Search in sources :

Example 26 with NoRestartStrategy

use of org.apache.flink.runtime.executiongraph.restart.NoRestartStrategy in project flink by apache.

the class PointwisePatternTest method testHighToLow.

private void testHighToLow(int highDop, int lowDop) throws Exception {
    if (highDop < lowDop) {
        throw new IllegalArgumentException();
    }
    final int factor = highDop / lowDop;
    final int delta = highDop % lowDop == 0 ? 0 : 1;
    JobVertex v1 = new JobVertex("vertex1");
    JobVertex v2 = new JobVertex("vertex2");
    v1.setParallelism(highDop);
    v2.setParallelism(lowDop);
    v1.setInvokableClass(AbstractInvokable.class);
    v2.setInvokableClass(AbstractInvokable.class);
    v2.connectNewDataSetAsInput(v1, DistributionPattern.POINTWISE, ResultPartitionType.PIPELINED);
    List<JobVertex> ordered = new ArrayList<JobVertex>(Arrays.asList(v1, v2));
    ExecutionGraph eg = new ExecutionGraph(TestingUtils.defaultExecutor(), TestingUtils.defaultExecutor(), jobId, jobName, cfg, new SerializedValue<>(new ExecutionConfig()), AkkaUtils.getDefaultTimeout(), new NoRestartStrategy(), new Scheduler(TestingUtils.defaultExecutionContext()));
    try {
        eg.attachJobGraph(ordered);
    } catch (JobException e) {
        e.printStackTrace();
        fail("Job failed with exception: " + e.getMessage());
    }
    ExecutionJobVertex target = eg.getAllVertices().get(v2.getID());
    int[] timesUsed = new int[highDop];
    for (ExecutionVertex ev : target.getTaskVertices()) {
        assertEquals(1, ev.getNumberOfInputs());
        ExecutionEdge[] inEdges = ev.getInputEdges(0);
        assertTrue(inEdges.length >= factor && inEdges.length <= factor + delta);
        for (ExecutionEdge ee : inEdges) {
            timesUsed[ee.getSource().getPartitionNumber()]++;
        }
    }
    for (int used : timesUsed) {
        assertEquals(1, used);
    }
}
Also used : Scheduler(org.apache.flink.runtime.jobmanager.scheduler.Scheduler) ArrayList(java.util.ArrayList) ExecutionConfig(org.apache.flink.api.common.ExecutionConfig) NoRestartStrategy(org.apache.flink.runtime.executiongraph.restart.NoRestartStrategy) JobException(org.apache.flink.runtime.JobException) JobVertex(org.apache.flink.runtime.jobgraph.JobVertex)

Example 27 with NoRestartStrategy

use of org.apache.flink.runtime.executiongraph.restart.NoRestartStrategy in project flink by apache.

the class RescalePartitionerTest method testExecutionGraphGeneration.

@Test
public void testExecutionGraphGeneration() throws Exception {
    final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
    env.setParallelism(4);
    // get input data
    DataStream<String> text = env.addSource(new ParallelSourceFunction<String>() {

        private static final long serialVersionUID = 7772338606389180774L;

        @Override
        public void run(SourceContext<String> ctx) throws Exception {
        }

        @Override
        public void cancel() {
        }
    }).setParallelism(2);
    DataStream<Tuple2<String, Integer>> counts = text.rescale().flatMap(new FlatMapFunction<String, Tuple2<String, Integer>>() {

        private static final long serialVersionUID = -5255930322161596829L;

