Examples with StageId com.facebook.presto.execution.StageId used on opensource projects

Example 1 with StageId

use of com.facebook.presto.execution.StageId in project presto by prestodb.

the class PrestoSparkQueryExecutionFactory method createStageInfo.

private static StageInfo createStageInfo(QueryId queryId, SubPlan plan, ListMultimap<PlanFragmentId, TaskInfo> taskInfoMap) {
    PlanFragmentId planFragmentId = plan.getFragment().getId();
    StageId stageId = new StageId(queryId, planFragmentId.getId());
    List<TaskInfo> taskInfos = taskInfoMap.get(planFragmentId);
    long peakUserMemoryReservationInBytes = 0;
    long peakNodeTotalMemoryReservationInBytes = 0;
    for (TaskInfo taskInfo : taskInfos) {
        long taskPeakUserMemoryInBytes = taskInfo.getStats().getUserMemoryReservationInBytes();
        peakUserMemoryReservationInBytes += taskPeakUserMemoryInBytes;
        peakNodeTotalMemoryReservationInBytes = max(taskInfo.getStats().getPeakNodeTotalMemoryInBytes(), peakNodeTotalMemoryReservationInBytes);
    }
    StageExecutionInfo stageExecutionInfo = StageExecutionInfo.create(new StageExecutionId(stageId, 0), // TODO: figure out a way to know what exactly stage has caused a failure
    StageExecutionState.FINISHED, Optional.empty(), taskInfos, DateTime.now(), new Distribution().snapshot(), succinctBytes(peakUserMemoryReservationInBytes), succinctBytes(peakNodeTotalMemoryReservationInBytes), 1, 1);
    return new StageInfo(stageId, URI.create("http://fake.invalid/stage/" + stageId), Optional.of(plan.getFragment()), stageExecutionInfo, ImmutableList.of(), plan.getChildren().stream().map(child -> createStageInfo(queryId, child, taskInfoMap)).collect(toImmutableList()), false);
}

Example 2 with StageId

use of com.facebook.presto.execution.StageId in project presto by prestodb.

the class PrestoSparkTaskExecutorFactory method doCreate.

