Examples with SchedulingExecutionVertex org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex used on opensource projects

Example 1 with SchedulingExecutionVertex

use of org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex in project flink by apache.

the class RestartPipelinedRegionFailoverStrategy method getTasksNeedingRestart.

// ------------------------------------------------------------------------
// task failure handling
// ------------------------------------------------------------------------
/**
 * Returns a set of IDs corresponding to the set of vertices that should be restarted. In this
 * strategy, all task vertices in 'involved' regions are proposed to be restarted. The
 * 'involved' regions are calculated with rules below: 1. The region containing the failed task
 * is always involved 2. If an input result partition of an involved region is not available,
 * i.e. Missing or Corrupted, the region containing the partition producer task is involved 3.
 * If a region is involved, all of its consumer regions are involved
 *
 * @param executionVertexId ID of the failed task
 * @param cause cause of the failure
 * @return set of IDs of vertices to restart
 */
@Override
public Set<ExecutionVertexID> getTasksNeedingRestart(ExecutionVertexID executionVertexId, Throwable cause) {
    LOG.info("Calculating tasks to restart to recover the failed task {}.", executionVertexId);
    final SchedulingPipelinedRegion failedRegion = topology.getPipelinedRegionOfVertex(executionVertexId);
    if (failedRegion == null) {
        // TODO: show the task name in the log
        throw new IllegalStateException("Can not find the failover region for task " + executionVertexId, cause);
    }
    // if the failure cause is data consumption error, mark the corresponding data partition to
    // be failed,
    // so that the failover process will try to recover it
    Optional<PartitionException> dataConsumptionException = ExceptionUtils.findThrowable(cause, PartitionException.class);
    if (dataConsumptionException.isPresent()) {
        resultPartitionAvailabilityChecker.markResultPartitionFailed(dataConsumptionException.get().getPartitionId().getPartitionId());
    }
    // calculate the tasks to restart based on the result of regions to restart
    Set<ExecutionVertexID> tasksToRestart = new HashSet<>();
    for (SchedulingPipelinedRegion region : getRegionsToRestart(failedRegion)) {
        for (SchedulingExecutionVertex vertex : region.getVertices()) {
            // we do not need to restart tasks which are already in the initial state
            if (vertex.getState() != ExecutionState.CREATED) {
                tasksToRestart.add(vertex.getId());
            }
        }
    }
    // the previous failed partition will be recovered. remove its failed state from the checker
    if (dataConsumptionException.isPresent()) {
        resultPartitionAvailabilityChecker.removeResultPartitionFromFailedState(dataConsumptionException.get().getPartitionId().getPartitionId());
    }
    LOG.info("{} tasks should be restarted to recover the failed task {}. ", tasksToRestart.size(), executionVertexId);
    return tasksToRestart;
}

Example 2 with SchedulingExecutionVertex

use of org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex in project flink by apache.

the class RestartPipelinedRegionFailoverStrategy method getConsumerVerticesToVisit.

private Iterable<ExecutionVertexID> getConsumerVerticesToVisit(SchedulingPipelinedRegion regionToRestart, Set<ConsumerVertexGroup> visitedConsumerVertexGroups) {
    final List<ConsumerVertexGroup> consumerVertexGroupsToVisit = new ArrayList<>();
    for (SchedulingExecutionVertex vertex : regionToRestart.getVertices()) {
        for (SchedulingResultPartition producedPartition : vertex.getProducedResults()) {
            final Optional<ConsumerVertexGroup> consumerVertexGroup = producedPartition.getConsumerVertexGroup();
            if (consumerVertexGroup.isPresent() && !visitedConsumerVertexGroups.contains(consumerVertexGroup.get())) {
                visitedConsumerVertexGroups.add(consumerVertexGroup.get());
                consumerVertexGroupsToVisit.add(consumerVertexGroup.get());
            }
        }
    }
    return IterableUtils.flatMap(consumerVertexGroupsToVisit, Function.identity());
}

Example 3 with SchedulingExecutionVertex

use of org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex in project flink by apache.

the class RestartPipelinedRegionFailoverStrategy method getConsumedPartitionsToVisit.

