聊聊storm trident的coordinator

本文主要研究一下storm trident的coordinator

实例

代码示例

    @Test
    public void testDebugTopologyBuild(){
        FixedBatchSpout spout = new FixedBatchSpout(new Fields("user", "score"), 3,
                new Values("nickt1", 4),
                new Values("nickt2", 7),
                new Values("nickt3", 8),
                new Values("nickt4", 9),
                new Values("nickt5", 7),
                new Values("nickt6", 11),
                new Values("nickt7", 5)
        );
        spout.setCycle(false);
        TridentTopology topology = new TridentTopology();
        Stream stream1 = topology.newStream("spout1",spout)
                .each(new Fields("user", "score"), new BaseFunction() {
                    @Override
                    public void execute(TridentTuple tuple, TridentCollector collector) {
                        System.out.println("tuple:"+tuple);
                    }
                },new Fields());

        topology.build();
    }
  • 这里使用的spout为FixedBatchSpout,它是IBatchSpout类型

拓扑图

MasterBatchCoordinator

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/topology/MasterBatchCoordinator.java

public class MasterBatchCoordinator extends BaseRichSpout { 
    public static final Logger LOG = LoggerFactory.getLogger(MasterBatchCoordinator.class);
    
    public static final long INIT_TXID = 1L;
    
    
    public static final String BATCH_STREAM_ID = "$batch";
    public static final String COMMIT_STREAM_ID = "$commit";
    public static final String SUCCESS_STREAM_ID = "$success";

    private static final String CURRENT_TX = "currtx";
    private static final String CURRENT_ATTEMPTS = "currattempts";
    
    private List _states = new ArrayList();
    
    TreeMap _activeTx = new TreeMap();
    TreeMap _attemptIds;
    
    private SpoutOutputCollector _collector;
    Long _currTransaction;
    int _maxTransactionActive;
    
    List _coordinators = new ArrayList();
    
    
    List _managedSpoutIds;
    List _spouts;
    WindowedTimeThrottler _throttler;
    
    boolean _active = true;
    
    public MasterBatchCoordinator(List spoutIds, List spouts) {
        if(spoutIds.isEmpty()) {
            throw new IllegalArgumentException("Must manage at least one spout");
        }
        _managedSpoutIds = spoutIds;
        _spouts = spouts;
        LOG.debug("Created {}", this);
    }

    public List getManagedSpoutIds(){
        return _managedSpoutIds;
    }

    @Override
    public void activate() {
        _active = true;
    }

    @Override
    public void deactivate() {
        _active = false;
    }
        
    @Override
    public void open(Map conf, TopologyContext context, SpoutOutputCollector collector) {
        _throttler = new WindowedTimeThrottler((Number)conf.get(Config.TOPOLOGY_TRIDENT_BATCH_EMIT_INTERVAL_MILLIS), 1);
        for(String spoutId: _managedSpoutIds) {
            _states.add(TransactionalState.newCoordinatorState(conf, spoutId));
        }
        _currTransaction = getStoredCurrTransaction();

        _collector = collector;
        Number active = (Number) conf.get(Config.TOPOLOGY_MAX_SPOUT_PENDING);
        if(active==null) {
            _maxTransactionActive = 1;
        } else {
            _maxTransactionActive = active.intValue();
        }
        _attemptIds = getStoredCurrAttempts(_currTransaction, _maxTransactionActive);

        
        for(int i=0; i<_spouts.size(); i++) {
            String txId = _managedSpoutIds.get(i);
            _coordinators.add(_spouts.get(i).getCoordinator(txId, conf, context));
        }
        LOG.debug("Opened {}", this);
    }

    @Override
    public void close() {
        for(TransactionalState state: _states) {
            state.close();
        }
        LOG.debug("Closed {}", this);
    }
    
    @Override
    public void declareOutputFields(OutputFieldsDeclarer declarer) {
        // in partitioned example, in case an emitter task receives a later transaction than it's emitted so far,
        // when it sees the earlier txid it should know to emit nothing
        declarer.declareStream(BATCH_STREAM_ID, new Fields("tx"));
        declarer.declareStream(COMMIT_STREAM_ID, new Fields("tx"));
        declarer.declareStream(SUCCESS_STREAM_ID, new Fields("tx"));
    }

    @Override
    public Map getComponentConfiguration() {
        Config ret = new Config();
        ret.setMaxTaskParallelism(1);
        ret.registerSerialization(TransactionAttempt.class);
        return ret;
    }