public <T extends PrestoSparkTaskOutput> IPrestoSparkTaskExecutor<T> doCreate(int partitionId, int attemptNumber, SerializedPrestoSparkTaskDescriptor serializedTaskDescriptor, Iterator<SerializedPrestoSparkTaskSource> serializedTaskSources, PrestoSparkTaskInputs inputs, CollectionAccumulator<SerializedTaskInfo> taskInfoCollector, CollectionAccumulator<PrestoSparkShuffleStats> shuffleStatsCollector, Class<T> outputType) {
    PrestoSparkTaskDescriptor taskDescriptor = taskDescriptorJsonCodec.fromJson(serializedTaskDescriptor.getBytes());
    ImmutableMap.Builder<String, TokenAuthenticator> extraAuthenticators = ImmutableMap.builder();
    authenticatorProviders.forEach(provider -> extraAuthenticators.putAll(provider.getTokenAuthenticators()));
    Session session = taskDescriptor.getSession().toSession(sessionPropertyManager, taskDescriptor.getExtraCredentials(), extraAuthenticators.build());
    PlanFragment fragment = taskDescriptor.getFragment();
    StageId stageId = new StageId(session.getQueryId(), fragment.getId().getId());
    // Clear the cache if the cache does not have broadcast table for current stageId.
    // We will only cache 1 HT at any time. If the stageId changes, we will drop the old cached HT
    prestoSparkBroadcastTableCacheManager.removeCachedTablesForStagesOtherThan(stageId);
    // TODO: include attemptId in taskId
    TaskId taskId = new TaskId(new StageExecutionId(stageId, 0), partitionId);
    List<TaskSource> taskSources = getTaskSources(serializedTaskSources);
    log.info("Task [%s] received %d splits.", taskId, taskSources.stream().mapToInt(taskSource -> taskSource.getSplits().size()).sum());
    OptionalLong totalSplitSize = computeAllSplitsSize(taskSources);
    if (totalSplitSize.isPresent()) {
        log.info("Total split size: %s bytes.", totalSplitSize.getAsLong());
    }
    // TODO: Remove this once we can display the plan on Spark UI.
    log.info(PlanPrinter.textPlanFragment(fragment, functionAndTypeManager, session, true));
    DataSize maxUserMemory = new DataSize(min(nodeMemoryConfig.getMaxQueryMemoryPerNode().toBytes(), getQueryMaxMemoryPerNode(session).toBytes()), BYTE);
    DataSize maxTotalMemory = new DataSize(min(nodeMemoryConfig.getMaxQueryTotalMemoryPerNode().toBytes(), getQueryMaxTotalMemoryPerNode(session).toBytes()), BYTE);
    DataSize maxBroadcastMemory = getSparkBroadcastJoinMaxMemoryOverride(session);
    if (maxBroadcastMemory == null) {
        maxBroadcastMemory = new DataSize(min(nodeMemoryConfig.getMaxQueryBroadcastMemory().toBytes(), getQueryMaxBroadcastMemory(session).toBytes()), BYTE);
    }
    MemoryPool memoryPool = new MemoryPool(new MemoryPoolId("spark-executor-memory-pool"), maxTotalMemory);
    SpillSpaceTracker spillSpaceTracker = new SpillSpaceTracker(maxQuerySpillPerNode);
    QueryContext queryContext = new QueryContext(session.getQueryId(), maxUserMemory, maxTotalMemory, maxBroadcastMemory, maxRevocableMemory, memoryPool, new TestingGcMonitor(), notificationExecutor, yieldExecutor, maxQuerySpillPerNode, spillSpaceTracker, memoryReservationSummaryJsonCodec);
    queryContext.setVerboseExceededMemoryLimitErrorsEnabled(isVerboseExceededMemoryLimitErrorsEnabled(session));
    queryContext.setHeapDumpOnExceededMemoryLimitEnabled(isHeapDumpOnExceededMemoryLimitEnabled(session));
    String heapDumpFilePath = Paths.get(getHeapDumpFileDirectory(session), format("%s_%s.hprof", session.getQueryId().getId(), stageId.getId())).toString();
    queryContext.setHeapDumpFilePath(heapDumpFilePath);
    TaskStateMachine taskStateMachine = new TaskStateMachine(taskId, notificationExecutor);
    TaskContext taskContext = queryContext.addTaskContext(taskStateMachine, session, // Plan has to be retained only if verbose memory exceeded errors are requested
    isVerboseExceededMemoryLimitErrorsEnabled(session) ? Optional.of(fragment.getRoot()) : Optional.empty(), perOperatorCpuTimerEnabled, cpuTimerEnabled, perOperatorAllocationTrackingEnabled, allocationTrackingEnabled, false);
    final double memoryRevokingThreshold = getMemoryRevokingThreshold(session);
    final double memoryRevokingTarget = getMemoryRevokingTarget(session);
    checkArgument(memoryRevokingTarget <= memoryRevokingThreshold, "memoryRevokingTarget should be less than or equal memoryRevokingThreshold, but got %s and %s respectively", memoryRevokingTarget, memoryRevokingThreshold);
    if (isSpillEnabled(session)) {
        memoryPool.addListener((pool, queryId, totalMemoryReservationBytes) -> {
            if (totalMemoryReservationBytes > queryContext.getPeakNodeTotalMemory()) {
                queryContext.setPeakNodeTotalMemory(totalMemoryReservationBytes);
            }
            if (totalMemoryReservationBytes > pool.getMaxBytes() * memoryRevokingThreshold && memoryRevokeRequestInProgress.compareAndSet(false, true)) {
                memoryRevocationExecutor.execute(() -> {
                    try {
                        AtomicLong remainingBytesToRevoke = new AtomicLong(totalMemoryReservationBytes - (long) (memoryRevokingTarget * pool.getMaxBytes()));
                        remainingBytesToRevoke.addAndGet(-MemoryRevokingSchedulerUtils.getMemoryAlreadyBeingRevoked(ImmutableList.of(taskContext), remainingBytesToRevoke.get()));
                        taskContext.accept(new VoidTraversingQueryContextVisitor<AtomicLong>() {

