`
dalan_123
  • 浏览: 87037 次
  • 性别: Icon_minigender_1
  • 来自: 郑州
社区版块
存档分类
最新评论

jstorm源码之TridentTopology

 
阅读更多
在jstorm中对应TridentTopology的源码如下,TridentTopology的主要作用是用来创建trident的
一、一般使用的方式如下
// 1、创建一个topology对象
TridentTopology topology = new TridentTopology();
// 2、设置相关参数: each分片、group by分组、聚合aggregate
//  返回TridentState
         TridentState wordCounts = topology.newStream("spout1", spout).parallelismHint(16)
                .each(new Fields("sentence"), new Split(), new Fields("word")).groupBy(new Fields("word"))
                .persistentAggregate(new MemoryMapState.Factory(), new Count(), new Fields("count"))
                .parallelismHint(16);
//  3、传输
        topology.newDRPCStream("words", drpc).each(new Fields("args"), new Split(), new Fields("word"))
                .groupBy(new Fields("word"))
                .stateQuery(wordCounts, new Fields("word"), new MapGet(), new Fields("count"))
                .each(new Fields("count"), new FilterNull())
                .aggregate(new Fields("count"), new Sum(), new Fields("sum"));
// 4、build topology
        return topology.build();

二、具体源码:

package storm.trident;

import java.io.Serializable;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Collection;
import java.util.HashMap;
import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.Set;
import java.util.TreeMap;

import org.jgrapht.DirectedGraph;
import org.jgrapht.UndirectedGraph;
import org.jgrapht.alg.ConnectivityInspector;
import org.jgrapht.graph.DefaultDirectedGraph;
import org.jgrapht.graph.Pseudograph;

import com.alibaba.jstorm.client.ConfigExtension;

import backtype.storm.Config;
import backtype.storm.ILocalDRPC;
import backtype.storm.drpc.DRPCSpout;
import backtype.storm.generated.GlobalStreamId;
import backtype.storm.generated.Grouping;
import backtype.storm.generated.StormTopology;
import backtype.storm.grouping.CustomStreamGrouping;
import backtype.storm.topology.BoltDeclarer;
import backtype.storm.topology.IRichSpout;
import backtype.storm.tuple.Fields;
import backtype.storm.utils.Utils;
import storm.trident.drpc.ReturnResultsReducer;
import storm.trident.fluent.GroupedStream;
import storm.trident.fluent.IAggregatableStream;
import storm.trident.fluent.UniqueIdGen;
import storm.trident.graph.GraphGrouper;
import storm.trident.graph.Group;
import storm.trident.operation.GroupedMultiReducer;
import storm.trident.operation.MultiReducer;
import storm.trident.operation.impl.FilterExecutor;
import storm.trident.operation.impl.GroupedMultiReducerExecutor;
import storm.trident.operation.impl.IdentityMultiReducer;
import storm.trident.operation.impl.JoinerMultiReducer;
import storm.trident.operation.impl.TrueFilter;
import storm.trident.partition.IdentityGrouping;
import storm.trident.planner.Node;
import storm.trident.planner.NodeStateInfo;
import storm.trident.planner.PartitionNode;
import storm.trident.planner.ProcessorNode;
import storm.trident.planner.SpoutNode;
import storm.trident.planner.SubtopologyBolt;
import storm.trident.planner.processor.EachProcessor;
import storm.trident.planner.processor.MultiReducerProcessor;
import storm.trident.spout.BatchSpoutExecutor;
import storm.trident.spout.IBatchSpout;
import storm.trident.spout.IOpaquePartitionedTridentSpout;
import storm.trident.spout.IPartitionedTridentSpout;
import storm.trident.spout.ITridentSpout;
import storm.trident.spout.OpaquePartitionedTridentSpoutExecutor;
import storm.trident.spout.PartitionedTridentSpoutExecutor;
import storm.trident.spout.RichSpoutBatchExecutor;
import storm.trident.state.StateFactory;
import storm.trident.state.StateSpec;
import storm.trident.topology.TridentTopologyBuilder;
import storm.trident.util.ErrorEdgeFactory;
import storm.trident.util.IndexedEdge;
import storm.trident.util.TridentUtils;


// graph with 3 kinds of nodes:
// operation, partition, or spout
// all operations have finishBatch and can optionally be committers
public class TridentTopology {

    //TODO: add a method for drpc stream, needs to know how to automatically do returnresults, etc
    // is it too expensive to do a batch per drpc request?
    // 有序无环图 主要记录该topology中对应的spout、bolt以及spout和bolt之间的stream(tuples)关联关系
    DefaultDirectedGraph<Node, IndexedEdge> _graph;
    // 记录topology包含的node(spout/bolt)
    Map<String, List<Node>> _colocate = new HashMap();
    // 唯一全局Id
    UniqueIdGen _gen;
  