private Iterable<IntermediateResultPartitionID> getConsumedPartitionsToVisit(SchedulingPipelinedRegion regionToRestart, Set<ConsumedPartitionGroup> visitedConsumedResultGroups) {
    final List<ConsumedPartitionGroup> consumedPartitionGroupsToVisit = new ArrayList<>();
    for (SchedulingExecutionVertex vertex : regionToRestart.getVertices()) {
        for (ConsumedPartitionGroup consumedPartitionGroup : vertex.getConsumedPartitionGroups()) {
            if (!visitedConsumedResultGroups.contains(consumedPartitionGroup)) {
                visitedConsumedResultGroups.add(consumedPartitionGroup);
                consumedPartitionGroupsToVisit.add(consumedPartitionGroup);
            }
        }
    }
    return IterableUtils.flatMap(consumedPartitionGroupsToVisit, Function.identity());
}

Example 4 with SchedulingExecutionVertex

use of org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex in project flink by apache.

the class SchedulingPipelinedRegionComputeUtil method buildOutEdgesDesc.

private static List<List<Integer>> buildOutEdgesDesc(final Map<SchedulingExecutionVertex, Set<SchedulingExecutionVertex>> vertexToRegion, final List<Set<SchedulingExecutionVertex>> regionList, final Function<ExecutionVertexID, ? extends SchedulingExecutionVertex> executionVertexRetriever) {
    final Map<Set<SchedulingExecutionVertex>, Integer> regionIndices = new IdentityHashMap<>();
    for (int i = 0; i < regionList.size(); i++) {
        regionIndices.put(regionList.get(i), i);
    }
    final List<List<Integer>> outEdges = new ArrayList<>(regionList.size());
    for (Set<SchedulingExecutionVertex> currentRegion : regionList) {
        final List<Integer> currentRegionOutEdges = new ArrayList<>();
        for (SchedulingExecutionVertex vertex : currentRegion) {
            for (SchedulingResultPartition producedResult : vertex.getProducedResults()) {
                if (!producedResult.getResultType().isReconnectable()) {
                    continue;
                }
                final Optional<ConsumerVertexGroup> consumerVertexGroup = producedResult.getConsumerVertexGroup();
                if (!consumerVertexGroup.isPresent()) {
                    continue;
                }
                for (ExecutionVertexID consumerVertexId : consumerVertexGroup.get()) {
                    SchedulingExecutionVertex consumerVertex = executionVertexRetriever.apply(consumerVertexId);
                    // regions and cannot be merged
                    if (!vertexToRegion.containsKey(consumerVertex)) {
                        break;
                    }
                    if (!currentRegion.contains(consumerVertex)) {
                        currentRegionOutEdges.add(regionIndices.get(vertexToRegion.get(consumerVertex)));
                    }
                }
            }
        }
        outEdges.add(currentRegionOutEdges);
    }
    return outEdges;
}

Example 5 with SchedulingExecutionVertex

use of org.apache.flink.runtime.scheduler.strategy.SchedulingExecutionVertex in project flink by apache.

the class SchedulingPipelinedRegionComputeUtilTest method testPipelinedApproximateDifferentRegions.

@Test
public void testPipelinedApproximateDifferentRegions() {
    TestingSchedulingTopology topology = new TestingSchedulingTopology();
    TestingSchedulingExecutionVertex v1 = topology.newExecutionVertex();
    TestingSchedulingExecutionVertex v2 = topology.newExecutionVertex();
    TestingSchedulingExecutionVertex v3 = topology.newExecutionVertex();
    TestingSchedulingExecutionVertex v4 = topology.newExecutionVertex();
    topology.connect(v1, v2, ResultPartitionType.PIPELINED_APPROXIMATE).connect(v1, v3, ResultPartitionType.PIPELINED_APPROXIMATE).connect(v2, v4, ResultPartitionType.PIPELINED_APPROXIMATE).connect(v3, v4, ResultPartitionType.PIPELINED_APPROXIMATE);
    Map<ExecutionVertexID, Set<SchedulingExecutionVertex>> pipelinedRegionByVertex = computePipelinedRegionByVertex(topology);
    Set<SchedulingExecutionVertex> r1 = pipelinedRegionByVertex.get(v1.getId());
    Set<SchedulingExecutionVertex> r2 = pipelinedRegionByVertex.get(v2.getId());
    Set<SchedulingExecutionVertex> r3 = pipelinedRegionByVertex.get(v3.getId());
    Set<SchedulingExecutionVertex> r4 = pipelinedRegionByVertex.get(v4.getId());
    assertDistinctRegions(r1, r2, r3, r4);
}