                            @Override
                            public Void visitOperatorContext(OperatorContext operatorContext, AtomicLong remainingBytesToRevoke) {
                                if (remainingBytesToRevoke.get() > 0) {
                                    long revokedBytes = operatorContext.requestMemoryRevoking();
                                    if (revokedBytes > 0) {
                                        memoryRevokePending.set(true);
                                        remainingBytesToRevoke.addAndGet(-revokedBytes);
                                    }
                                }
                                return null;
                            }
                        }, remainingBytesToRevoke);
                        memoryRevokeRequestInProgress.set(false);
                    } catch (Exception e) {
                        log.error(e, "Error requesting memory revoking");
                    }
                });
            }
            // Get the latest memory reservation info since it might have changed due to revoke
            long totalReservedMemory = pool.getQueryMemoryReservation(queryId) + pool.getQueryRevocableMemoryReservation(queryId);
            // If total memory usage is over maxTotalMemory and memory revoke request is not pending, fail the query with EXCEEDED_MEMORY_LIMIT error
            if (totalReservedMemory > maxTotalMemory.toBytes() && !memoryRevokeRequestInProgress.get() && !isMemoryRevokePending(taskContext)) {
                throw exceededLocalTotalMemoryLimit(maxTotalMemory, queryContext.getAdditionalFailureInfo(totalReservedMemory, 0) + format("Total reserved memory: %s, Total revocable memory: %s", succinctBytes(pool.getQueryMemoryReservation(queryId)), succinctBytes(pool.getQueryRevocableMemoryReservation(queryId))), isHeapDumpOnExceededMemoryLimitEnabled(session), Optional.ofNullable(heapDumpFilePath));
            }
        });
    }
    ImmutableMap.Builder<PlanNodeId, List<PrestoSparkShuffleInput>> shuffleInputs = ImmutableMap.builder();
    ImmutableMap.Builder<PlanNodeId, List<java.util.Iterator<PrestoSparkSerializedPage>>> pageInputs = ImmutableMap.builder();
    ImmutableMap.Builder<PlanNodeId, List<?>> broadcastInputs = ImmutableMap.builder();
    for (RemoteSourceNode remoteSource : fragment.getRemoteSourceNodes()) {
        List<PrestoSparkShuffleInput> remoteSourceRowInputs = new ArrayList<>();
        List<java.util.Iterator<PrestoSparkSerializedPage>> remoteSourcePageInputs = new ArrayList<>();
        List<List<?>> broadcastInputsList = new ArrayList<>();
        for (PlanFragmentId sourceFragmentId : remoteSource.getSourceFragmentIds()) {
            Iterator<Tuple2<MutablePartitionId, PrestoSparkMutableRow>> shuffleInput = inputs.getShuffleInputs().get(sourceFragmentId.toString());
            Broadcast<?> broadcastInput = inputs.getBroadcastInputs().get(sourceFragmentId.toString());
            List<PrestoSparkSerializedPage> inMemoryInput = inputs.getInMemoryInputs().get(sourceFragmentId.toString());
            if (shuffleInput != null) {
                checkArgument(broadcastInput == null, "single remote source is not expected to accept different kind of inputs");
                checkArgument(inMemoryInput == null, "single remote source is not expected to accept different kind of inputs");
                remoteSourceRowInputs.add(new PrestoSparkShuffleInput(sourceFragmentId.getId(), shuffleInput));
                continue;
            }
            if (broadcastInput != null) {
                checkArgument(inMemoryInput == null, "single remote source is not expected to accept different kind of inputs");
