[case44]聊聊storm trident的operations

序

本文主要研究一下storm trident的operations

function filter projection

Function

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/operation/Function.java

public interface Function extends EachOperation {
    /**
     * Performs the function logic on an individual tuple and emits 0 or more tuples.
     *
     * @param tuple The incoming tuple
     * @param collector A collector instance that can be used to emit tuples
     */
    void execute(TridentTuple tuple, TridentCollector collector);
}

• Function定义了execute方法，它发射的字段会追加到input tuple中

Filter

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/operation/Filter.java

public interface Filter extends EachOperation {

    /**
     * Determines if a tuple should be filtered out of a stream
     *
     * @param tuple the tuple being evaluated
     * @return `false` to drop the tuple, `true` to keep the tuple
     */
    boolean isKeep(TridentTuple tuple);
}

• Filter提供一个isKeep方法，用来决定该tuple是否输出

projection

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * Filters out fields from a stream, resulting in a Stream containing only the fields specified by `keepFields`.
     *
     * For example, if you had a Stream `mystream` containing the fields `["a", "b", "c","d"]`, calling"
     *
     * ```java
     * mystream.project(new Fields("b", "d"))
     * ```
     *
     * would produce a stream containing only the fields `["b", "d"]`.
     *
     *
     * @param keepFields The fields in the Stream to keep
     * @return
     */
    public Stream project(Fields keepFields) {
        projectionValidation(keepFields);
        return _topology.addSourcedNode(this, new ProcessorNode(_topology.getUniqueStreamId(), _name, keepFields, new Fields(), new ProjectedProcessor(keepFields)));
    }

• 这里使用了ProjectedProcessor来进行projection操作

repartitioning operations

partition

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * ## Repartitioning Operation
     *
     * @param partitioner
     * @return
     */
    public Stream partition(CustomStreamGrouping partitioner) {
        return partition(Grouping.custom_serialized(Utils.javaSerialize(partitioner)));
    }

• 这里使用了CustomStreamGrouping

partitionBy

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * ## Repartitioning Operation
     *
     * @param fields
     * @return
     */
    public Stream partitionBy(Fields fields) {
        projectionValidation(fields);
        return partition(Grouping.fields(fields.toList()));
    }

• 这里使用Grouping.fields

identityPartition

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * ## Repartitioning Operation
     *
     * @return
     */
    public Stream identityPartition() {
        return partition(new IdentityGrouping());
    }

• 这里使用IdentityGrouping

shuffle

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * ## Repartitioning Operation
     *
     * Use random round robin algorithm to evenly redistribute tuples across all target partitions
     *
     * @return
     */
    public Stream shuffle() {
        return partition(Grouping.shuffle(new NullStruct()));
    }

• 这里使用Grouping.shuffle

localOrShuffle

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * ## Repartitioning Operation
     *
     * Use random round robin algorithm to evenly redistribute tuples across all target partitions, with a preference
     * for local tasks.
     *
     * @return
     */
    public Stream localOrShuffle() {
        return partition(Grouping.local_or_shuffle(new NullStruct()));
    }

• 这里使用Grouping.local_or_shuffle

global

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * ## Repartitioning Operation
     *
     * All tuples are sent to the same partition. The same partition is chosen for all batches in the stream.
     * @return
     */
    public Stream global() {
        // use this instead of storm's built in one so that we can specify a singleemitbatchtopartition
        // without knowledge of storm's internals
        return partition(new GlobalGrouping());
    }

• 这里使用GlobalGrouping

batchGlobal

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * ## Repartitioning Operation
     *
     *  All tuples in the batch are sent to the same partition. Different batches in the stream may go to different
     *  partitions.
     *
     * @return
     */
    public Stream batchGlobal() {
        // the first field is the batch id
        return partition(new IndexHashGrouping(0));
    }

• 这里使用IndexHashGrouping，是对整个batch维度的repartition

broadcast

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * ## Repartitioning Operation
     *
     * Every tuple is replicated to all target partitions. This can useful during DRPC – for example, if you need to do
     * a stateQuery on every partition of data.
     *
     * @return
     */
    public Stream broadcast() {
        return partition(Grouping.all(new NullStruct()));
    }

