use of io.prestosql.execution.StageState in project hetu-core by openlookeng.
the class SqlQueryScheduler method schedule.
private void schedule() {
try (SetThreadName ignored = new SetThreadName("Query-%s", queryStateMachine.getQueryId())) {
Set<StageId> completedStages = new HashSet<>();
ExecutionSchedule executionSchedule = executionPolicy.createExecutionSchedule(stages.values());
while (!executionSchedule.isFinished()) {
List<ListenableFuture<?>> blockedStages = new ArrayList<>();
for (SqlStageExecution stage : executionSchedule.getStagesToSchedule()) {
if (isReuseTableScanEnabled(session) && !SqlStageExecution.getReuseTableScanMappingIdStatus(stage.getStateMachine())) {
continue;
}
stage.beginScheduling();
// and then let it schedule 10% of splits.
if (queryStateMachine.isThrottlingEnabled() && !canScheduleMoreSplits()) {
try {
SECONDS.sleep(THROTTLE_SLEEP_TIMER[currentTimerLevel]);
} catch (InterruptedException e) {
throw new PrestoException(GENERIC_INTERNAL_ERROR, "interrupted while sleeping");
}
currentTimerLevel = Math.min(currentTimerLevel + 1, THROTTLE_SLEEP_TIMER.length - 1);
stage.setThrottledSchedule(true);
} else {
stage.setThrottledSchedule(false);
currentTimerLevel = 0;
}
// perform some scheduling work
/* Get groupSize specification from the ResourceGroupManager */
int maxSplitGroupSize = getOptimalSmallSplitGroupSize();
ScheduleResult result = stageSchedulers.get(stage.getStageId()).schedule(maxSplitGroupSize);
// modify parent and children based on the results of the scheduling
if (result.isFinished()) {
stage.schedulingComplete();
} else if (!result.getBlocked().isDone()) {
blockedStages.add(result.getBlocked());
}
stageLinkages.get(stage.getStageId()).processScheduleResults(stage.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:
case NO_ACTIVE_DRIVER_GROUP:
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)
for (SqlStageExecution stage : stages.values()) {
if (!completedStages.contains(stage.getStageId()) && stage.getState().isDone()) {
stageLinkages.get(stage.getStageId()).processScheduleResults(stage.getState(), ImmutableSet.of());
completedStages.add(stage.getStageId());
}
}
// 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 (SqlStageExecution stage : stages.values()) {
StageState state = stage.getState();
// Snapshot: if state is resumable_failure, then state of stage and query will change soon again. Don't treat as an error.
if (state != SCHEDULED && state != RUNNING && !state.isDone() && state != RESUMABLE_FAILURE) {
throw new PrestoException(GENERIC_INTERNAL_ERROR, format("Scheduling is complete, but stage %s is in state %s", stage.getStageId(), state));
}
}
} catch (Throwable t) {
queryStateMachine.transitionToFailed(t);
throw t;
} finally {
RuntimeException closeError = new RuntimeException();
for (StageScheduler scheduler : stageSchedulers.values()) {
try {
// Snapshot: when trying to reschedule, then don't close the scheduler (and more importantly, split sources in it)
QueryState state = queryStateMachine.getQueryState();
if (state != QueryState.RESCHEDULING && state != QueryState.RESUMING) {
scheduler.close();
}
} catch (Throwable t) {
queryStateMachine.transitionToFailed(t);
// Self-suppression not permitted
if (closeError != t) {
closeError.addSuppressed(t);
}
}
}
// Snpashot: if resuming, notify the new scheduler so it can start scheduling new stages
schedulingFuture.set(null);
if (closeError.getSuppressed().length > 0) {
throw closeError;
}
}
}
use of io.prestosql.execution.StageState in project hetu-core by openlookeng.
the class SqlQueryScheduler method createStages.