                // TODO: Enable NullifyingIterator once migrated to one task per JVM model
                // NullifyingIterator removes element from the list upon return
                // This allows GC to gradually reclaim memory
                // remoteSourcePageInputs.add(getNullifyingIterator(broadcastInput.value()));
                broadcastInputsList.add((List<?>) broadcastInput.value());
                continue;
            }
            if (inMemoryInput != null) {
                // for inmemory inputs pages can be released incrementally to save memory
                remoteSourcePageInputs.add(getNullifyingIterator(inMemoryInput));
                continue;
            }
            throw new IllegalArgumentException("Input not found for sourceFragmentId: " + sourceFragmentId);
        }
        if (!remoteSourceRowInputs.isEmpty()) {
            shuffleInputs.put(remoteSource.getId(), remoteSourceRowInputs);
        }
        if (!remoteSourcePageInputs.isEmpty()) {
            pageInputs.put(remoteSource.getId(), remoteSourcePageInputs);
        }
        if (!broadcastInputsList.isEmpty()) {
            broadcastInputs.put(remoteSource.getId(), broadcastInputsList);
        }
    }
    OutputBufferMemoryManager memoryManager = new OutputBufferMemoryManager(sinkMaxBufferSize.toBytes(), () -> queryContext.getTaskContextByTaskId(taskId).localSystemMemoryContext(), notificationExecutor);
    Optional<OutputPartitioning> preDeterminedPartition = Optional.empty();
    if (fragment.getPartitioningScheme().getPartitioning().getHandle().equals(FIXED_ARBITRARY_DISTRIBUTION)) {
        int partitionCount = getHashPartitionCount(session);
        preDeterminedPartition = Optional.of(new OutputPartitioning(new PreDeterminedPartitionFunction(partitionId % partitionCount, partitionCount), ImmutableList.of(), ImmutableList.of(), false, OptionalInt.empty()));
    }
    TempDataOperationContext tempDataOperationContext = new TempDataOperationContext(session.getSource(), session.getQueryId().getId(), session.getClientInfo(), Optional.of(session.getClientTags()), session.getIdentity());
    TempStorage tempStorage = tempStorageManager.getTempStorage(storageBasedBroadcastJoinStorage);
    Output<T> output = configureOutput(outputType, blockEncodingManager, memoryManager, getShuffleOutputTargetAverageRowSize(session), preDeterminedPartition, tempStorage, tempDataOperationContext, getStorageBasedBroadcastJoinWriteBufferSize(session));
    PrestoSparkOutputBuffer<?> outputBuffer = output.getOutputBuffer();
    LocalExecutionPlan localExecutionPlan = localExecutionPlanner.plan(taskContext, fragment.getRoot(), fragment.getPartitioningScheme(), fragment.getStageExecutionDescriptor(), fragment.getTableScanSchedulingOrder(), output.getOutputFactory(), new PrestoSparkRemoteSourceFactory(blockEncodingManager, shuffleInputs.build(), pageInputs.build(), broadcastInputs.build(), partitionId, shuffleStatsCollector, tempStorage, tempDataOperationContext, prestoSparkBroadcastTableCacheManager, stageId), taskDescriptor.getTableWriteInfo(), true);
    taskStateMachine.addStateChangeListener(state -> {
        if (state.isDone()) {
            outputBuffer.setNoMoreRows();
        }
    });
    PrestoSparkTaskExecution taskExecution = new PrestoSparkTaskExecution(taskStateMachine, taskContext, localExecutionPlan, taskExecutor, splitMonitor, notificationExecutor, memoryUpdateExecutor);
    taskExecution.start(taskSources);
    return new PrestoSparkTaskExecutor<>(taskContext, taskStateMachine, output.getOutputSupplier(), taskInfoCodec, taskInfoCollector, shuffleStatsCollector, executionExceptionFactory, output.getOutputBufferType(), outputBuffer, tempStorage, tempDataOperationContext);
}

