storm-kafka源码走读之KafkaSpout

from: http://blog.csdn.net/wzhg0508/article/details/40903919

（五）storm-kafka源码走读之KafkaSpout

原创 2014年11月08日 14:09:06

• 标签：
• Storm /
• kafka /
• 大数据 /
• 实时计算

• 3458

现在开始介绍KafkaSpout源码了。

开始时，在open方法中做一些初始化，

........................

        _state = new ZkState(stateConf);

        _connections = new DynamicPartitionConnections(_spoutConfig, KafkaUtils.makeBrokerReader(conf, _spoutConfig));

        // using TransactionalState like this is a hack
        int totalTasks = context.getComponentTasks(context.getThisComponentId()).size();
        if (_spoutConfig.hosts instanceof StaticHosts) {
            _coordinator = new StaticCoordinator(_connections, conf, _spoutConfig, _state, context.getThisTaskIndex(), totalTasks, _uuid);
        } else {
            _coordinator = new ZkCoordinator(_connections, conf, _spoutConfig, _state, context.getThisTaskIndex(), totalTasks, _uuid);
        }
............

前后省略了一些代码，关于metric这系列暂时不介绍。主要是初始化Zookeeper连接zkstate，把kafka Partition 与broker关系对应起来（初始化DynamicPartitionConnections），在DynamicPartitionConnections构造函数需要传入一个brokerReader，我们是zkHosts，看KafkaUtils代码就知道采用的是ZkBrokerReader，来看下ZkBrokerReader的构造函数代码

public ZkBrokerReader(Map conf, String topic, ZkHosts hosts) {
        try {
            reader = new DynamicBrokersReader(conf, hosts.brokerZkStr, hosts.brokerZkPath, topic);
            cachedBrokers = reader.getBrokerInfo();
            lastRefreshTimeMs = System.currentTimeMillis();
            refreshMillis = hosts.refreshFreqSecs * 1000L;
        } catch (java.net.SocketTimeoutException e) {
            LOG.warn("Failed to update brokers", e);
        }

    }

有一个refreshMillis参数，这个参数是定时更新zk中partition的信息，

//ZkBrokerReader
    @Override
    public GlobalPartitionInformation getCurrentBrokers() {
        long currTime = System.currentTimeMillis();
        if (currTime > lastRefreshTimeMs + refreshMillis) { // 当前时间大于和上次更新时间之差大于refreshMillis
            try {
                LOG.info("brokers need refreshing because " + refreshMillis + "ms have expired");
                cachedBrokers = reader.getBrokerInfo();
                lastRefreshTimeMs = currTime;
            } catch (java.net.SocketTimeoutException e) {
                LOG.warn("Failed to update brokers", e);
            }
        }
        return cachedBrokers;
    }
    // 下面是调用DynamicBrokersReader 的代码
    /**
     * Get all partitions with their current leaders
     */
    public GlobalPartitionInformation getBrokerInfo() throws SocketTimeoutException {
      GlobalPartitionInformation globalPartitionInformation = new GlobalPartitionInformation();
        try {
            int numPartitionsForTopic = getNumPartitions();
            String brokerInfoPath = brokerPath();
            for (int partition = 0; partition < numPartitionsForTopic; partition++) {
                int leader = getLeaderFor(partition);
                String path = brokerInfoPath + "/" + leader;
                try {
                    byte[] brokerData = _curator.getData().forPath(path);
                    Broker hp = getBrokerHost(brokerData);
                    globalPartitionInformation.addPartition(partition, hp);
                } catch (org.apache.zookeeper.KeeperException.NoNodeException e) {
                    LOG.error("Node {} does not exist ", path);
                }
            }
        } catch (SocketTimeoutException e) {
                    throw e;
        } catch (Exception e) {
            throw new RuntimeException(e);
        }
        LOG.info("Read partition info from zookeeper: " + globalPartitionInformation);
        return globalPartitionInformation;
    }