    // 构建一个有序无环图 初始化 DirectedGraph 以及 id记录仪(用于topology无component(spout和bolt))
    public TridentTopology() {
        _graph = new DefaultDirectedGraph(new ErrorEdgeFactory());
        _gen = new UniqueIdGen();
    }
  
    //  构建一个有序无环图 初始化 DirectedGraph 以及 id记录仪(用于topology有component(spout和bolt))
    private TridentTopology(DefaultDirectedGraph<Node, IndexedEdge> graph, Map<String, List<Node>> colocate, UniqueIdGen gen) {
        _graph = graph;
        _colocate = colocate;
        _gen = gen;
    }


    // automatically turn it into a batch spout, should take parameters as to how much to batch
//    public Stream newStream(IRichSpout spout) {
//        Node n = new SpoutNode(getUniqueStreamId(), TridentUtils.getSingleOutputStreamFields(spout), null, spout, SpoutNode.SpoutType.BATCH);
//        return addNode(n);
//    }
    
     // 根据每一个stream  在jstorm中都有一个对应的txId(关于这个txId都是UniqueIdGen来生成的从0开始递增);注意:spout有两种一种是IRichSpout另外一种是IBatchSpout
     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);
    }
    //  根据指定的stream的streamId 、spout输出fileds、事务txId、spout、Spout的Node Type构建SpoutNode
    //  并将该节点Node添加有序无环图_graph 中 同时并判断对应的_colocate是否存在该节点的信息(根据state info的id作为node的唯一标示)
    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));
    }
   
   // 创建Drpc的stream
    public Stream newDRPCStream(String function) {
        return newDRPCStream(new DRPCSpout(function));
    }

    public Stream newDRPCStream(String function, ILocalDRPC server) {
        DRPCSpout spout;
        if(server==null) {
            spout = new DRPCSpout(function);
        } else {
            spout = new DRPCSpout(function, server);
        }
        return newDRPCStream(spout);
    }
   
    // 根据指定的unique stream id 、单个single output fields(jstorm在trident topology中只支持single output fields集同时也不支持direct的stream)、不需要指定事务txId
    //  Spout Node Type(spout的节点类型)
    // 将SpoutNode添加到Ttopology的有序无环图_graph中
    // 同时调用stream的project将该节点添加到topology的source node中(换句话说也是在graph指定该节点node对应的source和IndexedEdge 向该节点发送数据的所有节点)
    private Stream newDRPCStream(DRPCSpout spout) {
        // TODO: consider adding a shuffle grouping after the spout to avoid so much routing of the args/return-info all over the place
        // (at least until its possible to just pack bolt logic into the spout itself)

        Node n = new SpoutNode(getUniqueStreamId(), TridentUtils.getSingleOutputStreamFields(spout), null, spout, SpoutNode.SpoutType.DRPC);
        Stream nextStream = addNode(n);
        // later on, this will be joined back with return-info and all the results
        return nextStream.project(new Fields("args"));
    }
   
    public TridentState newStaticState(StateFactory factory) {
        return newStaticState(new StateSpec(factory));
    }

    // 根据StateSpec参数----》唯一状态Id,通过Node构造一个new node 同时指定StateInfo
    // 并该节点添加到_colocate中
    // 并返回TridentState
    public TridentState newStaticState(StateSpec spec) {
        String stateId = getUniqueStateId();
        Node n = new Node(getUniqueStreamId(), null, new Fields());
        n.stateInfo = new NodeStateInfo(stateId, spec);
        registerNode(n);
        return new TridentState(this, n);
    }

    // # jstrom 里面提供的reduce和join方法只是用来完成 上游节点 输出Fields 采用指定function 完成合并(reduce)和扩展(join)
    // 再通过新的OutputFields输出
   // 将stream 1和stream2通过reducer处理将output fields 存放 outputFields 不再存在原有的stream1和stream2的outputfields
    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<Stream> streams, MultiReducer function, Fields outputFields) {
        List<String> names = new ArrayList<String>();
        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);
    }

    /**
     * process Group stream
     * @param inputFields
     * @param groupedStreams
     * @param function
     * @param outputFields
     * @return
     */
    public Stream multiReduce(List<Fields> inputFields, List<GroupedStream> groupedStreams, GroupedMultiReducer function, Fields outputFields) {
        List<Fields> fullInputFields = new ArrayList<Fields>();
        List<Stream> streams = new ArrayList<Stream>();
        List<Fields> fullGroupFields = new ArrayList<Fields>();
        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 merge(Fields outputFields, Stream... streams) {
        return merge(outputFields, Arrays.asList(streams));
    }

