Examples with SinkPlanNode org.apache.flink.optimizer.plan.SinkPlanNode used on opensource projects

Example 1 with SinkPlanNode

use of org.apache.flink.optimizer.plan.SinkPlanNode in project flink by apache.

the class PregelCompilerTest method testPregelCompilerWithBroadcastVariable.

@SuppressWarnings("serial")
@Test
public void testPregelCompilerWithBroadcastVariable() {
    final String broadcastSetName = "broadcast";
    ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
    env.setParallelism(DEFAULT_PARALLELISM);
    // compose test program
    {
        DataSet<Long> bcVar = env.fromElements(1L);
        DataSet<Vertex<Long, Long>> initialVertices = env.fromElements(new Tuple2<>(1L, 1L), new Tuple2<>(2L, 2L)).map(new Tuple2ToVertexMap<>());
        DataSet<Edge<Long, NullValue>> edges = env.fromElements(new Tuple2<>(1L, 2L)).map(new MapFunction<Tuple2<Long, Long>, Edge<Long, NullValue>>() {

            public Edge<Long, NullValue> map(Tuple2<Long, Long> edge) {
                return new Edge<>(edge.f0, edge.f1, NullValue.getInstance());
            }
        });
        Graph<Long, Long, NullValue> graph = Graph.fromDataSet(initialVertices, edges, env);
        VertexCentricConfiguration parameters = new VertexCentricConfiguration();
        parameters.addBroadcastSet(broadcastSetName, bcVar);
        DataSet<Vertex<Long, Long>> result = graph.runVertexCentricIteration(new CCCompute(), null, 100, parameters).getVertices();
        result.output(new DiscardingOutputFormat<>());
    }
    Plan p = env.createProgramPlan("Pregel Connected Components");
    OptimizedPlan op = compileNoStats(p);
    // check the sink
    SinkPlanNode sink = op.getDataSinks().iterator().next();
    assertEquals(ShipStrategyType.FORWARD, sink.getInput().getShipStrategy());
    assertEquals(DEFAULT_PARALLELISM, sink.getParallelism());
    // check the iteration
    WorksetIterationPlanNode iteration = (WorksetIterationPlanNode) sink.getInput().getSource();
    assertEquals(DEFAULT_PARALLELISM, iteration.getParallelism());
    // check the solution set delta
    PlanNode ssDelta = iteration.getSolutionSetDeltaPlanNode();
    assertTrue(ssDelta instanceof SingleInputPlanNode);
    SingleInputPlanNode ssFlatMap = (SingleInputPlanNode) ((SingleInputPlanNode) (ssDelta)).getInput().getSource();
    assertEquals(DEFAULT_PARALLELISM, ssFlatMap.getParallelism());
    assertEquals(ShipStrategyType.FORWARD, ssFlatMap.getInput().getShipStrategy());
    // check the computation coGroup
    DualInputPlanNode computationCoGroup = (DualInputPlanNode) (ssFlatMap.getInput().getSource());
    assertEquals(DEFAULT_PARALLELISM, computationCoGroup.getParallelism());
    assertEquals(ShipStrategyType.FORWARD, computationCoGroup.getInput1().getShipStrategy());
    assertEquals(ShipStrategyType.PARTITION_HASH, computationCoGroup.getInput2().getShipStrategy());
    assertTrue(computationCoGroup.getInput2().getTempMode().isCached());
    assertEquals(new FieldList(0), computationCoGroup.getInput2().getShipStrategyKeys());
    // check that the initial partitioning is pushed out of the loop
    assertEquals(ShipStrategyType.PARTITION_HASH, iteration.getInput1().getShipStrategy());
    assertEquals(new FieldList(0), iteration.getInput1().getShipStrategyKeys());
}

Example 2 with SinkPlanNode

use of org.apache.flink.optimizer.plan.SinkPlanNode in project flink by apache.

the class SpargelCompilerTest method testSpargelCompilerWithBroadcastVariable.