• 这里使用Grouping.all

groupBy

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

/**
     * ## Grouping Operation
     *
     * @param fields
     * @return
     */
    public GroupedStream groupBy(Fields fields) {
        projectionValidation(fields);
        return new GroupedStream(this, fields);
    }

• 这里返回的是GroupedStream

aggregators

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/Stream.java

//partition aggregate
    public Stream partitionAggregate(Aggregator agg, Fields functionFields) {
        return partitionAggregate(null, agg, functionFields);
    }

    public Stream partitionAggregate(CombinerAggregator agg, Fields functionFields) {
        return partitionAggregate(null, agg, functionFields);
    }

    public Stream partitionAggregate(Fields inputFields, CombinerAggregator agg, Fields functionFields) {
        projectionValidation(inputFields);
        return chainedAgg()
               .partitionAggregate(inputFields, agg, functionFields)
               .chainEnd();
    }

    public Stream partitionAggregate(ReducerAggregator agg, Fields functionFields) {
        return partitionAggregate(null, agg, functionFields);
    }

    public Stream partitionAggregate(Fields inputFields, ReducerAggregator agg, Fields functionFields) {
        projectionValidation(inputFields);
        return chainedAgg()
               .partitionAggregate(inputFields, agg, functionFields)
               .chainEnd();
    }

    //aggregate
    public Stream aggregate(Fields inputFields, Aggregator agg, Fields functionFields) {
        projectionValidation(inputFields);
        return chainedAgg()
               .aggregate(inputFields, agg, functionFields)
               .chainEnd();
    }

    public Stream aggregate(Fields inputFields, CombinerAggregator agg, Fields functionFields) {
        projectionValidation(inputFields);
        return chainedAgg()
               .aggregate(inputFields, agg, functionFields)
               .chainEnd();
    }

    public Stream aggregate(Fields inputFields, ReducerAggregator agg, Fields functionFields) {
        projectionValidation(inputFields);
        return chainedAgg()
                .aggregate(inputFields, agg, functionFields)
                .chainEnd();
    }

    //persistent aggregate
    public TridentState persistentAggregate(StateFactory stateFactory, CombinerAggregator agg, Fields functionFields) {
        return persistentAggregate(new StateSpec(stateFactory), agg, functionFields);
    }

    public TridentState persistentAggregate(StateSpec spec, CombinerAggregator agg, Fields functionFields) {
        return persistentAggregate(spec, null, agg, functionFields);
    }

    public TridentState persistentAggregate(StateFactory stateFactory, Fields inputFields, CombinerAggregator agg, Fields functionFields) {
        return persistentAggregate(new StateSpec(stateFactory), inputFields, agg, functionFields);
    }

    public TridentState persistentAggregate(StateSpec spec, Fields inputFields, CombinerAggregator agg, Fields functionFields) {
        projectionValidation(inputFields);
        // replaces normal aggregation here with a global grouping because it needs to be consistent across batches 
        return new ChainedAggregatorDeclarer(this, new GlobalAggScheme())
                .aggregate(inputFields, agg, functionFields)
                .chainEnd()
               .partitionPersist(spec, functionFields, new CombinerAggStateUpdater(agg), functionFields);
    }

    public TridentState persistentAggregate(StateFactory stateFactory, ReducerAggregator agg, Fields functionFields) {
        return persistentAggregate(new StateSpec(stateFactory), agg, functionFields);
    }

    public TridentState persistentAggregate(StateSpec spec, ReducerAggregator agg, Fields functionFields) {
        return persistentAggregate(spec, null, agg, functionFields);
    }

    public TridentState persistentAggregate(StateFactory stateFactory, Fields inputFields, ReducerAggregator agg, Fields functionFields) {
        return persistentAggregate(new StateSpec(stateFactory), inputFields, agg, functionFields);
    }

    public TridentState persistentAggregate(StateSpec spec, Fields inputFields, ReducerAggregator agg, Fields functionFields) {
        projectionValidation(inputFields);
        return global().partitionPersist(spec, inputFields, new ReducerAggStateUpdater(agg), functionFields);
    }