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

    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 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 multiReduce(strippedInputFields(streams, joinFields),
              groupedStreams(streams, joinFields),
              new JoinerMultiReducer(mixed, joinFields.get(0).size(), strippedInputFields(streams, joinFields)),
              outFields);
    }   

  // 注:由于transaction不支持jstorm的batch mode 需要关闭TASK_BATCH_TUPLE = false
  //
    public StormTopology build() {
        // Transaction is not compatible with jstorm batch mode(task.batch.tuple)
        // so we close batch mode via system property
        System.setProperty(ConfigExtension.TASK_BATCH_TUPLE, "false");

        DefaultDirectedGraph<Node, IndexedEdge> graph = (DefaultDirectedGraph) _graph.clone();

        //  在此处主要完成如下几点内容
        //  1、通过使用第三方的图形库  判断整个topology的graph中component之间的connection ,并将component之间有关联的放到一个set中 那么整个topology形成的graph的集合List<Set<Node>>
        //  2、 针对生成后的graph进行spout检查  spout不同时支持batch和drpc
        //  3、获取每个分支中的DrpcNode  并将该分支中数据源投递  通过drpc接着后续操作
        completeDRPC(graph, _colocate, _gen);

        List<SpoutNode> spoutNodes = new ArrayList<SpoutNode>();

        // can be regular nodes (static state) or processor nodes
        Set<Node> boltNodes = new HashSet<Node>();
        // 循环遍历graph的有序无环图中的各个node(bolt和spout) 判断 添加到不同的nodes中
        for(Node n: graph.vertexSet()) {
            if(n instanceof SpoutNode) {
                spoutNodes.add((SpoutNode) n);
            } else if(!(n instanceof PartitionNode)) {
                boltNodes.add(n);
            }
        }

        // 分组 group
        Set<Group> initialGroups = new HashSet<Group>();
        for(List<Node> colocate: _colocate.values()) {
            Group g = new Group(graph, colocate);
            boltNodes.removeAll(colocate);
            initialGroups.add(g);
        }
        for(Node n: boltNodes) {
            initialGroups.add(new Group(graph, n));
        }

        //
        GraphGrouper grouper = new GraphGrouper(graph, initialGroups);
        grouper.mergeFully();
        Collection<Group> mergedGroups = grouper.getAllGroups();



        // add identity partitions between groups
        // 如下代码目前没有完全搞透
        // 大体上要实现在sourceNode 和 targetNode 在不同group中 添加一个ThriftNode(PartitionNode)
        // 并将该节点 填充到有序无环图中 (添加顶点(Vertex)、Edge)
        for(IndexedEdge<Node> e: new HashSet<IndexedEdge>(graph.edgeSet())) {
            if(!(e.source instanceof PartitionNode) && !(e.target instanceof PartitionNode)) {               
                Group g1 = grouper.nodeGroup(e.source);
                Group g2 = grouper.nodeGroup(e.target);
                // g1 being null means the source is a spout node
                if(g1==null && !(e.source instanceof SpoutNode))
                    throw new RuntimeException("Planner exception: Null source group must indicate a spout node at this phase of planning");
                if(g1==null || !g1.equals(g2)) {
                    graph.removeEdge(e);
                    PartitionNode pNode = makeIdentityPartition(e.source);
                    graph.addVertex(pNode);
                    graph.addEdge(e.source, pNode, new IndexedEdge(e.source, pNode, 0));
                    graph.addEdge(pNode, e.target, new IndexedEdge(pNode, e.target, e.index));                   
                }
            }
        }
        // if one group subscribes to the same stream with same partitioning multiple times,
        // merge those together (otherwise can end up with many output streams created for that partitioning
        // if need to split into multiple output streams because of same input having different
        // partitioning to the group)

        // this is because can't currently merge splitting logic into a spout
        // not the most kosher algorithm here, since the grouper indexes are being trounced via the adding of nodes to random groups, but it
        // works out
        //  通过指定的Group获取其额外的PartitionNode
        //  再根据指定outputFields生成唯一的Node
        //  同时找出PartitionNode所对应的parent node 和 该节点的edge信息
        //  清除原有节点
        // 构建新的节点 idNode和newPartitonNode 设置相关的vertex 和 edges 重新添加到graph中
        //
        List<Node> forNewGroups = new ArrayList<Node>();
        for(Group g: mergedGroups) {
            for(PartitionNode n: extraPartitionInputs(g)) {
                Node idNode = makeIdentityNode(n.allOutputFields);
                Node newPartitionNode = new PartitionNode(idNode.streamId, n.name, idNode.allOutputFields, n.thriftGrouping);
                Node parentNode = TridentUtils.getParent(graph, n);
                Set<IndexedEdge> outgoing = graph.outgoingEdgesOf(n);
                graph.removeVertex(n);