    //......
}
  • prepare方法首先从Config.TOPOLOGY_TRIDENT_BATCH_EMIT_INTERVAL_MILLIS(topology.trident.batch.emit.interval.millis,在defaults.yaml默认为500)读取触发batch的频率配置,然后创建WindowedTimeThrottler,其maxAmt值为1
  • 这里使用TransactionalState在zookeeper上维护transactional状态
  • 之后读取Config.TOPOLOGY_MAX_SPOUT_PENDING(topology.max.spout.pending,在defaults.yaml中默认为null)设置_maxTransactionActive,如果为null,则设置为1

MasterBatchCoordinator.nextTuple

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/topology/MasterBatchCoordinator.java

    @Override
    public void nextTuple() {
        sync();
    }

    private void sync() {
        // note that sometimes the tuples active may be less than max_spout_pending, e.g.
        // max_spout_pending = 3
        // tx 1, 2, 3 active, tx 2 is acked. there won't be a commit for tx 2 (because tx 1 isn't committed yet),
        // and there won't be a batch for tx 4 because there's max_spout_pending tx active
        TransactionStatus maybeCommit = _activeTx.get(_currTransaction);
        if(maybeCommit!=null && maybeCommit.status == AttemptStatus.PROCESSED) {
            maybeCommit.status = AttemptStatus.COMMITTING;
            _collector.emit(COMMIT_STREAM_ID, new Values(maybeCommit.attempt), maybeCommit.attempt);
            LOG.debug("Emitted on [stream = {}], [tx_status = {}], [{}]", COMMIT_STREAM_ID, maybeCommit, this);
        }
        
        if(_active) {
            if(_activeTx.size() < _maxTransactionActive) {
                Long curr = _currTransaction;
                for(int i=0; i<_maxTransactionActive; i++) {
                    if(!_activeTx.containsKey(curr) && isReady(curr)) {
                        // by using a monotonically increasing attempt id, downstream tasks
                        // can be memory efficient by clearing out state for old attempts
                        // as soon as they see a higher attempt id for a transaction
                        Integer attemptId = _attemptIds.get(curr);
                        if(attemptId==null) {
                            attemptId = 0;
                        } else {
                            attemptId++;
                        }
                        _attemptIds.put(curr, attemptId);
                        for(TransactionalState state: _states) {
                            state.setData(CURRENT_ATTEMPTS, _attemptIds);
                        }
                        
                        TransactionAttempt attempt = new TransactionAttempt(curr, attemptId);
                        final TransactionStatus newTransactionStatus = new TransactionStatus(attempt);
                        _activeTx.put(curr, newTransactionStatus);
                        _collector.emit(BATCH_STREAM_ID, new Values(attempt), attempt);
                        LOG.debug("Emitted on [stream = {}], [tx_attempt = {}], [tx_status = {}], [{}]", BATCH_STREAM_ID, attempt, newTransactionStatus, this);
                        _throttler.markEvent();
                    }
                    curr = nextTransactionId(curr);
                }
            }
        }
    }
  • nextTuple就是调用sync方法,该方法在ack及fail中均有调用;sync方法首先根据事务状态,如果需要提交,则会往MasterBatchCoordinator.COMMIT_STREAM_ID($commit)发送tuple;之后根据_maxTransactionActive以及WindowedTimeThrottler限制,符合要求才启动新的TransactionAttempt,往MasterBatchCoordinator.BATCH_STREAM_ID($batch)发送tuple,同时对WindowedTimeThrottler标记下windowEvent数量

MasterBatchCoordinator.ack

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/topology/MasterBatchCoordinator.java

    @Override
    public void ack(Object msgId) {
        TransactionAttempt tx = (TransactionAttempt) msgId;
        TransactionStatus status = _activeTx.get(tx.getTransactionId());
        LOG.debug("Ack. [tx_attempt = {}], [tx_status = {}], [{}]", tx, status, this);
        if(status!=null && tx.equals(status.attempt)) {
            if(status.status==AttemptStatus.PROCESSING) {
                status.status = AttemptStatus.PROCESSED;
                LOG.debug("Changed status. [tx_attempt = {}] [tx_status = {}]", tx, status);
            } else if(status.status==AttemptStatus.COMMITTING) {
                _activeTx.remove(tx.getTransactionId());
                _attemptIds.remove(tx.getTransactionId());
                _collector.emit(SUCCESS_STREAM_ID, new Values(tx));
                _currTransaction = nextTransactionId(tx.getTransactionId());
                for(TransactionalState state: _states) {
                    state.setData(CURRENT_TX, _currTransaction);                    
                }
                LOG.debug("Emitted on [stream = {}], [tx_attempt = {}], [tx_status = {}], [{}]", SUCCESS_STREAM_ID, tx, status, this);
            }
            sync();
        }
    }
  • ack主要是根据当前事务状态进行不同操作,如果之前是AttemptStatus.PROCESSING状态,则更新为AttemptStatus.PROCESSED;如果之前是AttemptStatus.COMMITTING,则移除当前事务,然后往MasterBatchCoordinator.SUCCESS_STREAM_ID($success)发送tuple,更新_currTransaction为nextTransactionId;最后再调用sync触发新的TransactionAttempt