        @Override
        public void flatMap(String value, Collector<Tuple2<String, Integer>> out) throws Exception {
        }
    });
    counts.rescale().print().setParallelism(2);
    JobGraph jobGraph = env.getStreamGraph().getJobGraph();
    final JobID jobId = new JobID();
    final String jobName = "Semi-Rebalance Test Job";
    final Configuration cfg = new Configuration();
    List<JobVertex> jobVertices = jobGraph.getVerticesSortedTopologicallyFromSources();
    JobVertex sourceVertex = jobVertices.get(0);
    JobVertex mapVertex = jobVertices.get(1);
    JobVertex sinkVertex = jobVertices.get(2);
    assertEquals(2, sourceVertex.getParallelism());
    assertEquals(4, mapVertex.getParallelism());
    assertEquals(2, sinkVertex.getParallelism());
    ExecutionGraph eg = new ExecutionGraph(TestingUtils.defaultExecutor(), TestingUtils.defaultExecutor(), jobId, jobName, cfg, new SerializedValue<>(new ExecutionConfig()), AkkaUtils.getDefaultTimeout(), new NoRestartStrategy(), new ArrayList<BlobKey>(), new ArrayList<URL>(), new Scheduler(TestingUtils.defaultExecutionContext()), ExecutionGraph.class.getClassLoader(), new UnregisteredMetricsGroup());
    try {
        eg.attachJobGraph(jobVertices);
    } catch (JobException e) {
        e.printStackTrace();
        fail("Building ExecutionGraph failed: " + e.getMessage());
    }
    ExecutionJobVertex execSourceVertex = eg.getJobVertex(sourceVertex.getID());
    ExecutionJobVertex execMapVertex = eg.getJobVertex(mapVertex.getID());
    ExecutionJobVertex execSinkVertex = eg.getJobVertex(sinkVertex.getID());
    assertEquals(0, execSourceVertex.getInputs().size());
    assertEquals(1, execMapVertex.getInputs().size());
    assertEquals(4, execMapVertex.getParallelism());
    ExecutionVertex[] mapTaskVertices = execMapVertex.getTaskVertices();
    // verify that we have each parallel input partition exactly twice, i.e. that one source
    // sends to two unique mappers
    Map<Integer, Integer> mapInputPartitionCounts = new HashMap<>();
    for (ExecutionVertex mapTaskVertex : mapTaskVertices) {
        assertEquals(1, mapTaskVertex.getNumberOfInputs());
        assertEquals(1, mapTaskVertex.getInputEdges(0).length);
        ExecutionEdge inputEdge = mapTaskVertex.getInputEdges(0)[0];
        assertEquals(sourceVertex.getID(), inputEdge.getSource().getProducer().getJobvertexId());
        int inputPartition = inputEdge.getSource().getPartitionNumber();
        if (!mapInputPartitionCounts.containsKey(inputPartition)) {
            mapInputPartitionCounts.put(inputPartition, 1);
        } else {
            mapInputPartitionCounts.put(inputPartition, mapInputPartitionCounts.get(inputPartition) + 1);
        }
    }
    assertEquals(2, mapInputPartitionCounts.size());
    for (int count : mapInputPartitionCounts.values()) {
        assertEquals(2, count);
    }
    assertEquals(1, execSinkVertex.getInputs().size());
    assertEquals(2, execSinkVertex.getParallelism());
    ExecutionVertex[] sinkTaskVertices = execSinkVertex.getTaskVertices();
    // verify each sink instance has two inputs from the map and that each map subpartition
    // only occurs in one unique input edge
    Set<Integer> mapSubpartitions = new HashSet<>();
    for (ExecutionVertex sinkTaskVertex : sinkTaskVertices) {
        assertEquals(1, sinkTaskVertex.getNumberOfInputs());
        assertEquals(2, sinkTaskVertex.getInputEdges(0).length);
        ExecutionEdge inputEdge1 = sinkTaskVertex.getInputEdges(0)[0];
        ExecutionEdge inputEdge2 = sinkTaskVertex.getInputEdges(0)[1];
        assertEquals(mapVertex.getID(), inputEdge1.getSource().getProducer().getJobvertexId());
        assertEquals(mapVertex.getID(), inputEdge2.getSource().getProducer().getJobvertexId());
        int inputPartition1 = inputEdge1.getSource().getPartitionNumber();
        assertFalse(mapSubpartitions.contains(inputPartition1));
        mapSubpartitions.add(inputPartition1);
        int inputPartition2 = inputEdge2.getSource().getPartitionNumber();
        assertFalse(mapSubpartitions.contains(inputPartition2));
        mapSubpartitions.add(inputPartition2);
    }
    assertEquals(4, mapSubpartitions.size());
}
Also used : UnregisteredMetricsGroup(org.apache.flink.metrics.groups.UnregisteredMetricsGroup) Configuration(org.apache.flink.configuration.Configuration) HashMap(java.util.HashMap) Scheduler(org.apache.flink.runtime.jobmanager.scheduler.Scheduler) ExecutionConfig(org.apache.flink.api.common.ExecutionConfig) URL(java.net.URL) ExecutionVertex(org.apache.flink.runtime.executiongraph.ExecutionVertex) JobException(org.apache.flink.runtime.JobException) BlobKey(org.apache.flink.runtime.blob.BlobKey) ExecutionJobVertex(org.apache.flink.runtime.executiongraph.ExecutionJobVertex) ExecutionEdge(org.apache.flink.runtime.executiongraph.ExecutionEdge) ParallelSourceFunction(org.apache.flink.streaming.api.functions.source.ParallelSourceFunction) HashSet(java.util.HashSet) NoRestartStrategy(org.apache.flink.runtime.executiongraph.restart.NoRestartStrategy) JobException(org.apache.flink.runtime.JobException) JobGraph(org.apache.flink.runtime.jobgraph.JobGraph) JobVertex(org.apache.flink.runtime.jobgraph.JobVertex) ExecutionJobVertex(org.apache.flink.runtime.executiongraph.ExecutionJobVertex) Tuple2(org.apache.flink.api.java.tuple.Tuple2) ExecutionGraph(org.apache.flink.runtime.executiongraph.ExecutionGraph) StreamExecutionEnvironment(org.apache.flink.streaming.api.environment.StreamExecutionEnvironment) JobID(org.apache.flink.api.common.JobID) Test(org.junit.Test)

Aggregations

NoRestartStrategy (org.apache.flink.runtime.executiongraph.restart.NoRestartStrategy)27 ExecutionConfig (org.apache.flink.api.common.ExecutionConfig)25 JobVertex (org.apache.flink.runtime.jobgraph.JobVertex)25 Scheduler (org.apache.flink.runtime.jobmanager.scheduler.Scheduler)24 JobID (org.apache.flink.api.common.JobID)19 Test (org.junit.Test)19 Configuration (org.apache.flink.configuration.Configuration)18 ArrayList (java.util.ArrayList)17 JobException (org.apache.flink.runtime.JobException)16 UnregisteredMetricsGroup (org.apache.flink.metrics.groups.UnregisteredMetricsGroup)4 IntermediateDataSet (org.apache.flink.runtime.jobgraph.IntermediateDataSet)4 JobVertexID (org.apache.flink.runtime.jobgraph.JobVertexID)4 ActorTaskManagerGateway (org.apache.flink.runtime.jobmanager.slots.ActorTaskManagerGateway)4 JobGraph (org.apache.flink.runtime.jobgraph.JobGraph)3 DirectScheduledExecutorService (org.apache.flink.runtime.testutils.DirectScheduledExecutorService)3 HashMap (java.util.HashMap)2 HashSet (java.util.HashSet)2 Tuple2 (org.apache.flink.api.java.tuple.Tuple2)2 ExecutionGraph (org.apache.flink.runtime.executiongraph.ExecutionGraph)2 ExecutionJobVertex (org.apache.flink.runtime.executiongraph.ExecutionJobVertex)2