• trident的aggregators主要分为三类，分别是partitionAggregate、aggregate、persistentAggregate；aggregator操作会改变输出
• partitionAggregate其作用的粒度为每个partition，而非整个batch
• aggregrate操作作用的粒度为batch，对每个batch，它先使用global操作将该batch的tuple从所有partition合并到一个partition，最后再对batch进行aggregation操作；这里提供了三类参数，分别是Aggregator、CombinerAggregator、ReducerAggregator；调用stream.aggregrate方法时，相当于一次global aggregation，此时使用Aggregator或ReducerAggregator时，stream会先将tuple划分到一个partition，然后再进行aggregate操作；而使用CombinerAggregator时，trident会进行优化，先对每个partition进行局部的aggregate操作，然后再划分到一个partition，最后再进行aggregate操作，因而相对Aggregator或ReducerAggregator可以节省网络传输耗时
• persistentAggregate操作会对stream上所有batch的tuple进行aggretation，然后将结果存储在state中

Aggregator

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/operation/Aggregator.java

public interface Aggregator<T> extends Operation {
    T init(Object batchId, TridentCollector collector);
    void aggregate(T val, TridentTuple tuple, TridentCollector collector);
    void complete(T val, TridentCollector collector);
}

• Aggregator首先会调用init进行初始化，然后通过参数传递给aggregate以及complete方法
• 对于batch partition中的每个tuple执行一次aggregate；当batch partition中的tuple执行完aggregate之后执行complete方法
• 假设自定义Aggregator为累加操作，那么对于[4]、[7]、[8]这批tuple，init为0，对于[4]，val=0，0+4=4；对于[7]，val=4，4+7=11；对于[8]，val=11，11+8=19；然后batch结束，val=19，此时执行complete，可以使用collector发射数据

CombinerAggregator

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/operation/CombinerAggregator.java

public interface CombinerAggregator<T> extends Serializable {
    T init(TridentTuple tuple);
    T combine(T val1, T val2);
    T zero();
}

• CombinerAggregator每收到一个tuple，就调用init获取当前tuple的值，调用combine操作使用前一个combine的结果(没有的话取zero的值)与init取得的值进行新的combine操作，如果该partition中没有tuple，则返回zero方法的值
• 假设combine为累加操作，zero返回0，那么对于[4]、[7]、[8]这批tuple，init值分别是4、7、8，对于[4]，没有前一个combine结果，于是val1=0，val2=4，combine结果为4；对于[7]，val1=4，val2=7，combine结果为11；对于[8]，val1为11，val2为8，combine结果为19
• CombinerAggregator操作的网络开销相对较低，因此性能比其他两类aggratator好

ReducerAggregator

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/operation/ReducerAggregator.java

public interface ReducerAggregator<T> extends Serializable {
    T init();
    T reduce(T curr, TridentTuple tuple);
}

• ReducerAggregator在对一批tuple进行计算时，先调用一次init获取初始值，然后再执行reduce操作，curr值为前一次reduce操作的值，没有的话，就是init值
• 假设reduce为累加操作，init返回0，那么对于[4]、[7]、[8]这批tuple，对于[4]，init为0，然后curr=0，先是0+4=4；对于[7]，curr为4，就是4+7=11；对于[8]，curr为11，最后就是11+8=19

topology stream operations

join

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/TridentTopology.java

public Stream join(Stream s1, Fields joinFields1, Stream s2, Fields joinFields2, Fields outFields) {
        return join(Arrays.asList(s1, s2), Arrays.asList(joinFields1, joinFields2), outFields);        
    }
    
    public Stream join(List<Stream> streams, List<Fields> joinFields, Fields outFields) {
        return join(streams, joinFields, outFields, JoinType.INNER);        
    }

    public Stream join(Stream s1, Fields joinFields1, Stream s2, Fields joinFields2, Fields outFields, JoinType type) {
        return join(Arrays.asList(s1, s2), Arrays.asList(joinFields1, joinFields2), outFields, type);        
    }
    
    public Stream join(List<Stream> streams, List<Fields> joinFields, Fields outFields, JoinType type) {
        return join(streams, joinFields, outFields, repeat(streams.size(), type));
    }

    public Stream join(Stream s1, Fields joinFields1, Stream s2, Fields joinFields2, Fields outFields, List<JoinType> mixed) {
        return join(Arrays.asList(s1, s2), Arrays.asList(joinFields1, joinFields2), outFields, mixed);
        