GlobalPartitionInformation是一个Iterator类，存放了paritition与broker之间的对应关系，DynamicPartitionConnections中维护Kafka Consumer与parittion之间的关系，每个Consumer读取哪些paritition信息。这个COnnectionInfo信息会在storm.kafka.ZkCoordinator中会被初始化和更新，需要提到的一点是一个KafkaSpout包含一个SimpleConsumer

//storm.kafka.DynamicPartitionConnections
    static class ConnectionInfo {
        SimpleConsumer consumer;
        Set<Integer> partitions = new HashSet();

        public ConnectionInfo(SimpleConsumer consumer) {
            this.consumer = consumer;
        }
    }

再看ZkCoordinator类，看其构造函数

//storm.kafka.ZkCoordinator
    public ZkCoordinator(DynamicPartitionConnections connections, Map stormConf, SpoutConfig spoutConfig, ZkState state, int taskIndex, int totalTasks, String topologyInstanceId, DynamicBrokersReader reader) {
        _spoutConfig = spoutConfig;
        _connections = connections;
        _taskIndex = taskIndex;
        _totalTasks = totalTasks;
        _topologyInstanceId = topologyInstanceId;
        _stormConf = stormConf;
        _state = state;
        ZkHosts brokerConf = (ZkHosts) spoutConfig.hosts;
        _refreshFreqMs = brokerConf.refreshFreqSecs * 1000;
        _reader = reader;
    }

_refreshFreqMs就是定时更新zk partition到本地的操作，在kafkaSpout中nextTuple方法中每次都会去调用ZkCoordinator的getMyManagedPartitions方法。该方法根据_refreshFreqMs参数定时更新partition信息

//storm.kafka.ZkCoordinator
    @Override
    public List<PartitionManager> getMyManagedPartitions() {
        if (_lastRefreshTime == null || (System.currentTimeMillis() - _lastRefreshTime) > _refreshFreqMs) {
            refresh();
            _lastRefreshTime = System.currentTimeMillis();
        }
        return _cachedList;
    }

    @Override
    public void refresh() {
        try {
            LOG.info(taskId(_taskIndex, _totalTasks) + "Refreshing partition manager connections");
            GlobalPartitionInformation brokerInfo = _reader.getBrokerInfo();
            List<Partition> mine = KafkaUtils.calculatePartitionsForTask(brokerInfo, _totalTasks, _taskIndex);

            Set<Partition> curr = _managers.keySet();
            Set<Partition> newPartitions = new HashSet<Partition>(mine);
            newPartitions.removeAll(curr);

            Set<Partition> deletedPartitions = new HashSet<Partition>(curr);
            deletedPartitions.removeAll(mine);

            LOG.info(taskId(_taskIndex, _totalTasks) + "Deleted partition managers: " + deletedPartitions.toString());

            for (Partition id : deletedPartitions) {
                PartitionManager man = _managers.remove(id);
                man.close();
            }
            LOG.info(taskId(_taskIndex, _totalTasks) + "New partition managers: " + newPartitions.toString());

            for (Partition id : newPartitions) {
                PartitionManager man = new PartitionManager(_connections, _topologyInstanceId, _state, _stormConf, _spoutConfig, id);
                _managers.put(id, man);
            }

        } catch (Exception e) {
            throw new RuntimeException(e);
        }
        _cachedList = new ArrayList<PartitionManager>(_managers.values());
        LOG.info(taskId(_taskIndex, _totalTasks) + "Finished refreshing");
    }

其中每个Consumer分配partition的算法是KafkaUtils.calculatePartitionsForTask(brokerInfo, _totalTasks, _taskIndex);