private List<SqlStageExecution> createStages(ExchangeLocationsConsumer parent, AtomicInteger nextStageId, LocationFactory locationFactory, StageExecutionPlan plan, NodeScheduler nodeScheduler, RemoteTaskFactory remoteTaskFactory, Session session, int splitBatchSize, BiFunction<PartitioningHandle, Integer, NodePartitionMap> partitioningCache, NodePartitioningManager nodePartitioningManager, ExecutorService queryExecutor, ScheduledExecutorService schedulerExecutor, FailureDetector failureDetector, NodeTaskMap nodeTaskMap, ImmutableMap.Builder<StageId, StageScheduler> stageSchedulers, ImmutableMap.Builder<StageId, StageLinkage> stageLinkages, boolean isSnapshotEnabled, QuerySnapshotManager snapshotManager, Map<StageId, Integer> stageTaskCounts) {
ImmutableList.Builder<SqlStageExecution> localStages = ImmutableList.builder();
StageId stageId = new StageId(queryStateMachine.getQueryId(), nextStageId.getAndIncrement());
SqlStageExecution stageExecution = createSqlStageExecution(stageId, locationFactory.createStageLocation(stageId), plan.getFragment(), plan.getTables(), remoteTaskFactory, session, summarizeTaskInfo, nodeTaskMap, queryExecutor, failureDetector, schedulerStats, dynamicFilterService, snapshotManager);
localStages.add(stageExecution);
Optional<int[]> bucketToPartition;
PartitioningHandle partitioningHandle = plan.getFragment().getPartitioning();
boolean keepConsumerOnFeederNodes = !plan.getFragment().getFeederCTEId().isPresent() && plan.getFragment().getFeederCTEParentId().isPresent();
if (partitioningHandle.equals(SOURCE_DISTRIBUTION)) {
// nodes are selected dynamically based on the constraints of the splits and the system load
Entry<PlanNodeId, SplitSource> entry = Iterables.getOnlyElement(plan.getSplitSources().entrySet());
PlanNodeId planNodeId = entry.getKey();
SplitSource splitSource = entry.getValue();
CatalogName catalogName = splitSource.getCatalogName();
if (isInternalSystemConnector(catalogName)) {
catalogName = null;
}
NodeSelector nodeSelector = nodeScheduler.createNodeSelector(catalogName, keepConsumerOnFeederNodes, feederScheduledNodes);
if (isSnapshotEnabled) {
// When snapshot is enabled, then no task can be added after the query started running,
// otherwise assumptions about how many "input channels" may be broken.
nodeSelector.lockDownNodes();
}
SplitPlacementPolicy placementPolicy = new DynamicSplitPlacementPolicy(nodeSelector, stageExecution::getAllTasks);
checkArgument(!plan.getFragment().getStageExecutionDescriptor().isStageGroupedExecution());
stageSchedulers.put(stageId, newSourcePartitionedSchedulerAsStageScheduler(stageExecution, planNodeId, splitSource, placementPolicy, splitBatchSize, session, heuristicIndexerManager));
bucketToPartition = Optional.of(new int[1]);
} else if (partitioningHandle.equals(SCALED_WRITER_DISTRIBUTION)) {
bucketToPartition = Optional.of(new int[1]);
} else {
Map<PlanNodeId, SplitSource> splitSources = plan.getSplitSources();
if (!splitSources.isEmpty()) {
// contains local source
List<PlanNodeId> schedulingOrder = plan.getFragment().getPartitionedSources();
CatalogName catalogName = partitioningHandle.getConnectorId().orElseThrow(IllegalStateException::new);
List<ConnectorPartitionHandle> connectorPartitionHandles;
boolean groupedExecutionForStage = plan.getFragment().getStageExecutionDescriptor().isStageGroupedExecution();
if (groupedExecutionForStage) {
connectorPartitionHandles = nodePartitioningManager.listPartitionHandles(session, partitioningHandle);
checkState(!ImmutableList.of(NOT_PARTITIONED).equals(connectorPartitionHandles));
} else {
connectorPartitionHandles = ImmutableList.of(NOT_PARTITIONED);
}
BucketNodeMap bucketNodeMap;
List<InternalNode> stageNodeList;
if (plan.getFragment().getRemoteSourceNodes().stream().allMatch(node -> node.getExchangeType() == REPLICATE)) {
// no remote source
boolean dynamicLifespanSchedule = plan.getFragment().getStageExecutionDescriptor().isDynamicLifespanSchedule();
if (isSnapshotEnabled) {
NodeSelector nodeSelector = nodeScheduler.createNodeSelector(catalogName, keepConsumerOnFeederNodes, feederScheduledNodes);
int nodeCount;
if (stageTaskCounts != null) {
// Resuming: need to create same number of tasks as old stage.