@SuppressWarnings("serial")
@Test
public void testSpargelCompilerWithBroadcastVariable() {
    final String broadcastVariableName = "broadcast variable";
    ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
    env.setParallelism(DEFAULT_PARALLELISM);
    // compose test program
    DataSet<Long> bcVar = env.fromElements(1L);
    DataSet<Vertex<Long, Long>> initialVertices = env.fromElements(new Tuple2<>(1L, 1L), new Tuple2<>(2L, 2L)).map(new Tuple2ToVertexMap<>());
    DataSet<Edge<Long, NullValue>> edges = env.fromElements(new Tuple2<>(1L, 2L)).map(new MapFunction<Tuple2<Long, Long>, Edge<Long, NullValue>>() {

        public Edge<Long, NullValue> map(Tuple2<Long, Long> edge) {
            return new Edge<>(edge.f0, edge.f1, NullValue.getInstance());
        }
    });
    Graph<Long, Long, NullValue> graph = Graph.fromDataSet(initialVertices, edges, env);
    ScatterGatherConfiguration parameters = new ScatterGatherConfiguration();
    parameters.addBroadcastSetForScatterFunction(broadcastVariableName, bcVar);
    parameters.addBroadcastSetForGatherFunction(broadcastVariableName, bcVar);
    DataSet<Vertex<Long, Long>> result = graph.runScatterGatherIteration(new ConnectedComponents.CCMessenger<>(BasicTypeInfo.LONG_TYPE_INFO), new ConnectedComponents.CCUpdater<>(), 100).getVertices();
    result.output(new DiscardingOutputFormat<>());
    Plan p = env.createProgramPlan("Spargel Connected Components");
    OptimizedPlan op = compileNoStats(p);
    // check the sink
    SinkPlanNode sink = op.getDataSinks().iterator().next();
    assertEquals(ShipStrategyType.FORWARD, sink.getInput().getShipStrategy());
    assertEquals(DEFAULT_PARALLELISM, sink.getParallelism());
    // check the iteration
    WorksetIterationPlanNode iteration = (WorksetIterationPlanNode) sink.getInput().getSource();
    assertEquals(DEFAULT_PARALLELISM, iteration.getParallelism());
    // check the solution set join and the delta
    PlanNode ssDelta = iteration.getSolutionSetDeltaPlanNode();
    assertTrue(ssDelta instanceof // this is only true if the update functions preserves
    DualInputPlanNode);
    // the partitioning
    DualInputPlanNode ssJoin = (DualInputPlanNode) ssDelta;
    assertEquals(DEFAULT_PARALLELISM, ssJoin.getParallelism());
    assertEquals(ShipStrategyType.PARTITION_HASH, ssJoin.getInput1().getShipStrategy());
    assertEquals(new FieldList(0), ssJoin.getInput1().getShipStrategyKeys());
    // check the workset set join
    DualInputPlanNode edgeJoin = (DualInputPlanNode) ssJoin.getInput1().getSource();
    assertEquals(DEFAULT_PARALLELISM, edgeJoin.getParallelism());
    assertEquals(ShipStrategyType.PARTITION_HASH, edgeJoin.getInput1().getShipStrategy());
    assertEquals(ShipStrategyType.FORWARD, edgeJoin.getInput2().getShipStrategy());
    assertTrue(edgeJoin.getInput1().getTempMode().isCached());
    assertEquals(new FieldList(0), edgeJoin.getInput1().getShipStrategyKeys());
    // check that the initial partitioning is pushed out of the loop
    assertEquals(ShipStrategyType.PARTITION_HASH, iteration.getInput1().getShipStrategy());
    assertEquals(ShipStrategyType.PARTITION_HASH, iteration.getInput2().getShipStrategy());
    assertEquals(new FieldList(0), iteration.getInput1().getShipStrategyKeys());
    assertEquals(new FieldList(0), iteration.getInput2().getShipStrategyKeys());
}

Example 3 with SinkPlanNode

use of org.apache.flink.optimizer.plan.SinkPlanNode in project flink by apache.

the class GSACompilerTest method testGSACompiler.

