Examples with SpoutNode storm.trident.planner.SpoutNode used on opensource projects

Example 1 with SpoutNode

use of storm.trident.planner.SpoutNode in project storm by nathanmarz.

the class TridentTopology method build.

public StormTopology build() {
    DefaultDirectedGraph<Node, IndexedEdge> graph = (DefaultDirectedGraph) _graph.clone();
    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>();
    for (Node n : graph.vertexSet()) {
        if (n instanceof SpoutNode) {
            spoutNodes.add((SpoutNode) n);
        } else if (!(n instanceof PartitionNode)) {
            boltNodes.add(n);
        }
    }
    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
    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
    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);
            }
        }
    }
    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));
    }
    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);
    Map<Group, Integer> parallelisms = getGroupParallelisms(graph, grouper, mergedGroups);
    TridentTopologyBuilder builder = new TridentTopologyBuilder();
    Map<Node, String> spoutIds = genSpoutIds(spoutNodes);
    Map<Group, String> boltIds = genBoltIds(mergedGroups);
    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));
        }
    }
    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);
            }
        }
    }
    return builder.buildTopology();
}

Example 2 with SpoutNode

use of storm.trident.planner.SpoutNode in project storm by nathanmarz.

the class TridentTopology method completeDRPC.

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());
        }
    }
}

Example 3 with SpoutNode

use of storm.trident.planner.SpoutNode in project storm by nathanmarz.

the class TridentTopology method checkValidJoins.

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");
    }
}

Example 4 with SpoutNode

use of storm.trident.planner.SpoutNode in project storm by nathanmarz.

the class TridentTopology method newDRPCStream.

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"));
}

Example 5 with SpoutNode

use of storm.trident.planner.SpoutNode in project jstorm by alibaba.

the class TridentTopology method genSpoutIds.

private static Map<Node, String> genSpoutIds(Collection<SpoutNode> spoutNodes) {
    Map<Node, String> ret = new HashMap<>();
    int ctr = 0;
    for (SpoutNode n : spoutNodes) {
        if (n.type == SpoutNode.SpoutType.BATCH) {
            // if Batch spout then id contains txId
            ret.put(n, "spout-" + n.txId);
        } else if (n.type == SpoutNode.SpoutType.DRPC) {
            //if DRPC spout then id contains function
            ret.put(n, "spout-" + ((DRPCSpout) n.spout).get_function() + ctr);
            ctr++;
        } else {
            ret.put(n, "spout" + ctr);
            ctr++;
        }
    }
    return ret;
}