MasterBatchCoordinator.fail

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/topology/MasterBatchCoordinator.java

    @Override
    public void fail(Object msgId) {
        TransactionAttempt tx = (TransactionAttempt) msgId;
        TransactionStatus stored = _activeTx.remove(tx.getTransactionId());
        LOG.debug("Fail. [tx_attempt = {}], [tx_status = {}], [{}]", tx, stored, this);
        if(stored!=null && tx.equals(stored.attempt)) {
            _activeTx.tailMap(tx.getTransactionId()).clear();
            sync();
        }
    }
  • fail方法将当前事务从_activeTx中移除,然后清空_activeTx中txId大于这个失败txId的数据,最后再调用sync判断是否该触发新的TransactionAttempt(注意这里没有变更_currTransaction,因而sync方法触发新的TransactionAttempt的_txid还是当前这个失败的_currTransaction)

TridentSpoutCoordinator

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/spout/TridentSpoutCoordinator.java

public class TridentSpoutCoordinator implements IBasicBolt {
    public static final Logger LOG = LoggerFactory.getLogger(TridentSpoutCoordinator.class);
    private static final String META_DIR = "meta";

    ITridentSpout _spout;
    ITridentSpout.BatchCoordinator _coord;
    RotatingTransactionalState _state;
    TransactionalState _underlyingState;
    String _id;

    
    public TridentSpoutCoordinator(String id, ITridentSpout spout) {
        _spout = spout;
        _id = id;
    }
    
    @Override
    public void prepare(Map conf, TopologyContext context) {
        _coord = _spout.getCoordinator(_id, conf, context);
        _underlyingState = TransactionalState.newCoordinatorState(conf, _id);
        _state = new RotatingTransactionalState(_underlyingState, META_DIR);
    }

    @Override
    public void execute(Tuple tuple, BasicOutputCollector collector) {
        TransactionAttempt attempt = (TransactionAttempt) tuple.getValue(0);

        if(tuple.getSourceStreamId().equals(MasterBatchCoordinator.SUCCESS_STREAM_ID)) {
            _state.cleanupBefore(attempt.getTransactionId());
            _coord.success(attempt.getTransactionId());
        } else {
            long txid = attempt.getTransactionId();
            Object prevMeta = _state.getPreviousState(txid);
            Object meta = _coord.initializeTransaction(txid, prevMeta, _state.getState(txid));
            _state.overrideState(txid, meta);
            collector.emit(MasterBatchCoordinator.BATCH_STREAM_ID, new Values(attempt, meta));
        }
                
    }

    @Override
    public void cleanup() {
        _coord.close();
        _underlyingState.close();
    }

    @Override
    public void declareOutputFields(OutputFieldsDeclarer declarer) {
        declarer.declareStream(MasterBatchCoordinator.BATCH_STREAM_ID, new Fields("tx", "metadata"));
    }

    @Override
    public Map getComponentConfiguration() {
        Config ret = new Config();
        ret.setMaxTaskParallelism(1);
        return ret;
    }   
} 
 
  • TridentSpoutCoordinator的nextTuple根据streamId分别做不同的处理
  • 如果是MasterBatchCoordinator.SUCCESS_STREAM_ID($success)则表示master那边接收到了ack已经成功了,然后coordinator就清除该txId之前的数据,然后回调ITridentSpout.BatchCoordinator的success方法
  • 如果是MasterBatchCoordinator.BATCH_STREAM_ID($batch)则要启动新的TransactionAttempt,则往MasterBatchCoordinator.BATCH_STREAM_ID($batch)发送tuple,该tuple会被下游的bolt接收(在本实例就是使用TridentSpoutExecutor包装了用户spout的TridentBoltExecutor)

TridentBoltExecutor

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/topology/TridentBoltExecutor.java

public class TridentBoltExecutor implements IRichBolt {
    public static final String COORD_STREAM_PREFIX = "$coord-";
    
    public static String COORD_STREAM(String batch) {
        return COORD_STREAM_PREFIX + batch;
    }

    RotatingMap _batches;