@Test
public void testGSACompiler() {
    ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
    env.setParallelism(DEFAULT_PARALLELISM);
    // compose test program
    DataSet<Edge<Long, NullValue>> edges = env.fromElements(new Tuple3<>(1L, 2L, NullValue.getInstance())).map(new Tuple3ToEdgeMap<>());
    Graph<Long, Long, NullValue> graph = Graph.fromDataSet(edges, new InitVertices(), env);
    DataSet<Vertex<Long, Long>> result = graph.runGatherSumApplyIteration(new GatherNeighborIds(), new SelectMinId(), new UpdateComponentId(), 100).getVertices();
    result.output(new DiscardingOutputFormat<>());
    Plan p = env.createProgramPlan("GSA Connected Components");
    OptimizedPlan op = compileNoStats(p);
    // check the sink
    SinkPlanNode sink = op.getDataSinks().iterator().next();
    assertEquals(ShipStrategyType.FORWARD, sink.getInput().getShipStrategy());
    assertEquals(DEFAULT_PARALLELISM, sink.getParallelism());
    assertEquals(PartitioningProperty.HASH_PARTITIONED, sink.getGlobalProperties().getPartitioning());
    // check the iteration
    WorksetIterationPlanNode iteration = (WorksetIterationPlanNode) sink.getInput().getSource();
    assertEquals(DEFAULT_PARALLELISM, iteration.getParallelism());
    // check the solution set join and the delta
    PlanNode ssDelta = iteration.getSolutionSetDeltaPlanNode();
    assertTrue(ssDelta instanceof // this is only true if the update function preserves
    DualInputPlanNode);
    // the partitioning
    DualInputPlanNode ssJoin = (DualInputPlanNode) ssDelta;
    assertEquals(DEFAULT_PARALLELISM, ssJoin.getParallelism());
    assertEquals(ShipStrategyType.PARTITION_HASH, ssJoin.getInput1().getShipStrategy());
    assertEquals(new FieldList(0), ssJoin.getInput1().getShipStrategyKeys());
    // check the workset set join
    SingleInputPlanNode sumReducer = (SingleInputPlanNode) ssJoin.getInput1().getSource();
    SingleInputPlanNode gatherMapper = (SingleInputPlanNode) sumReducer.getInput().getSource();
    DualInputPlanNode edgeJoin = (DualInputPlanNode) gatherMapper.getInput().getSource();
    assertEquals(DEFAULT_PARALLELISM, edgeJoin.getParallelism());
    // input1 is the workset
    assertEquals(ShipStrategyType.FORWARD, edgeJoin.getInput1().getShipStrategy());
    // input2 is the edges
    assertEquals(ShipStrategyType.PARTITION_HASH, edgeJoin.getInput2().getShipStrategy());
    assertTrue(edgeJoin.getInput2().getTempMode().isCached());
    assertEquals(new FieldList(0), edgeJoin.getInput2().getShipStrategyKeys());
}

Example 4 with SinkPlanNode

use of org.apache.flink.optimizer.plan.SinkPlanNode in project flink by apache.

the class PartitionOperatorTest method testRangePartitionOperatorPreservesFields2.

@Test
public void testRangePartitionOperatorPreservesFields2() {
    try {
        ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
        DataSet<Tuple2<Long, Long>> data = env.fromCollection(Collections.singleton(new Tuple2<>(0L, 0L)));
        PartitionOperator<Tuple2<Long, Long>> rangePartitioned = data.partitionByRange(1);
        rangePartitioned.groupBy(1).reduceGroup(new IdentityGroupReducerCombinable<Tuple2<Long, Long>>()).output(new DiscardingOutputFormat<Tuple2<Long, Long>>());
        data.groupBy(0).aggregate(Aggregations.SUM, 1).map(new MapFunction<Tuple2<Long, Long>, Long>() {

            @Override
            public Long map(Tuple2<Long, Long> value) throws Exception {
                return value.f1;
            }
        }).output(new DiscardingOutputFormat<Long>());
        rangePartitioned.filter(new FilterFunction<Tuple2<Long, Long>>() {