nodeCount = stageTaskCounts.get(stageId);
} else {
// Scheduling: reserve some nodes for resuming
nodeCount = calculateTaskCount(nodeSelector.selectableNodeCount());
}
stageNodeList = new ArrayList<>(nodeSelector.selectRandomNodes(nodeCount));
checkCondition(stageNodeList.size() == nodeCount, NO_NODES_AVAILABLE, "Snapshot: not enough worker nodes to resume expected number of tasks: " + nodeCount);
// Make sure bucketNodeMap uses the same node list
bucketNodeMap = nodePartitioningManager.getBucketNodeMap(session, partitioningHandle, dynamicLifespanSchedule, stageNodeList);
} else {
bucketNodeMap = nodePartitioningManager.getBucketNodeMap(session, partitioningHandle, dynamicLifespanSchedule);
stageNodeList = new ArrayList<>(nodeScheduler.createNodeSelector(catalogName, keepConsumerOnFeederNodes, feederScheduledNodes).allNodes());
}
// verify execution is consistent with planner's decision on dynamic lifespan schedule
verify(bucketNodeMap.isDynamic() == dynamicLifespanSchedule);
Collections.shuffle(stageNodeList);
bucketToPartition = Optional.empty();
} else {
// cannot use dynamic lifespan schedule
verify(!plan.getFragment().getStageExecutionDescriptor().isDynamicLifespanSchedule());
// remote source requires nodePartitionMap
NodePartitionMap nodePartitionMap = partitioningCache.apply(plan.getFragment().getPartitioning(), stageTaskCounts == null ? null : stageTaskCounts.get(stageId));
if (groupedExecutionForStage) {
checkState(connectorPartitionHandles.size() == nodePartitionMap.getBucketToPartition().length);
}
stageNodeList = nodePartitionMap.getPartitionToNode();
bucketNodeMap = nodePartitionMap.asBucketNodeMap();
bucketToPartition = Optional.of(nodePartitionMap.getBucketToPartition());
}
stageSchedulers.put(stageId, new FixedSourcePartitionedScheduler(stageExecution, splitSources, plan.getFragment().getStageExecutionDescriptor(), schedulingOrder, stageNodeList, bucketNodeMap, splitBatchSize, getConcurrentLifespansPerNode(session), nodeScheduler.createNodeSelector(catalogName, keepConsumerOnFeederNodes, feederScheduledNodes), connectorPartitionHandles, session, heuristicIndexerManager));
} else {
// all sources are remote
NodePartitionMap nodePartitionMap = partitioningCache.apply(plan.getFragment().getPartitioning(), stageTaskCounts == null ? null : stageTaskCounts.get(stageId));
List<InternalNode> partitionToNode = nodePartitionMap.getPartitionToNode();
// todo this should asynchronously wait a standard timeout period before failing
checkCondition(!partitionToNode.isEmpty(), NO_NODES_AVAILABLE, "No worker nodes available");
stageSchedulers.put(stageId, new FixedCountScheduler(stageExecution, partitionToNode));
bucketToPartition = Optional.of(nodePartitionMap.getBucketToPartition());
}
}
ImmutableSet.Builder<SqlStageExecution> childStagesBuilder = ImmutableSet.builder();
for (StageExecutionPlan subStagePlan : plan.getSubStages()) {
if (visitedPlanFrags.contains(subStagePlan.getFragment().getId())) {
continue;
}
visitedPlanFrags.add(subStagePlan.getFragment().getId());
List<SqlStageExecution> subTree = createStages(stageExecution::addExchangeLocations, nextStageId, locationFactory, subStagePlan.withBucketToPartition(bucketToPartition), nodeScheduler, remoteTaskFactory, session, splitBatchSize, partitioningCache, nodePartitioningManager, queryExecutor, schedulerExecutor, failureDetector, nodeTaskMap, stageSchedulers, stageLinkages, isSnapshotEnabled, snapshotManager, stageTaskCounts);
localStages.addAll(subTree);
SqlStageExecution childStage = subTree.get(0);
childStagesBuilder.add(childStage);
Optional<RemoteSourceNode> parentNode = plan.getFragment().getRemoteSourceNodes().stream().filter(x -> x.getSourceFragmentIds().contains(childStage.getFragment().getId())).findAny();
checkArgument(parentNode.isPresent(), "Couldn't find parent of a CTE node");
childStage.setParentId(parentNode.get().getId());
}
Set<SqlStageExecution> childStages = childStagesBuilder.build();
stageExecution.addStateChangeListener(newState -> {
if (newState.isDone() && newState != StageState.RESCHEDULING) {
// Snapshot: For "rescheduling", tasks are already cancelled (for resume)
childStages.forEach(SqlStageExecution::cancel);
}
});
stageLinkages.put(stageId, new StageLinkage(plan.getFragment().getId(), parent, childStages));
if (partitioningHandle.equals(SCALED_WRITER_DISTRIBUTION)) {
Supplier<Collection<TaskStatus>> sourceTasksProvider = () -> childStages.stream().map(SqlStageExecution::getAllTasks).flatMap(Collection::stream).map(RemoteTask::getTaskStatus).collect(toList());
Supplier<Collection<TaskStatus>> writerTasksProvider = () -> stageExecution.getAllTasks().stream().map(RemoteTask::getTaskStatus).collect(toList());
ScaledWriterScheduler scheduler = new ScaledWriterScheduler(stageExecution, sourceTasksProvider, writerTasksProvider, nodeScheduler.createNodeSelector(null, keepConsumerOnFeederNodes, feederScheduledNodes), schedulerExecutor, getWriterMinSize(session), isSnapshotEnabled, stageTaskCounts != null ? stageTaskCounts.get(stageId) : null);
whenAllStages(childStages, StageState::isDone).addListener(scheduler::finish, directExecutor());
stageSchedulers.put(stageId, scheduler);
}
return localStages.build();
}
Aggregations