    @Override
    public void prepare(Map conf, TopologyContext context, OutputCollector collector) {        
        _messageTimeoutMs = context.maxTopologyMessageTimeout() * 1000L;
        _lastRotate = System.currentTimeMillis();
        _batches = new RotatingMap<>(2);
        _context = context;
        _collector = collector;
        _coordCollector = new CoordinatedOutputCollector(collector);
        _coordOutputCollector = new BatchOutputCollectorImpl(new OutputCollector(_coordCollector));
                
        _coordConditions = (Map) context.getExecutorData("__coordConditions");
        if(_coordConditions==null) {
            _coordConditions = new HashMap<>();
            for(String batchGroup: _coordSpecs.keySet()) {
                CoordSpec spec = _coordSpecs.get(batchGroup);
                CoordCondition cond = new CoordCondition();
                cond.commitStream = spec.commitStream;
                cond.expectedTaskReports = 0;
                for(String comp: spec.coords.keySet()) {
                    CoordType ct = spec.coords.get(comp);
                    if(ct.equals(CoordType.single())) {
                        cond.expectedTaskReports+=1;
                    } else {
                        cond.expectedTaskReports+=context.getComponentTasks(comp).size();
                    }
                }
                cond.targetTasks = new HashSet<>();
                for(String component: Utils.get(context.getThisTargets(),
                                        COORD_STREAM(batchGroup),
                                        new HashMap()).keySet()) {
                    cond.targetTasks.addAll(context.getComponentTasks(component));
                }
                _coordConditions.put(batchGroup, cond);
            }
            context.setExecutorData("_coordConditions", _coordConditions);
        }
        _bolt.prepare(conf, context, _coordOutputCollector);
    }

    //......

    @Override
    public void cleanup() {
        _bolt.cleanup();
    }

    @Override
    public void declareOutputFields(OutputFieldsDeclarer declarer) {
        _bolt.declareOutputFields(declarer);
        for(String batchGroup: _coordSpecs.keySet()) {
            declarer.declareStream(COORD_STREAM(batchGroup), true, new Fields("id", "count"));
        }
    }

    @Override
    public Map getComponentConfiguration() {
        Map ret = _bolt.getComponentConfiguration();
        if(ret==null) ret = new HashMap<>();
        ret.put(Config.TOPOLOGY_TICK_TUPLE_FREQ_SECS, 5);
        // TODO: Need to be able to set the tick tuple time to the message timeout, ideally without parameterization
        return ret;
    }
}
  • prepare的时候,先创建了CoordinatedOutputCollector,之后用OutputCollector包装,再最后包装为BatchOutputCollectorImpl,调用ITridentBatchBolt.prepare方法,ITridentBatchBolt这里头使用的实现类为TridentSpoutExecutor
  • prepare初始化了RotatingMap _batches = new RotatingMap<>(2);
  • prepare主要做的是构建CoordCondition,这里主要是计算expectedTaskReports以及targetTasks

TridentBoltExecutor.execute

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/topology/TridentBoltExecutor.java

    @Override
    public void execute(Tuple tuple) {
        if(TupleUtils.isTick(tuple)) {
            long now = System.currentTimeMillis();
            if(now - _lastRotate > _messageTimeoutMs) {
                _batches.rotate();
                _lastRotate = now;
            }
            return;
        }
        String batchGroup = _batchGroupIds.get(tuple.getSourceGlobalStreamId());
        if(batchGroup==null) {
            // this is so we can do things like have simple DRPC that doesn't need to use batch processing
            _coordCollector.setCurrBatch(null);
            _bolt.execute(null, tuple);
            _collector.ack(tuple);
            return;
        }
        IBatchID id = (IBatchID) tuple.getValue(0);
        //get transaction id
        //if it already exists and attempt id is greater than the attempt there
        
        
        TrackedBatch tracked = (TrackedBatch) _batches.get(id.getId());
//        if(_batches.size() > 10 && _context.getThisTaskIndex() == 0) {
//            System.out.println("Received in " + _context.getThisComponentId() + " " + _context.getThisTaskIndex()
//                    + " (" + _batches.size() + ")" +
//                    "\ntuple: " + tuple +
//                    "\nwith tracked " + tracked +
//                    "\nwith id " + id + 
//                    "\nwith group " + batchGroup
//                    + "\n");
//            
//        }
        //System.out.println("Num tracked: " + _batches.size() + " " + _context.getThisComponentId() + " " + _context.getThisTaskIndex());
        