Example 3 with StageId

use of com.facebook.presto.execution.StageId in project presto by prestodb.

the class LegacySqlQueryScheduler method schedule.

private void schedule() {
    if (!scheduling.compareAndSet(false, true)) {
        // still scheduling the previous batch of stages
        return;
    }
    List<StageExecutionAndScheduler> scheduledStageExecutions = new ArrayList<>();
    try (SetThreadName ignored = new SetThreadName("Query-%s", queryStateMachine.getQueryId())) {
        Set<StageId> completedStages = new HashSet<>();
        List<ExecutionSchedule> sectionExecutionSchedules = new LinkedList<>();
        while (!Thread.currentThread().isInterrupted()) {
            // remove finished section
            sectionExecutionSchedules.removeIf(ExecutionSchedule::isFinished);
            // try to pull more section that are ready to be run
            List<StreamingPlanSection> sectionsReadyForExecution = getSectionsReadyForExecution();
            // all finished
            if (sectionsReadyForExecution.isEmpty() && sectionExecutionSchedules.isEmpty()) {
                break;
            }
            List<List<StageExecutionAndScheduler>> sectionStageExecutions = getStageExecutions(sectionsReadyForExecution);
            sectionStageExecutions.forEach(scheduledStageExecutions::addAll);
            sectionStageExecutions.stream().map(executionInfos -> executionInfos.stream().collect(toImmutableList())).map(executionPolicy::createExecutionSchedule).forEach(sectionExecutionSchedules::add);
            while (sectionExecutionSchedules.stream().noneMatch(ExecutionSchedule::isFinished)) {
                List<ListenableFuture<?>> blockedStages = new ArrayList<>();
                List<StageExecutionAndScheduler> executionsToSchedule = sectionExecutionSchedules.stream().flatMap(schedule -> schedule.getStagesToSchedule().stream()).collect(toImmutableList());
                for (StageExecutionAndScheduler stageExecutionAndScheduler : executionsToSchedule) {
                    SqlStageExecution stageExecution = stageExecutionAndScheduler.getStageExecution();
                    StageId stageId = stageExecution.getStageExecutionId().getStageId();
                    stageExecution.beginScheduling();
                    // perform some scheduling work
                    ScheduleResult result = stageExecutionAndScheduler.getStageScheduler().schedule();
                    // Track leaf tasks if partial results are enabled
                    if (isPartialResultsEnabled(session) && stageExecutionAndScheduler.getStageExecution().getFragment().isLeaf()) {
                        for (RemoteTask task : result.getNewTasks()) {
                            partialResultQueryTaskTracker.trackTask(task);
                            task.addFinalTaskInfoListener(partialResultQueryTaskTracker::recordTaskFinish);
                        }
                    }
                    // modify parent and children based on the results of the scheduling
                    if (result.isFinished()) {
                        stageExecution.schedulingComplete();
                    } else if (!result.getBlocked().isDone()) {
                        blockedStages.add(result.getBlocked());
                    }
                    stageExecutionAndScheduler.getStageLinkage().processScheduleResults(stageExecution.getState(), result.getNewTasks());
                    schedulerStats.getSplitsScheduledPerIteration().add(result.getSplitsScheduled());
                    if (result.getBlockedReason().isPresent()) {
                        switch(result.getBlockedReason().get()) {
                            case WRITER_SCALING:
                                // no-op
                                break;
                            case WAITING_FOR_SOURCE:
                                schedulerStats.getWaitingForSource().update(1);
                                break;
                            case SPLIT_QUEUES_FULL:
                                schedulerStats.getSplitQueuesFull().update(1);
                                break;
                            case MIXED_SPLIT_QUEUES_FULL_AND_WAITING_FOR_SOURCE:
                                schedulerStats.getMixedSplitQueuesFullAndWaitingForSource().update(1);
                                break;
                            case NO_ACTIVE_DRIVER_GROUP:
                                schedulerStats.getNoActiveDriverGroup().update(1);
                                break;
                            default:
                                throw new UnsupportedOperationException("Unknown blocked reason: " + result.getBlockedReason().get());
                        }
                    }
                }
                // make sure to update stage linkage at least once per loop to catch async state changes (e.g., partial cancel)
                boolean stageFinishedExecution = false;
                for (StageExecutionAndScheduler stageExecutionInfo : scheduledStageExecutions) {
                    SqlStageExecution stageExecution = stageExecutionInfo.getStageExecution();
                    StageId stageId = stageExecution.getStageExecutionId().getStageId();
                    if (!completedStages.contains(stageId) && stageExecution.getState().isDone()) {
                        stageExecutionInfo.getStageLinkage().processScheduleResults(stageExecution.getState(), ImmutableSet.of());
                        completedStages.add(stageId);
                        stageFinishedExecution = true;
                    }
                }
                // if any stage has just finished execution try to pull more sections for scheduling
                if (stageFinishedExecution) {
                    break;
                }
                // wait for a state change and then schedule again
                if (!blockedStages.isEmpty()) {
                    try (TimeStat.BlockTimer timer = schedulerStats.getSleepTime().time()) {
                        tryGetFutureValue(whenAnyComplete(blockedStages), 1, SECONDS);
                    }
                    for (ListenableFuture<?> blockedStage : blockedStages) {
                        blockedStage.cancel(true);
                    }
                }
            }
        }
        for (StageExecutionAndScheduler stageExecutionInfo : scheduledStageExecutions) {
            StageExecutionState state = stageExecutionInfo.getStageExecution().getState();
            if (state != SCHEDULED && state != RUNNING && !state.isDone()) {
                throw new PrestoException(GENERIC_INTERNAL_ERROR, format("Scheduling is complete, but stage execution %s is in state %s", stageExecutionInfo.getStageExecution().getStageExecutionId(), state));
            }
        }
        scheduling.set(false);
        // Inform the tracker that task scheduling has completed
        partialResultQueryTaskTracker.completeTaskScheduling();
        if (!getSectionsReadyForExecution().isEmpty()) {
            startScheduling();
        }
    } catch (Throwable t) {
        scheduling.set(false);
        queryStateMachine.transitionToFailed(t);
        throw t;
    } finally {
        RuntimeException closeError = new RuntimeException();
        for (StageExecutionAndScheduler stageExecutionInfo : scheduledStageExecutions) {
            try {
                stageExecutionInfo.getStageScheduler().close();
            } catch (Throwable t) {
                queryStateMachine.transitionToFailed(t);
                // Self-suppression not permitted
                if (closeError != t) {
                    closeError.addSuppressed(t);
                }
            }
        }
        if (closeError.getSuppressed().length > 0) {
            throw closeError;
        }
    }
}