    }
    
    public Stream join(List<Stream> streams, List<Fields> joinFields, Fields outFields, List<JoinType> mixed) {
        return join(streams, joinFields, outFields, mixed, JoinOutFieldsMode.COMPACT);
    }

    public Stream join(Stream s1, Fields joinFields1, Stream s2, Fields joinFields2, Fields outFields, JoinOutFieldsMode mode) {
        return join(Arrays.asList(s1, s2), Arrays.asList(joinFields1, joinFields2), outFields, mode);
    }

    public Stream join(List<Stream> streams, List<Fields> joinFields, Fields outFields, JoinOutFieldsMode mode) {
        return join(streams, joinFields, outFields, JoinType.INNER, mode);
    }

    public Stream join(Stream s1, Fields joinFields1, Stream s2, Fields joinFields2, Fields outFields, JoinType type, JoinOutFieldsMode mode) {
        return join(Arrays.asList(s1, s2), Arrays.asList(joinFields1, joinFields2), outFields, type, mode);
    }

    public Stream join(List<Stream> streams, List<Fields> joinFields, Fields outFields, JoinType type, JoinOutFieldsMode mode) {
        return join(streams, joinFields, outFields, repeat(streams.size(), type), mode);
    }

    public Stream join(Stream s1, Fields joinFields1, Stream s2, Fields joinFields2, Fields outFields, List<JoinType> mixed, JoinOutFieldsMode mode) {
        return join(Arrays.asList(s1, s2), Arrays.asList(joinFields1, joinFields2), outFields, mixed, mode);

    }

    public Stream join(List<Stream> streams, List<Fields> joinFields, Fields outFields, List<JoinType> mixed, JoinOutFieldsMode mode) {
        switch (mode) {
            case COMPACT:
                return multiReduce(strippedInputFields(streams, joinFields),
                        groupedStreams(streams, joinFields),
                        new JoinerMultiReducer(mixed, joinFields.get(0).size(), strippedInputFields(streams, joinFields)),
                        outFields);

            case PRESERVE:
                return multiReduce(strippedInputFields(streams, joinFields),
                        groupedStreams(streams, joinFields),
                        new PreservingFieldsOrderJoinerMultiReducer(mixed, joinFields.get(0).size(),
                                getAllOutputFields(streams), joinFields, strippedInputFields(streams, joinFields)),
                        outFields);

            default:
                throw new IllegalArgumentException("Unsupported out-fields mode: " + mode);
        }
    }

• 可以看到join最后调用了multiReduce，对于COMPACT类型使用的GroupedMultiReducer是JoinerMultiReducer，对于PRESERVE类型使用的GroupedMultiReducer是PreservingFieldsOrderJoinerMultiReducer

merge

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/TridentTopology.java

public Stream merge(Fields outputFields, Stream... streams) {
        return merge(outputFields, Arrays.asList(streams));
    }
    
    public Stream merge(Stream... streams) {
        return merge(Arrays.asList(streams));
    }
    
    public Stream merge(List<Stream> streams) {
        return merge(streams.get(0).getOutputFields(), streams);
    } 

    public Stream merge(Fields outputFields, List<Stream> streams) {
        return multiReduce(streams, new IdentityMultiReducer(), outputFields);
    }

• 可以看到merge最后是调用了multiReduce，使用的MultiReducer是IdentityMultiReducer

multiReduce

storm-core-1.2.2-sources.jar!/org/apache/storm/trident/TridentTopology.java

public Stream multiReduce(Stream s1, Stream s2, MultiReducer function, Fields outputFields) {
        return multiReduce(Arrays.asList(s1, s2), function, outputFields);        
    }

    public Stream multiReduce(Fields inputFields1, Stream s1, Fields inputFields2, Stream s2, MultiReducer function, Fields outputFields) {
        return multiReduce(Arrays.asList(inputFields1, inputFields2), Arrays.asList(s1, s2), function, outputFields);        
    }    
    
    public Stream multiReduce(GroupedStream s1, GroupedStream s2, GroupedMultiReducer function, Fields outputFields) {
        return multiReduce(Arrays.asList(s1, s2), function, outputFields);        
    }
    
