 "key1" 192.168.1.137:6390> set {aaa}1 "1" OK 192.168.1.137:6390> set {aaa}2 "2" OK 192.168.1.137:6390> mget {aaa}1 {aaa}2 1) "1" 2) "2"
因此,暂且不能在RedisCluster模式下,通过增加中间层来对批量请求进行分组,并处理到对应的slot中,理想很好,但是不能够实现,因为服务端会进行一定的限制。
只能通过HASH_TAG来实现cluster模式下的mget/mset批量操作,我们可以在命令行中通过cluster keyslot ${key}来查看某个key对应的slot,可以从Jedis客户端的源码查看对应的key slot算法:
public static int getSlot(String key) { int s = key.indexOf("{"); if (s > -1) { int e = key.indexOf("}", s + 1); if (e > -1 && e != s + 1) { key = key.substring(s + 1, e); } } // optimization with modulo operator with power of 2 // equivalent to getCRC16(key) % 16384 return getCRC16(key) & (16384 - 1); }
可以看出,keySlot算法中,如果key包含{},就会使用第一个{}内部的字符串作为hash key,这样就可以保证拥有同样{}内部字符串的key就会拥有相同slot。