Example 4 with StageId

use of com.facebook.presto.execution.StageId in project presto by prestodb.

the class LegacySqlQueryScheduler method updateStageExecutions.

/**
 * Utility function that rebuild a StreamingPlanSection, re-create stageExecutionAndScheduler for each of its stage, and finally update the stageExecutions map.
 */
private void updateStageExecutions(StreamingPlanSection section, Map<PlanFragment, PlanFragment> oldToNewFragment) {
    StreamingPlanSection newSection = new StreamingPlanSection(rewriteStreamingSubPlan(section.getPlan(), oldToNewFragment), section.getChildren());
    PlanFragment sectionRootFragment = newSection.getPlan().getFragment();
    Optional<int[]> bucketToPartition;
    OutputBuffers outputBuffers;
    ExchangeLocationsConsumer locationsConsumer;
    if (isRootFragment(sectionRootFragment)) {
        bucketToPartition = Optional.of(new int[1]);
        outputBuffers = createInitialEmptyOutputBuffers(sectionRootFragment.getPartitioningScheme().getPartitioning().getHandle()).withBuffer(new OutputBufferId(0), BROADCAST_PARTITION_ID).withNoMoreBufferIds();
        OutputBufferId rootBufferId = getOnlyElement(outputBuffers.getBuffers().keySet());
        locationsConsumer = (fragmentId, tasks, noMoreExchangeLocations) -> updateQueryOutputLocations(queryStateMachine, rootBufferId, tasks, noMoreExchangeLocations);
    } else {
        bucketToPartition = Optional.empty();
        outputBuffers = createDiscardingOutputBuffers();
        locationsConsumer = (fragmentId, tasks, noMoreExchangeLocations) -> {
        };
    }
    SectionExecution sectionExecution = sectionExecutionFactory.createSectionExecutions(session, newSection, locationsConsumer, bucketToPartition, outputBuffers, summarizeTaskInfo, remoteTaskFactory, splitSourceFactory, 0);
    addStateChangeListeners(sectionExecution);
    Map<StageId, StageExecutionAndScheduler> updatedStageExecutions = sectionExecution.getSectionStages().stream().collect(toImmutableMap(execution -> execution.getStageExecution().getStageExecutionId().getStageId(), identity()));
    synchronized (this) {
        stageExecutions.putAll(updatedStageExecutions);
    }
}