                graph.addVertex(idNode);
                graph.addVertex(newPartitionNode);
                addEdge(graph, parentNode, idNode, 0);
                addEdge(graph, idNode, newPartitionNode, 0);
                for(IndexedEdge e: outgoing) {
                    addEdge(graph, newPartitionNode, e.target, e.index);
                }
                Group parentGroup = grouper.nodeGroup(parentNode);
                if(parentGroup==null) {
                    forNewGroups.add(idNode);
                } else {
                    parentGroup.nodes.add(idNode);
                }
            }
        }
        // TODO: in the future, want a way to include this logic in the spout itself,
        // or make it unecessary by having storm include metadata about which grouping a tuple
        // came from

        for(Node n: forNewGroups) {
            grouper.addGroup(new Group(graph, n));
        }

        // add in spouts as groups so we can get parallelisms
        for(Node n: spoutNodes) {
            grouper.addGroup(new Group(graph, n));
        }
        // 每次添加新的group之后需要重新生成index
        grouper.reindex();
        mergedGroups = grouper.getAllGroups();


        Map<Node, String> batchGroupMap = new HashMap();
        //
        List<Set<Node>> connectedComponents = new ConnectivityInspector<Node, IndexedEdge>(graph).connectedSets();
        for(int i=0; i<connectedComponents.size(); i++) {
            String groupId = "bg" + i;
            for(Node n: connectedComponents.get(i)) {
                batchGroupMap.put(n, groupId);
            }
        }

//        System.out.println("GRAPH:");
//        System.out.println(graph);
        // 计算parallelism
        Map<Group, Integer> parallelisms = getGroupParallelisms(graph, grouper, mergedGroups);
        // 创建 topology builder
        TridentTopologyBuilder builder = new TridentTopologyBuilder();

        Map<Node, String> spoutIds = genSpoutIds(spoutNodes);
        Map<Group, String> boltIds = genBoltIds(mergedGroups);
         // 设置 spout
        for(SpoutNode sn: spoutNodes) {
            Integer parallelism = parallelisms.get(grouper.nodeGroup(sn));
            if(sn.type == SpoutNode.SpoutType.DRPC) {
                builder.setBatchPerTupleSpout(spoutIds.get(sn), sn.streamId,
                        (IRichSpout) sn.spout, parallelism, batchGroupMap.get(sn));
            } else {
                ITridentSpout s;
                if(sn.spout instanceof IBatchSpout) {
                    s = new BatchSpoutExecutor((IBatchSpout)sn.spout);
                } else if(sn.spout instanceof ITridentSpout) {
                    s = (ITridentSpout) sn.spout;
                } else {
                    throw new RuntimeException("Regular rich spouts not supported yet... try wrapping in a RichSpoutBatchExecutor");
                    // TODO: handle regular rich spout without batches (need lots of updates to support this throughout)
                }
                builder.setSpout(spoutIds.get(sn), sn.streamId, sn.txId, s, parallelism, batchGroupMap.get(sn));
            }
        }
         //  设置bolt
        for(Group g: mergedGroups) {
            if(!isSpoutGroup(g)) {
                Integer p = parallelisms.get(g);
                Map<String, String> streamToGroup = getOutputStreamBatchGroups(g, batchGroupMap);
                BoltDeclarer d = builder.setBolt(boltIds.get(g), new SubtopologyBolt(graph, g.nodes, batchGroupMap), p,
                        committerBatches(g, batchGroupMap), streamToGroup);
                Collection<PartitionNode> inputs = uniquedSubscriptions(externalGroupInputs(g));
                for(PartitionNode n: inputs) {
                    Node parent = TridentUtils.getParent(graph, n);
                    String componentId;
                    if(parent instanceof SpoutNode) {
                        componentId = spoutIds.get(parent);
                    } else {
                        componentId = boltIds.get(grouper.nodeGroup(parent));
                    }
                    d.grouping(new GlobalStreamId(componentId, n.streamId), n.thriftGrouping);
                }
            }
        }
        // 构建topology
        return builder.buildTopology();
    }

    private static void completeDRPC(DefaultDirectedGraph<Node, IndexedEdge> graph, Map<String, List<Node>> colocate, UniqueIdGen gen) {
        List<Set<Node>> connectedComponents = new ConnectivityInspector<Node, IndexedEdge>(graph).connectedSets();

        for(Set<Node> g: connectedComponents) {
            checkValidJoins(g);
        }