        // this code here ensures that only one attempt is ever tracked for a batch, so when
        // failures happen you don't get an explosion in memory usage in the tasks
        if(tracked!=null) {
            if(id.getAttemptId() > tracked.attemptId) {
                _batches.remove(id.getId());
                tracked = null;
            } else if(id.getAttemptId() < tracked.attemptId) {
                // no reason to try to execute a previous attempt than we've already seen
                return;
            }
        }
        
        if(tracked==null) {
            tracked = new TrackedBatch(new BatchInfo(batchGroup, id, _bolt.initBatchState(batchGroup, id)), _coordConditions.get(batchGroup), id.getAttemptId());
            _batches.put(id.getId(), tracked);
        }
        _coordCollector.setCurrBatch(tracked);
        
        //System.out.println("TRACKED: " + tracked + " " + tuple);
        
        TupleType t = getTupleType(tuple, tracked);
        if(t==TupleType.COMMIT) {
            tracked.receivedCommit = true;
            checkFinish(tracked, tuple, t);
        } else if(t==TupleType.COORD) {
            int count = tuple.getInteger(1);
            tracked.reportedTasks++;
            tracked.expectedTupleCount+=count;
            checkFinish(tracked, tuple, t);
        } else {
            tracked.receivedTuples++;
            boolean success = true;
            try {
                _bolt.execute(tracked.info, tuple);
                if(tracked.condition.expectedTaskReports==0) {
                    success = finishBatch(tracked, tuple);
                }
            } catch(FailedException e) {
                failBatch(tracked, e);
            }
            if(success) {
                _collector.ack(tuple);                   
            } else {
                _collector.fail(tuple);
            }
        }
        _coordCollector.setCurrBatch(null);
    }

    private TupleType getTupleType(Tuple tuple, TrackedBatch batch) {
        CoordCondition cond = batch.condition;
        if(cond.commitStream!=null
                && tuple.getSourceGlobalStreamId().equals(cond.commitStream)) {
            return TupleType.COMMIT;
        } else if(cond.expectedTaskReports > 0
                && tuple.getSourceStreamId().startsWith(COORD_STREAM_PREFIX)) {
            return TupleType.COORD;
        } else {
            return TupleType.REGULAR;
        }
    }

    private void failBatch(TrackedBatch tracked, FailedException e) {
        if(e!=null && e instanceof ReportedFailedException) {
            _collector.reportError(e);
        }
        tracked.failed = true;
        if(tracked.delayedAck!=null) {
            _collector.fail(tracked.delayedAck);
            tracked.delayedAck = null;
        }
    }
  • TridentBoltExecutor的execute方法首先判断是否是tickTuple,如果是判断距离_lastRotate的时间(prepare的时候初始化为当时的时间)是否超过_messageTimeoutMs,如果是则进行_batches.rotate()操作;tickTuple的发射频率为Config.TOPOLOGY_TICK_TUPLE_FREQ_SECS(topology.tick.tuple.freq.secs),在TridentBoltExecutor中它被设置为5秒;_messageTimeoutMs为context.maxTopologyMessageTimeout() * 1000L,它从整个topology的component的Config.TOPOLOGY_MESSAGE_TIMEOUT_SECS(topology.message.timeout.secs,defaults.yaml中默认为30)最大值*1000
  • _batches按TransactionAttempt的txId来存储TrackedBatch信息,如果没有则创建一个新的TrackedBatch;创建TrackedBatch时,会回调_bolt的initBatchState方法
  • 之后判断tuple的类型,这里分为TupleType.COMMIT、TupleType.COORD、TupleType.REGULAR;如果是TupleType.COMMIT类型,则设置tracked.receivedCommit为true,然后调用checkFinish方法;如果是TupleType.COORD类型,则更新reportedTasks及expectedTupleCount计数,再调用checkFinish方法;如果是TupleType.REGULAR类型(coordinator发送过来的batch信息),则更新receivedTuples计数,然后调用_bolt.execute方法(这里的_bolt为TridentSpoutExecutor),对于tracked.condition.expectedTaskReports==0的则立马调用finishBatch,将该batch从_batches中移除;如果有FailedException则直接failBatch上报error信息,之后对tuple进行ack或者fail;如果下游是each操作,一个batch中如果是部分抛出FailedException异常,则需要等到所有batch中的tuple执行完,等到TupleType.COORD触发检测checkFinish,这个时候才能fail通知到master,也就是有一些滞后性,比如这个batch中有3个tuple,第二个tuple抛出FailedException,还会继续执行第三个tuple,最后该batch的tuple都处理完了,才收到TupleType.COORD触发检测checkFinish。