Example 5 with StageId

use of com.facebook.presto.execution.StageId in project presto by prestodb.

the class SqlQueryScheduler method schedule.

private void schedule() {
    if (!scheduling.compareAndSet(false, true)) {
        // still scheduling the previous batch of stages
        return;
    }
    List<StageExecutionAndScheduler> scheduledStageExecutions = new ArrayList<>();
    try (SetThreadName ignored = new SetThreadName("Query-%s", queryStateMachine.getQueryId())) {
        Set<StageId> completedStages = new HashSet<>();
        List<ExecutionSchedule> executionSchedules = new LinkedList<>();
        while (!Thread.currentThread().isInterrupted()) {
            // remove finished section
            executionSchedules.removeIf(ExecutionSchedule::isFinished);
            // try to pull more section that are ready to be run
            List<StreamingPlanSection> sectionsReadyForExecution = getSectionsReadyForExecution();
            // all finished
            if (sectionsReadyForExecution.isEmpty() && executionSchedules.isEmpty()) {
                break;
            }
            // Apply runtime CBO on the ready sections before creating SectionExecutions.
            List<SectionExecution> sectionExecutions = createStageExecutions(sectionsReadyForExecution.stream().map(this::tryCostBasedOptimize).collect(toImmutableList()));
            if (queryStateMachine.isDone()) {
                sectionExecutions.forEach(SectionExecution::abort);
                break;
            }
            sectionExecutions.forEach(sectionExecution -> scheduledStageExecutions.addAll(sectionExecution.getSectionStages()));
            sectionExecutions.stream().map(SectionExecution::getSectionStages).map(executionPolicy::createExecutionSchedule).forEach(executionSchedules::add);
            while (!executionSchedules.isEmpty() && executionSchedules.stream().noneMatch(ExecutionSchedule::isFinished)) {
                List<ListenableFuture<?>> blockedStages = new ArrayList<>();
                List<StageExecutionAndScheduler> executionsToSchedule = executionSchedules.stream().flatMap(schedule -> schedule.getStagesToSchedule().stream()).collect(toImmutableList());
                for (StageExecutionAndScheduler executionAndScheduler : executionsToSchedule) {
                    executionAndScheduler.getStageExecution().beginScheduling();
                    // perform some scheduling work
                    ScheduleResult result = executionAndScheduler.getStageScheduler().schedule();
                    // Track leaf tasks if partial results are enabled
                    if (isPartialResultsEnabled(session) && executionAndScheduler.getStageExecution().getFragment().isLeaf()) {
                        for (RemoteTask task : result.getNewTasks()) {
                            partialResultQueryTaskTracker.trackTask(task);
                            task.addFinalTaskInfoListener(partialResultQueryTaskTracker::recordTaskFinish);
                        }
                    }
                    // modify parent and children based on the results of the scheduling
                    if (result.isFinished()) {
                        executionAndScheduler.getStageExecution().schedulingComplete();
                    } else if (!result.getBlocked().isDone()) {
                        blockedStages.add(result.getBlocked());
                    }
                    executionAndScheduler.getStageLinkage().processScheduleResults(executionAndScheduler.getStageExecution().getState(), result.getNewTasks());
                    schedulerStats.getSplitsScheduledPerIteration().add(result.getSplitsScheduled());
                    if (result.getBlockedReason().isPresent()) {
                        switch(result.getBlockedReason().get()) {
                            case WRITER_SCALING:
                                // no-op
                                break;
                            case WAITING_FOR_SOURCE:
                                schedulerStats.getWaitingForSource().update(1);
                                break;
                            case SPLIT_QUEUES_FULL:
                                schedulerStats.getSplitQueuesFull().update(1);
                                break;
                            case MIXED_SPLIT_QUEUES_FULL_AND_WAITING_FOR_SOURCE:
                                schedulerStats.getMixedSplitQueuesFullAndWaitingForSource().update(1);
                                break;
                            case NO_ACTIVE_DRIVER_GROUP:
                                schedulerStats.getNoActiveDriverGroup().update(1);
                                break;
                            default:
                                throw new UnsupportedOperationException("Unknown blocked reason: " + result.getBlockedReason().get());
                        }
                    }
                }
                // make sure to update stage linkage at least once per loop to catch async state changes (e.g., partial cancel)
                boolean stageFinishedExecution = false;
                for (StageExecutionAndScheduler stageExecutionAndScheduler : scheduledStageExecutions) {
                    SqlStageExecution stageExecution = stageExecutionAndScheduler.getStageExecution();
                    StageId stageId = stageExecution.getStageExecutionId().getStageId();
                    if (!completedStages.contains(stageId) && stageExecution.getState().isDone()) {
                        stageExecutionAndScheduler.getStageLinkage().processScheduleResults(stageExecution.getState(), ImmutableSet.of());
                        completedStages.add(stageId);
                        stageFinishedExecution = true;
                    }
                }
                // if any stage has just finished execution try to pull more sections for scheduling
                if (stageFinishedExecution) {
                    break;
                }
                // wait for a state change and then schedule again
                if (!blockedStages.isEmpty()) {
                    try (TimeStat.BlockTimer timer = schedulerStats.getSleepTime().time()) {
                        tryGetFutureValue(whenAnyComplete(blockedStages), 1, SECONDS);
                    }
                    for (ListenableFuture<?> blockedStage : blockedStages) {
                        blockedStage.cancel(true);
                    }
                }
            }
        }
        for (StageExecutionAndScheduler stageExecutionAndScheduler : scheduledStageExecutions) {
            StageExecutionState state = stageExecutionAndScheduler.getStageExecution().getState();
            if (state != SCHEDULED && state != RUNNING && !state.isDone()) {
                throw new PrestoException(GENERIC_INTERNAL_ERROR, format("Scheduling is complete, but stage execution %s is in state %s", stageExecutionAndScheduler.getStageExecution().getStageExecutionId(), state));
            }
        }
        scheduling.set(false);
        // Inform the tracker that task scheduling has completed
        partialResultQueryTaskTracker.completeTaskScheduling();
        if (!getSectionsReadyForExecution().isEmpty()) {
            startScheduling();
        }
    } catch (Throwable t) {
        scheduling.set(false);
        queryStateMachine.transitionToFailed(t);
        throw t;
    } finally {
        RuntimeException closeError = new RuntimeException();
        for (StageExecutionAndScheduler stageExecutionAndScheduler : scheduledStageExecutions) {
            try {
                stageExecutionAndScheduler.getStageScheduler().close();
            } catch (Throwable t) {
                queryStateMachine.transitionToFailed(t);
                // Self-suppression not permitted
                if (closeError != t) {
                    closeError.addSuppressed(t);
                }
            }
        }
        if (closeError.getSuppressed().length > 0) {
            throw closeError;
        }
    }
}