        TridentTopology helper = new TridentTopology(graph, colocate, gen);
        for(Set<Node> g: connectedComponents) {
            SpoutNode drpcNode = getDRPCSpoutNode(g);
            if(drpcNode!=null) {
                Stream lastStream = new Stream(helper, null, getLastAddedNode(g));
                Stream s = new Stream(helper, null, drpcNode);
                helper.multiReduce(
                        s.project(new Fields("return-info"))
                         .batchGlobal(),
                        lastStream.batchGlobal(),
                        new ReturnResultsReducer(),
                        new Fields());
            }
        }               
    }

    private static Node getLastAddedNode(Collection<Node> g) {
        Node ret = null;
        for(Node n: g) {
            if(ret==null || n.creationIndex > ret.creationIndex) {
                ret = n;
            }
        }
        return ret;
    }

    //returns null if it's not a drpc group
    private static SpoutNode getDRPCSpoutNode(Collection<Node> g) {
        for(Node n: g) {
            if(n instanceof SpoutNode) {
                SpoutNode.SpoutType type = ((SpoutNode) n).type;
                if(type==SpoutNode.SpoutType.DRPC) {
                    return (SpoutNode) n;
                }
            }
        }
        return null;
    }

    private static void checkValidJoins(Collection<Node> g) {
        boolean hasDRPCSpout = false;
        boolean hasBatchSpout = false;
        for(Node n: g) {
            if(n instanceof SpoutNode) {
                SpoutNode.SpoutType type = ((SpoutNode) n).type;
                if(type==SpoutNode.SpoutType.BATCH) {
                    hasBatchSpout = true;
                } else if(type==SpoutNode.SpoutType.DRPC) {
                    hasDRPCSpout = true;
                }
            }
        }
        if(hasBatchSpout && hasDRPCSpout) {
            throw new RuntimeException("Cannot join DRPC stream with streams originating from other spouts");
        }
    }

    private static boolean isSpoutGroup(Group g) {
        return g.nodes.size() == 1 && g.nodes.iterator().next() instanceof SpoutNode;
    }

    private static Collection<PartitionNode> uniquedSubscriptions(Set<PartitionNode> subscriptions) {
        Map<String, PartitionNode> ret = new HashMap();
        for(PartitionNode n: subscriptions) {
            PartitionNode curr = ret.get(n.streamId);
            if(curr!=null && !curr.thriftGrouping.equals(n.thriftGrouping)) {
                throw new RuntimeException("Multiple subscriptions to the same stream with different groupings. Should be impossible since that is explicitly guarded against.");
            }
            ret.put(n.streamId, n);
        }
        return ret.values();
    }

    private static Map<Node, String> genSpoutIds(Collection<SpoutNode> spoutNodes) {
        Map<Node, String> ret = new HashMap();
        int ctr = 0;
        for(SpoutNode n: spoutNodes) {
            ret.put(n, "spout" + ctr);
            ctr++;
        }
        return ret;
    }

    private static Map<Group, String> genBoltIds(Collection<Group> groups) {
        Map<Group, String> ret = new HashMap();
        int ctr = 0;
        for(Group g: groups) {
            if(!isSpoutGroup(g)) {
                List<String> name = new ArrayList();
                name.add("b");
                name.add("" + ctr);
                String groupName = getGroupName(g);
                if(groupName!=null && !groupName.isEmpty()) {
                    name.add(getGroupName(g));               
                }
                ret.put(g, Utils.join(name, "-"));
                ctr++;
            }
        }
        return ret;
    }

    private static String getGroupName(Group g) {
        TreeMap<Integer, String> sortedNames = new TreeMap();
        for(Node n: g.nodes) {
            if(n.name!=null) {
                sortedNames.put(n.creationIndex, n.name);
            }
        }
        List<String> names = new ArrayList<String>();
        String prevName = null;
        for(String n: sortedNames.values()) {
            if(prevName==null || !n.equals(prevName)) {
                prevName = n;
                names.add(n);
            }
        }
        return Utils.join(names, "-");
    }

    private static Map<String, String> getOutputStreamBatchGroups(Group g, Map<Node, String> batchGroupMap) {
        Map<String, String> ret = new HashMap();
        Set<PartitionNode> externalGroupOutputs = externalGroupOutputs(g);
        for(PartitionNode n: externalGroupOutputs) {
            ret.put(n.streamId, batchGroupMap.get(n));
        }       
        return ret;
    }

    private static Set<String> committerBatches(Group g, Map<Node, String> batchGroupMap) {
        Set<String> ret = new HashSet();
        for(Node n: g.nodes) {
           if(n instanceof ProcessorNode) {
               if(((ProcessorNode) n).committer) {
                   ret.add(batchGroupMap.get(n));
               }
           }
        }
        return ret;
    }

    private static Map<Group, Integer> getGroupParallelisms(DirectedGraph<Node, IndexedEdge> graph, GraphGrouper grouper, Collection<Group> groups) {
        UndirectedGraph<Group, Object> equivs = new Pseudograph<Group, Object>(Object.class);
        for(Group g: groups) {
            equivs.addVertex(g);
        }
        for(Group g: groups) {
            for(PartitionNode n: externalGroupInputs(g)) {
                if(isIdentityPartition(n)) {
                    Node parent = TridentUtils.getParent(graph, n);
                    Group parentGroup = grouper.nodeGroup(parent);
                    if(parentGroup!=null && !parentGroup.equals(g)) {
                        equivs.addEdge(parentGroup, g);
                    }
                }
            }           
        }