TridentBoltExecutor.checkFinish

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/topology/TridentBoltExecutor.java

   private void checkFinish(TrackedBatch tracked, Tuple tuple, TupleType type) {
        if(tracked.failed) {
            failBatch(tracked);
            _collector.fail(tuple);
            return;
        }
        CoordCondition cond = tracked.condition;
        boolean delayed = tracked.delayedAck==null &&
                              (cond.commitStream!=null && type==TupleType.COMMIT
                               || cond.commitStream==null);
        if(delayed) {
            tracked.delayedAck = tuple;
        }
        boolean failed = false;
        if(tracked.receivedCommit && tracked.reportedTasks == cond.expectedTaskReports) {
            if(tracked.receivedTuples == tracked.expectedTupleCount) {
                finishBatch(tracked, tuple);                
            } else {
                //TODO: add logging that not all tuples were received
                failBatch(tracked);
                _collector.fail(tuple);
                failed = true;
            }
        }
        
        if(!delayed && !failed) {
            _collector.ack(tuple);
        }
        
    }

    private void failBatch(TrackedBatch tracked) {
        failBatch(tracked, null);
    }

    private void failBatch(TrackedBatch tracked, FailedException e) {
        if(e!=null && e instanceof ReportedFailedException) {
            _collector.reportError(e);
        }
        tracked.failed = true;
        if(tracked.delayedAck!=null) {
            _collector.fail(tracked.delayedAck);
            tracked.delayedAck = null;
        }
    }
  • TridentBoltExecutor在execute的时候,在tuple是TupleType.COMMIT以及TupleType.COORD的时候都会调用checkFinish
  • 一旦_bolt.execute(tracked.info, tuple)方法抛出FailedException,则会调用failBatch,它会标记tracked.failed为true
  • checkFinish在发现tracked.failed为true的时候,会调用_collector.fail(tuple),然后回调MasterBatchCoordinator的fail方法

TridentSpoutExecutor

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/spout/TridentSpoutExecutor.java

public class TridentSpoutExecutor implements ITridentBatchBolt {
    public static final String ID_FIELD = "$tx";
    
    public static final Logger LOG = LoggerFactory.getLogger(TridentSpoutExecutor.class);

    AddIdCollector _collector;
    ITridentSpout _spout;
    ITridentSpout.Emitter _emitter;
    String _streamName;
    String _txStateId;
    
    TreeMap _activeBatches = new TreeMap<>();

    public TridentSpoutExecutor(String txStateId, String streamName, ITridentSpout spout) {
        _txStateId = txStateId;
        _spout = spout;
        _streamName = streamName;
    }
    
    @Override
    public void prepare(Map conf, TopologyContext context, BatchOutputCollector collector) {
        _emitter = _spout.getEmitter(_txStateId, conf, context);
        _collector = new AddIdCollector(_streamName, collector);
    }

    @Override
    public void execute(BatchInfo info, Tuple input) {
        // there won't be a BatchInfo for the success stream
        TransactionAttempt attempt = (TransactionAttempt) input.getValue(0);
        if(input.getSourceStreamId().equals(MasterBatchCoordinator.COMMIT_STREAM_ID)) {
            if(attempt.equals(_activeBatches.get(attempt.getTransactionId()))) {
                ((ICommitterTridentSpout.Emitter) _emitter).commit(attempt);
                _activeBatches.remove(attempt.getTransactionId());
            } else {
                 throw new FailedException("Received commit for different transaction attempt");
            }
        } else if(input.getSourceStreamId().equals(MasterBatchCoordinator.SUCCESS_STREAM_ID)) {
            // valid to delete before what's been committed since 
            // those batches will never be accessed again
            _activeBatches.headMap(attempt.getTransactionId()).clear();
            _emitter.success(attempt);
        } else {            
            _collector.setBatch(info.batchId);
            _emitter.emitBatch(attempt, input.getValue(1), _collector);
            _activeBatches.put(attempt.getTransactionId(), attempt);
        }
    }

    @Override
    public void cleanup() {
        _emitter.close();
    }

    @Override
    public void declareOutputFields(OutputFieldsDeclarer declarer) {
        List fields = new ArrayList<>(_spout.getOutputFields().toList());
        fields.add(0, ID_FIELD);
        declarer.declareStream(_streamName, new Fields(fields));
    }

    @Override
    public Map getComponentConfiguration() {
        return _spout.getComponentConfiguration();
    }

    @Override
    public void finishBatch(BatchInfo batchInfo) {
    }

    @Override
    public Object initBatchState(String batchGroup, Object batchId) {
        return null;
    }
} 
 