    public Stream multiReduce(Fields inputFields1, GroupedStream s1, Fields inputFields2, GroupedStream s2, GroupedMultiReducer function, Fields outputFields) {
        return multiReduce(Arrays.asList(inputFields1, inputFields2), Arrays.asList(s1, s2), function, outputFields);        
    } 
    
    public Stream multiReduce(List<Stream> streams, MultiReducer function, Fields outputFields) {
        return multiReduce(getAllOutputFields(streams), streams, function, outputFields);
    }
        
    public Stream multiReduce(List<GroupedStream> streams, GroupedMultiReducer function, Fields outputFields) {
        return multiReduce(getAllOutputFields(streams), streams, function, outputFields);        
    }    
    
    public Stream multiReduce(List<Fields> inputFields, List<GroupedStream> groupedStreams, GroupedMultiReducer function, Fields outputFields) {
        List<Fields> fullInputFields = new ArrayList<>();
        List<Stream> streams = new ArrayList<>();
        List<Fields> fullGroupFields = new ArrayList<>();
        for(int i=0; i<groupedStreams.size(); i++) {
            GroupedStream gs = groupedStreams.get(i);
            Fields groupFields = gs.getGroupFields();
            fullGroupFields.add(groupFields);
            streams.add(gs.toStream().partitionBy(groupFields));
            fullInputFields.add(TridentUtils.fieldsUnion(groupFields, inputFields.get(i)));
            
        }
        return multiReduce(fullInputFields, streams, new GroupedMultiReducerExecutor(function, fullGroupFields, inputFields), outputFields);
    }

    public Stream multiReduce(List<Fields> inputFields, List<Stream> streams, MultiReducer function, Fields outputFields) {
        List<String> names = new ArrayList<>();
        for(Stream s: streams) {
            if(s._name!=null) {
                names.add(s._name);
            }
        }
        Node n = new ProcessorNode(getUniqueStreamId(), Utils.join(names, "-"), outputFields, outputFields, new MultiReducerProcessor(inputFields, function));
        return addSourcedNode(streams, n);
    }

• multiReduce方法有个MultiReducer参数，join与merge虽然都调用了multiReduce，但是他们传的MultiReducer值不一样

小结

• trident的操作主要有几类，一类是基本的function、filter、projection操作；一类是repartitioning操作，主要是一些grouping；一类是aggregate操作，包括aggregate、partitionAggregate、persistentAggregate；一类是在topology对stream的join、merge操作
• function的话，若有emit字段会追加到原始的tuple上；filter用于过滤tuple；projection用于提取字段
• repartitioning操作有Grouping.local_or_shuffle、Grouping.shuffle、Grouping.all、GlobalGrouping、CustomStreamGrouping、IdentityGrouping、IndexHashGrouping等；partition操作可以理解为将输入的tuple分配到task上，也可以理解为是对stream进行grouping
• aggregate操作的话，普通的aggregate操作有3类接口，分别是Aggregator、CombinerAggregator、ReducerAggregator，其中Aggregator是最为通用的，它继承了Operation接口，而且在方法参数里头可以使用到collector，这是CombinerAggregator与ReducerAggregator所没有的；而CombinerAggregator与Aggregator及ReducerAggregator不同的是，调用stream.aggregrate方法时，trident会优先在partition进行局部聚合，然后再归一到一个partition做最后聚合，相对来说比较节省网络传输耗时，但是如果将CombinerAggregator与非CombinerAggregator的进行chaining的话，就享受不到这个优化；partitionAggregate主要是在partition维度上进行操作；而persistentAggregate则是在整个stream的维度上对所有batch的tuple进行操作，结果持久化在state上
• 对于stream的join及merge操作，其最后都是依赖multiReduce来实现，只是传递的MultiReducer值不一样；join的话join的话需要字段来进行匹配(字段名可以不一样)，可以选择JoinType，是INNER还是OUTER，不过join是对于spout的small batch来进行join的；merge的话，就是纯粹的几个stream进行tuple的归总。

doc

• Trident API Overview
• Trident API 综述
• JStorm Trident 入门精华一页纸
• Everything You Need to Know about Apache Storm
• batch and partition - differences