        Map<Group, Integer> ret = new HashMap();
        List<Set<Group>> equivGroups = new ConnectivityInspector<Group, Object>(equivs).connectedSets();
        for(Set<Group> equivGroup: equivGroups) {
            Integer fixedP = getFixedParallelism(equivGroup);
            Integer maxP = getMaxParallelism(equivGroup);
            if(fixedP!=null && maxP!=null && maxP < fixedP) {
                throw new RuntimeException("Parallelism is fixed to " + fixedP + " but max parallelism is less than that: " + maxP);
            }


            Integer p = 1;
            for(Group g: equivGroup) {
                for(Node n: g.nodes) {
                    if(n.parallelismHint!=null) {
                        p = Math.max(p, n.parallelismHint);
                    }
                }
            }
            if(maxP!=null) p = Math.min(maxP, p);

            if(fixedP!=null) p = fixedP;
            for(Group g: equivGroup) {
                ret.put(g, p);
            }
        }
        return ret;
    }

    private static Integer getMaxParallelism(Set<Group> groups) {
        Integer ret = null;
        for(Group g: groups) {
            if(isSpoutGroup(g)) {
                SpoutNode n = (SpoutNode) g.nodes.iterator().next();
                Map conf = getSpoutComponentConfig(n.spout);
                if(conf==null) conf = new HashMap();
                Number maxP = (Number) conf.get(Config.TOPOLOGY_MAX_TASK_PARALLELISM);
                if(maxP!=null) {
                    if(ret==null) ret = maxP.intValue();
                    else ret = Math.min(ret, maxP.intValue());
                }
            }
        }
        return ret;
    }

    private static Map getSpoutComponentConfig(Object spout) {
        if(spout instanceof IRichSpout) {
            return ((IRichSpout) spout).getComponentConfiguration();
        } else if (spout instanceof IBatchSpout) {
            return ((IBatchSpout) spout).getComponentConfiguration();
        } else {
            return ((ITridentSpout) spout).getComponentConfiguration();
        }
    }

    private static Integer getFixedParallelism(Set<Group> groups) {
        Integer ret = null;
        for(Group g: groups) {
            for(Node n: g.nodes) {
                if(n.stateInfo != null && n.stateInfo.spec.requiredNumPartitions!=null) {
                    int reqPartitions = n.stateInfo.spec.requiredNumPartitions;
                    if(ret!=null && ret!=reqPartitions) {
                        throw new RuntimeException("Cannot have one group have fixed parallelism of two different values");
                    }
                    ret = reqPartitions;
                }
            }
        }
        return ret;
    }

    private static boolean isIdentityPartition(PartitionNode n) {
        Grouping g = n.thriftGrouping;
        if(g.is_set_custom_serialized()) {
            CustomStreamGrouping csg = (CustomStreamGrouping) Utils.javaDeserialize(g.get_custom_serialized(), Serializable.class);
            return csg instanceof IdentityGrouping;
        }
        return false;
    }

    private static void addEdge(DirectedGraph g, Object source, Object target, int index) {
        g.addEdge(source, target, new IndexedEdge(source, target, index));
    }

    private Node makeIdentityNode(Fields allOutputFields) {
        return new ProcessorNode(getUniqueStreamId(), null, allOutputFields, new Fields(),
                new EachProcessor(new Fields(), new FilterExecutor(new TrueFilter())));
    }

    private static List<PartitionNode> extraPartitionInputs(Group g) {
        List<PartitionNode> ret = new ArrayList();
        Set<PartitionNode> inputs = externalGroupInputs(g);
        Map<String, List<PartitionNode>> grouped = new HashMap();
        for(PartitionNode n: inputs) {
            if(!grouped.containsKey(n.streamId)) {
                grouped.put(n.streamId, new ArrayList());
            }
            grouped.get(n.streamId).add(n);
        }
        for(List<PartitionNode> group: grouped.values()) {
            PartitionNode anchor = group.get(0);
            for(int i=1; i<group.size(); i++) {
                PartitionNode n = group.get(i);
                if(!n.thriftGrouping.equals(anchor.thriftGrouping)) {
                    ret.add(n);
                }
            }
        }
        return ret;
    }

    private static Set<PartitionNode> externalGroupInputs(Group g) {
        Set<PartitionNode> ret = new HashSet();
        for(Node n: g.incomingNodes()) {
            if(n instanceof PartitionNode) {
                ret.add((PartitionNode) n);
            }
        }
        return ret;
    }