  • TridentSpoutExecutor使用的BatchOutputCollector为TridentBoltExecutor在prepare方法构造的,经过几层包装,先是CoordinatedOutputCollector,然后是OutputCollector,最后是BatchOutputCollectorImpl;这里最主要的是CoordinatedOutputCollector包装,它维护每个taskId发出的tuple的数量;而在这个executor的prepare方法里头,该collector又被包装为AddIdCollector,主要是添加了batchId信息(即TransactionAttempt信息)
  • TridentSpoutExecutor的ITridentSpout就是包装了用户设置的原始spout(IBatchSpout类型)的BatchSpoutExecutor(假设原始spout是IBatchSpout类型的,因而会通过BatchSpoutExecutor包装为ITridentSpout类型),其execute方法根据不同stream类型进行不同处理,如果是master发过来的MasterBatchCoordinator.COMMIT_STREAM_ID($commit)则调用emitter的commit方法提交当前TransactionAttempt(本文的实例没有commit信息),然后将该tx从_activeBatches中移除;如果是master发过来的MasterBatchCoordinator.SUCCESS_STREAM_ID($success)则先把_activeBatches中txId小于该txId的TransactionAttempt移除,然后调用emitter的success方法,标记TransactionAttempt成功,该方法回调原始spout(IBatchSpout类型)的ack方法
  • 非MasterBatchCoordinator.COMMIT_STREAM_ID($commit)及MasterBatchCoordinator.SUCCESS_STREAM_ID($success)类型的tuple,则是启动batch的消息,这里设置batchId,然后调用emitter的emitBatch进行数据发送(这里传递的batchId就是TransactionAttempt的txId),同时将该TransactionAttempt放入_activeBatches中(这里的batch相当于TransactionAttempt)

FixedBatchSpout

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/testing/FixedBatchSpout.java

public class FixedBatchSpout implements IBatchSpout {

    Fields fields;
    List[] outputs;
    int maxBatchSize;
    HashMap>> batches = new HashMap>>();
    
    public FixedBatchSpout(Fields fields, int maxBatchSize, List... outputs) {
        this.fields = fields;
        this.outputs = outputs;
        this.maxBatchSize = maxBatchSize;
    }
    
    int index = 0;
    boolean cycle = false;
    
    public void setCycle(boolean cycle) {
        this.cycle = cycle;
    }
    
    @Override
    public void open(Map conf, TopologyContext context) {
        index = 0;
    }

    @Override
    public void emitBatch(long batchId, TridentCollector collector) {
        List> batch = this.batches.get(batchId);
        if(batch == null){
            batch = new ArrayList>();
            if(index>=outputs.length && cycle) {
                index = 0;
            }
            for(int i=0; index < outputs.length && i < maxBatchSize; index++, i++) {
                batch.add(outputs[index]);
            }
            this.batches.put(batchId, batch);
        }
        for(List list : batch){
            collector.emit(list);
        }
    }

    @Override
    public void ack(long batchId) {
        this.batches.remove(batchId);
    }

    @Override
    public void close() {
    }

    @Override
    public Map getComponentConfiguration() {
        Config conf = new Config();
        conf.setMaxTaskParallelism(1);
        return conf;
    }

    @Override
    public Fields getOutputFields() {
        return fields;
    }
    
} 
 
  • 用户使用的spout是IBatchSpout类型,这里缓存了每个batchId对应的tuple数据,实现的是transactional spout的语义

TridentTopology.newStream

storm-1.2.2/storm-core/src/jvm/org/apache/storm/trident/TridentTopology.java

     public Stream newStream(String txId, IRichSpout spout) {
        return newStream(txId, new RichSpoutBatchExecutor(spout));
    }
    
    public Stream newStream(String txId, IBatchSpout spout) {
        Node n = new SpoutNode(getUniqueStreamId(), spout.getOutputFields(), txId, spout, SpoutNode.SpoutType.BATCH);
        return addNode(n);
    }
    
    public Stream newStream(String txId, ITridentSpout spout) {
        Node n = new SpoutNode(getUniqueStreamId(), spout.getOutputFields(), txId, spout, SpoutNode.SpoutType.BATCH);
        return addNode(n);
    }
    
    public Stream newStream(String txId, IPartitionedTridentSpout spout) {
        return newStream(txId, new PartitionedTridentSpoutExecutor(spout));
    }
    
    public Stream newStream(String txId, IOpaquePartitionedTridentSpout spout) {
        return newStream(txId, new OpaquePartitionedTridentSpoutExecutor(spout));
    }