主要做的工作就是获取并行的task数，与当前partition做比较，得出一个COnsumer要负责哪些parititons的读取，具体算法去kafka文档吧

以上在KafkaSpout中做完了初始化操作，下面开始取数据发射数据了，来看nextTuple方法

// storm.kafka.KafkaSpout
    @Override
    public void nextTuple() {
        List<PartitionManager> managers = _coordinator.getMyManagedPartitions();
        for (int i = 0; i < managers.size(); i++) {

            try {
                // in case the number of managers decreased
                _currPartitionIndex = _currPartitionIndex % managers.size();
                EmitState state = managers.get(_currPartitionIndex).next(_collector);
                if (state != EmitState.EMITTED_MORE_LEFT) {
                    _currPartitionIndex = (_currPartitionIndex + 1) % managers.size();
                }
                if (state != EmitState.NO_EMITTED) {
                    break;
                }
            } catch (FailedFetchException e) {
                LOG.warn("Fetch failed", e);
                _coordinator.refresh();
            }
        }

        long now = System.currentTimeMillis();
        if ((now - _lastUpdateMs) > _spoutConfig.stateUpdateIntervalMs) {
            commit();
        }
    }

看完上述代码可知，所有的操作都是在PartitionManager中进行的，PartitionManager中会读取message信息，然后进行发射，主要逻辑在PartitionManager的next方法中

//returns false if it's reached the end of current batch
    public EmitState next(SpoutOutputCollector collector) {
        if (_waitingToEmit.isEmpty()) {
            fill();
        }
        while (true) {
            MessageAndRealOffset toEmit = _waitingToEmit.pollFirst();
            if (toEmit == null) {
                return EmitState.NO_EMITTED;
            }
            Iterable<List<Object>> tups = KafkaUtils.generateTuples(_spoutConfig, toEmit.msg);
            if (tups != null) {
                for (List<Object> tup : tups) {
                    collector.emit(tup, new KafkaMessageId(_partition, toEmit.offset));
                }
                break;
            } else {
                ack(toEmit.offset);
            }
        }
        if (!_waitingToEmit.isEmpty()) {
            return EmitState.EMITTED_MORE_LEFT;
        } else {
            return EmitState.EMITTED_END;
        }
    }

如果_waitingToEmit列表为空，则去读取msg，然后进行逐条发射，每发射一条，break一下，返回EMIT_MORE_LEFT给KafkaSpout的nextTuple方法中，，然后进行判断是否该paritition读取的一次读取的message buffer size是否已发射完毕，如果发射完毕就进行下一个partition 数据读取和发射，

注意的一点是，并不是一次把该partition的所有待发射的msg都发射完再commit offset到zk，而是发射一条，判断一下是否到了该commit的时候了（开始时设置的定时commit时间间隔），笔者认为这样做的原因是为了好控制fail

KafkaSpout中的ack，fail，commit操作全部交给了PartitionManager来做，看代码

@Override
    public void ack(Object msgId) {
        KafkaMessageId id = (KafkaMessageId) msgId;
        PartitionManager m = _coordinator.getManager(id.partition);
        if (m != null) {
            m.ack(id.offset);
        }
    }

    @Override
    public void fail(Object msgId) {
        KafkaMessageId id = (KafkaMessageId) msgId;
        PartitionManager m = _coordinator.getManager(id.partition);
        if (m != null) {
            m.fail(id.offset);
        }
    }

    @Override
    public void deactivate() {
        commit();
    }

    @Override
    public void declareOutputFields(OutputFieldsDeclarer declarer) {
        declarer.declare(_spoutConfig.scheme.getOutputFields());
    }

    private void commit() {
        _lastUpdateMs = System.currentTimeMillis();
        for (PartitionManager manager : _coordinator.getMyManagedPartitions()) {
            manager.commit();
        }
    }

所以PartitionManager是KafkaSpout的核心，很晚了，都3点多了，后续会不上PartitionManager的分析，晚安

版权声明：本文为博主原创文章，未经博主允许不得转载。

• 本文已收录于以下专栏：
• Storm-kafka源码浅谈