    private static Set<PartitionNode> externalGroupOutputs(Group g) {
        Set<PartitionNode> ret = new HashSet();
        for(Node n: g.outgoingNodes()) {
            if(n instanceof PartitionNode) {
                ret.add((PartitionNode) n);
            }
        }
        return ret;
    }   

    private static PartitionNode makeIdentityPartition(Node basis) {
        return new PartitionNode(basis.streamId, basis.name, basis.allOutputFields,
            Grouping.custom_serialized(Utils.javaSerialize(new IdentityGrouping())));
    }


    protected String getUniqueStreamId() {
        return _gen.getUniqueStreamId();
    }

    protected String getUniqueStateId() {
        return _gen.getUniqueStateId();
    }

    protected void registerNode(Node n) {
        _graph.addVertex(n);
        if(n.stateInfo!=null) {
            String id = n.stateInfo.id;
            if(!_colocate.containsKey(id)) {
                _colocate.put(id, new ArrayList());
            }
            _colocate.get(id).add(n);
        }
    }

    protected Stream addNode(Node n) {
        registerNode(n);
        return new Stream(this, n.name, n);
    }

    protected void registerSourcedNode(List<Stream> sources, Node newNode) {
        registerNode(newNode);
        int streamIndex = 0;
        for(Stream s: sources) {
            _graph.addEdge(s._node, newNode, new IndexedEdge(s._node, newNode, streamIndex));
            streamIndex++;
        }       
    }

    protected Stream addSourcedNode(List<Stream> sources, Node newNode) {
        registerSourcedNode(sources, newNode);
        return new Stream(this, newNode.name, newNode);
    }

    protected TridentState addSourcedStateNode(List<Stream> sources, Node newNode) {
        registerSourcedNode(sources, newNode);
        return new TridentState(this, newNode);
    }   

    protected Stream addSourcedNode(Stream source, Node newNode) {
        return addSourcedNode(Arrays.asList(source), newNode);
    }

    protected TridentState addSourcedStateNode(Stream source, Node newNode) {
        return addSourcedStateNode(Arrays.asList(source), newNode);
    }      

    private static List<Fields> getAllOutputFields(List streams) {
        List<Fields> ret = new ArrayList<Fields>();
        for(Object o: streams) {
            ret.add(((IAggregatableStream) o).getOutputFields());
        }
        return ret;
    }


    private static List<GroupedStream> groupedStreams(List<Stream> streams, List<Fields> joinFields) {
        List<GroupedStream> ret = new ArrayList<GroupedStream>();
        for(int i=0; i<streams.size(); i++) {
            ret.add(streams.get(i).groupBy(joinFields.get(i)));
        }
        return ret;
    }

    private static List<Fields> strippedInputFields(List<Stream> streams, List<Fields> joinFields) {
        List<Fields> ret = new ArrayList<Fields>();
        for(int i=0; i<streams.size(); i++) {
            ret.add(TridentUtils.fieldsSubtract(streams.get(i).getOutputFields(), joinFields.get(i)));
        }
        return ret;
    }

    private static List<JoinType> repeat(int n, JoinType type) {
        List<JoinType> ret = new ArrayList<JoinType>();
        for(int i=0; i<n; i++) {
            ret.add(type);
        }
        return ret;
    }
}
分享到:
评论

相关推荐

    jstorm源码解析之bolt异常处理方法.docx

    在JStorm的源码中,我们可以看到Bolt执行的核心逻辑位于`BasicBoltExecutor`类的`execute()`方法。该方法首先设置上下文,然后尝试执行Bolt的业务逻辑,最后确认处理成功。如果在执行过程中抛出`FailedException`,...

    Storm 源码分析

    - **Nimbus**:Nimbus是Storm集群的核心组件之一,负责整个集群的管理和协调工作,包括任务调度、故障恢复等。Nimbus通过Zookeeper来实现状态同步和集群协调。 - **Supervisor**:Supervisor运行在每个Worker节点上...

    jstorm源码解析之bolt异常处理方法

    在本篇文章中,我们将深入探讨JStorm源码中关于Bolt异常处理方法的实现细节。首先,我们了解到如果在Bolt代码中出现未被捕获的异常,所在的工作进程会直接退出。这一点与Storm的设计相似,旨在避免问题被掩盖而无法...

    JStorm:JStorm原始码学习-源码包

    JStorm原始码学习:主要包含Storm重新启动,Nimbus启动,Supervisor启动,Executor创建和启动 风暴编程模型 Nimbus:负责资源分配和任务调度。 主管:负责接受nimbus分配的任务,启动和停止属于自己管理的worker...

    jstorm storm入门demo

    1. **JStorm与Storm的关系**:讲解JStorm与Apache Storm的关系,包括JStorm的诞生背景、主要改进和优化之处。 2. **JStorm的核心概念**:介绍Spout、Bolt、拓扑等核心概念,解释它们在实时数据处理中的作用。 3. *...