    public Stream newStream(String txId, ITridentDataSource dataSource) {
        if (dataSource instanceof IBatchSpout) {
            return newStream(txId, (IBatchSpout) dataSource);
        } else if (dataSource instanceof ITridentSpout) {
            return newStream(txId, (ITridentSpout) dataSource);
        } else if (dataSource instanceof IPartitionedTridentSpout) {
            return newStream(txId, (IPartitionedTridentSpout) dataSource);
        } else if (dataSource instanceof IOpaquePartitionedTridentSpout) {
            return newStream(txId, (IOpaquePartitionedTridentSpout) dataSource);
        } else {
            throw new UnsupportedOperationException("Unsupported stream");
        }
    }
  • 用户在TridentTopology.newStream可以直接使用IBatchSpout类似的spout,使用它的好处就是TridentTopology在build的时候会使用BatchSpoutExecutor将其包装为ITridentSpout类型(省得用户再去实现ITridentSpout的相关接口,屏蔽trident spout的相关逻辑,使得之前一直使用普通topology的用户可以快速上手trident topology)
  • BatchSpoutExecutor实现了ITridentSpout接口,将IBatchSpout适配为ITridentSpout,使用的coordinator是EmptyCoordinator,使用的emitter是BatchSpoutEmitter
  • 如果用户在TridentTopology.newStream使用的spout是IPartitionedTridentSpout类型,则TridentTopology在newStream方法内部会使用PartitionedTridentSpoutExecutor将其包装为ITridentSpout类型,对于IOpaquePartitionedTridentSpout则使用OpaquePartitionedTridentSpoutExecutor将其包装为ITridentSpout类型

小结

  • TridentTopology在newStream或者build方法里头会将ITridentDataSource中不是ITridentSpout类型的IBatchSpout(在build方法)、IPartitionedTridentSpout(在newStream方法)、IOpaquePartitionedTridentSpout(在newStream方法)适配为ITridentSpout类型;分别使用BatchSpoutExecutor、PartitionedTridentSpoutExecutor、OpaquePartitionedTridentSpoutExecutor进行适配(TridentTopologyBuilder在buildTopology的时候,对于ITridentSpout类型的spout先用TridentSpoutExecutor包装,再用TridentBoltExecutor包装,最后转换为bolt,而整个TridentTopology真正的spout就是MasterBatchCoordinator;这里可以看到一个IBatchSpout的spout先经过BatchSpoutExecutor包装为ITridentSpout类型,之后再经过TridentSpoutExecutor及TridentBoltExecutor包装为bolt)
  • IBatchSpout的ack是针对batch维度的,也就是TransactionAttempt维度,注意这里没有fail方法,如果emitBatch方法抛出了FailedException异常,则TridentBoltExecutor会调用failBatch方法(一个batch的tuples会等所有tuple执行完再触发checkFinish),进行reportError以及标记TrackedBatch的failed为true,之后TridentBoltExecutor在checkFinish的时候,一旦发现tracked.failed为true的时候,会调用_collector.fail(tuple),然后回调MasterBatchCoordinator的fail方法
  • MasterBatchCoordinator的fail方法会将当前TransactionAttempt从_activeTx移除,然后一并移除txId大于失败的txId的数据,最后调用sync方法继续TransactionAttempt(注意这里没有更改_currTransaction值,因而会继续从失败的txId开始重试,只有在ack方法里头会更改_currTransaction为nextTransactionId)
  • TridentBoltExecutor的execute方法会根据tickTuple来检测距离上次rotate是否超过_messageTimeoutMs(取component中Config.TOPOLOGY_MESSAGE_TIMEOUT_SECS最大值*1000,这里*1000是将秒转换为毫秒),超过的话进行rotate操作,_batches的最后一个bucket将会被移除掉;这里的tickTuple的频率为5秒,Config.TOPOLOGY_MESSAGE_TIMEOUT_SECS按30秒算的话,_messageTimeoutMs为30*1000,相当于每5秒检测一下距离上次rotate时间是否超过30秒,如果超过则进行rotate,丢弃最后一个bucket的数据(TrackedBatch),这里相当于重置超时的TrackedBatch信息
  • 关于MasterBatchCoordinator的fail的情况,有几种情况,一种是下游componnent主动抛出FailException,这个时候会触发master的fail,再次重试TransactionAttempt;一种是下游component处理tuple时间超过Config.TOPOLOGY_MESSAGE_TIMEOUT_SECS(topology.message.timeout.secs,defaults.yaml中默认为30),这个时候ack会触发master的fail,导致该TransactionAttempt失败继续重试,目前没有对attempt的次数做限制,实际生产过程中要注意,因为只要该batchId的一个tuple失败,整个batchId的tuples都会重发,这个时候下游如果没有做好处理,可能会出现一个batchId中前面部分tuple成功,后面部分失败,导致成功的tuple不断重复处理(要避免失败的batch中tuples部分处理成功部分处理失败这个问题就需要配合使用Trident的State)。

doc

  • Trident Spouts
  • Trident State
  • 聊聊storm TridentTopology的构建

你可能感兴趣的:(storm)