            @Override
            public boolean filter(Tuple2<Long, Long> value) throws Exception {
                return value.f0 % 2 == 0;
            }
        }).output(new DiscardingOutputFormat<Tuple2<Long, Long>>());
        Plan p = env.createProgramPlan();
        OptimizedPlan op = compileNoStats(p);
        SinkPlanNode sink = op.getDataSinks().iterator().next();
        SingleInputPlanNode reducer = (SingleInputPlanNode) sink.getInput().getSource();
        SingleInputPlanNode partitionNode = (SingleInputPlanNode) reducer.getInput().getSource();
        SingleInputPlanNode partitionIDRemover = (SingleInputPlanNode) partitionNode.getInput().getSource();
        assertEquals(ShipStrategyType.FORWARD, reducer.getInput().getShipStrategy());
        assertEquals(ShipStrategyType.FORWARD, partitionNode.getInput().getShipStrategy());
        assertEquals(ShipStrategyType.PARTITION_CUSTOM, partitionIDRemover.getInput().getShipStrategy());
        SourcePlanNode sourcePlanNode = op.getDataSources().iterator().next();
        List<Channel> sourceOutgoingChannels = sourcePlanNode.getOutgoingChannels();
        assertEquals(3, sourceOutgoingChannels.size());
        assertEquals(ShipStrategyType.FORWARD, sourceOutgoingChannels.get(0).getShipStrategy());
        assertEquals(ShipStrategyType.FORWARD, sourceOutgoingChannels.get(1).getShipStrategy());
        assertEquals(ShipStrategyType.FORWARD, sourceOutgoingChannels.get(2).getShipStrategy());
        assertEquals(DataExchangeMode.PIPELINED, sourceOutgoingChannels.get(0).getDataExchangeMode());
        assertEquals(DataExchangeMode.PIPELINED, sourceOutgoingChannels.get(1).getDataExchangeMode());
        assertEquals(DataExchangeMode.BATCH, sourceOutgoingChannels.get(2).getDataExchangeMode());
        List<Channel> partitionOutputChannels = partitionNode.getOutgoingChannels();
        assertEquals(2, partitionOutputChannels.size());
        assertEquals(ShipStrategyType.FORWARD, partitionOutputChannels.get(0).getShipStrategy());
        assertEquals(ShipStrategyType.FORWARD, partitionOutputChannels.get(1).getShipStrategy());
        assertEquals(DataExchangeMode.PIPELINED, partitionOutputChannels.get(0).getDataExchangeMode());
        assertEquals(DataExchangeMode.PIPELINED, partitionOutputChannels.get(1).getDataExchangeMode());
    } catch (Exception e) {
        e.printStackTrace();
        fail(e.getMessage());
    }
}

Example 5 with SinkPlanNode

use of org.apache.flink.optimizer.plan.SinkPlanNode in project flink by apache.

the class ReduceCompilationTest method testGroupedReduceWithSelectorFunctionKey.

@Test
public void testGroupedReduceWithSelectorFunctionKey() {
    try {
        ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
        env.setParallelism(8);
        DataSet<Tuple2<String, Double>> data = env.readCsvFile("file:///will/never/be/read").types(String.class, Double.class).name("source").setParallelism(6);
        data.groupBy(new KeySelector<Tuple2<String, Double>, String>() {

            public String getKey(Tuple2<String, Double> value) {
                return value.f0;
            }
        }).reduce(new RichReduceFunction<Tuple2<String, Double>>() {

            @Override
            public Tuple2<String, Double> reduce(Tuple2<String, Double> value1, Tuple2<String, Double> value2) {
                return null;
            }
        }).name("reducer").output(new DiscardingOutputFormat<Tuple2<String, Double>>()).name("sink");
        Plan p = env.createProgramPlan();
        OptimizedPlan op = compileNoStats(p);
        OptimizerPlanNodeResolver resolver = getOptimizerPlanNodeResolver(op);
        // get the original nodes
        SourcePlanNode sourceNode = resolver.getNode("source");
        SingleInputPlanNode reduceNode = resolver.getNode("reducer");
        SinkPlanNode sinkNode = resolver.getNode("sink");
        // get the combiner
        SingleInputPlanNode combineNode = (SingleInputPlanNode) reduceNode.getInput().getSource();
        // get the key extractors and projectors
        SingleInputPlanNode keyExtractor = (SingleInputPlanNode) combineNode.getInput().getSource();
        SingleInputPlanNode keyProjector = (SingleInputPlanNode) sinkNode.getInput().getSource();
        // check wiring
        assertEquals(sourceNode, keyExtractor.getInput().getSource());
        assertEquals(keyProjector, sinkNode.getInput().getSource());
        // check the strategies
        assertEquals(DriverStrategy.SORTED_REDUCE, reduceNode.getDriverStrategy());
        assertEquals(DriverStrategy.SORTED_PARTIAL_REDUCE, combineNode.getDriverStrategy());
        // check the keys
        assertEquals(new FieldList(0), reduceNode.getKeys(0));
        assertEquals(new FieldList(0), combineNode.getKeys(0));
        assertEquals(new FieldList(0), reduceNode.getInput().getLocalStrategyKeys());
        // check parallelism
        assertEquals(6, sourceNode.getParallelism());
        assertEquals(6, keyExtractor.getParallelism());
        assertEquals(6, combineNode.getParallelism());
        assertEquals(8, reduceNode.getParallelism());
        assertEquals(8, keyProjector.getParallelism());
        assertEquals(8, sinkNode.getParallelism());
    } catch (Exception e) {
        System.err.println(e.getMessage());
        e.printStackTrace();
        fail(e.getClass().getSimpleName() + " in test: " + e.getMessage());
    }
}