    大数据技术分享 JStorm介绍 JStorm-分布式实时计算引擎 共40页.pptx

    **JStorm介绍** JStorm是由阿里巴巴开发的分布式实时计算引擎,它是基于Apache Storm的一个增强版本。JStorm的设计目标是提供比Storm更高的稳定性、更强的功能和更快的计算性能。这个技术分享主要涵盖了JStorm的...

    阿里中间件性能挑战赛初赛-RocketMQ+JStorm+Tair实时统计双11交易金额源码+学习说明.zip

    阿里中间件性能挑战赛初赛-RocketMQ+JStorm+Tair实时统计双11交易金额源码+学习说明.zip阿里中间件性能挑战赛初赛-RocketMQ+JStorm+Tair实时统计双11交易金额源码+学习说明.zip阿里中间件性能挑战赛初赛-RocketMQ+...

    jstorm集成kafka代码实例

    本示例将探讨如何将Apache JStorm与Kafka进行集成,以利用它们各自的优势。JStorm是一款分布式实时计算系统,它允许用户处理数据流,而Kafka则是一个高吞吐量的分布式消息系统,用于实时数据传递。接下来,我们将...

    jstorm集成kafka插件demo

    在IT行业中,分布式计算系统和实时数据处理是关键领域,其中JStorm和Kafka都是重要的组件。本示例探讨了如何将JStorm与Kafka进行集成,以实现高效的数据流处理。下面,我们将深入理解这两个工具,以及它们结合使用的...

    jstorm 阿里巴巴官方文档 pdf

    接着,文档会详细阐述JStorm的安装与配置过程,包括下载源码、编译、部署以及设置环境变量。对于初学者,这部分内容至关重要,因为正确配置JStorm环境是后续开发和运行的基础。 在JStorm的使用方法部分,你会了解到...

    JStorm 2.1.1 API

    **JStorm 2.1.1 API 深度解析** JStorm是阿里巴巴开源的一款分布式实时计算系统,它基于Apache Storm并进行了大量的优化,提供了一种高效、稳定且易用的流处理框架。JStorm 2.1.1是其一个重要的版本,此版本的API为...

    jstorm example

    **JStorm简介** JStorm是阿里巴巴开源的一款分布式实时计算系统,它是基于Apache Storm的一个高性能、高可用、热扩展的实时处理框架。JStorm的核心设计理念是简单、高效和稳定,能够处理大规模的数据流处理任务,...

    jstorm文档

    JStorm是中国淘宝团队开发的一款分布式实时计算系统,它是基于Apache Storm的设计理念,但在性能、稳定性和易用性上进行了大量的优化。本文档将详细介绍JStorm的基础知识,包括其核心概念、工作原理、安装配置以及...

    jstorm课程

    标题"jstorm课程"指的是关于JStorm的教育课程,JStorm是阿里巴巴开源的一个分布式实时计算框架,它是Apache Storm的Java版本,主要用于处理大规模数据流的实时计算。 描述中提到的"全套storm资料初学者必备 比较...

    jstorm-2.2.1

    《JStorm 2.2.1:分布式流处理框架详解》 JStorm是阿里巴巴开源的一款高性能、高可靠的分布式实时计算系统,它基于Apache Storm并针对大规模数据处理进行了优化。JStorm 2.2.1是该框架的一个稳定版本,提供了许多...

    storm-core源码

    storm-core-1.0.3-sources.jar 源码文件,1.0.3版本

    Jstorm 到 Apache Flink 的迁移实践.pdf

    1. **内存隔离与资源管理**:Flink作业运行在YARN之上,提供了内存隔离,确保了作业间的资源独立,同时YARN的队列机制便于进行资源分配和预算审核。 2. **统一资源管理**:通过YARN管理资源,减轻了运维负担,实现了...

    Jstorm介绍

    jstorm框架介绍,包含架构图、jstorm安装部署以及配置、如何在jstorm框架里写业务代码。

    alibaba-jstorm

    阿里巴巴JStorm是一款由阿里巴巴开源的分布式实时计算系统,它基于Apache Storm,但在性能、稳定性、易用性等多个方面进行了优化,是大数据处理领域的重要工具之一。JStorm的设计理念是“简单、稳定、低延迟”,旨在...

    JStorm 2.2.1下载 (百度网盘提取链接和密码)

    JStorm 2.2.1 百度网盘下载,Alibaba JStorm 是一个强大的企业级流式计算引擎

Global site tag (gtag.js) - Google Analytics