public SqlQueryScheduler(
QueryStateMachine queryStateMachine,
LocationFactory locationFactory,
StageExecutionPlan plan,
NodePartitioningManager nodePartitioningManager,
NodeScheduler nodeScheduler,
RemoteTaskFactory remoteTaskFactory,
Session session,
boolean summarizeTaskInfo,
int splitBatchSize,
ExecutorService executor,
OutputBuffers rootOutputBuffers,
NodeTaskMap nodeTaskMap,
ExecutionPolicy executionPolicy) {
this.queryStateMachine = requireNonNull(queryStateMachine, "queryStateMachine is null");
this.executionPolicy = requireNonNull(executionPolicy, "schedulerPolicyFactory is null");
this.summarizeTaskInfo = summarizeTaskInfo;
// todo come up with a better way to build this, or eliminate this map
ImmutableMap.Builder<StageId, StageScheduler> stageSchedulers = ImmutableMap.builder();
ImmutableMap.Builder<StageId, StageLinkage> stageLinkages = ImmutableMap.builder();
// Only fetch a distribution once per query to assure all stages see the same machine
// assignments
Map<PartitioningHandle, NodePartitionMap> partitioningCache = new HashMap<>();
List<SqlStageExecution> stages =
createStages(
Optional.empty(),
new AtomicInteger(),
locationFactory,
plan.withBucketToPartition(Optional.of(new int[1])),
nodeScheduler,
remoteTaskFactory,
session,
splitBatchSize,
partitioningHandle ->
partitioningCache.computeIfAbsent(
partitioningHandle,
handle -> nodePartitioningManager.getNodePartitioningMap(session, handle)),
executor,
nodeTaskMap,
stageSchedulers,
stageLinkages);
SqlStageExecution rootStage = stages.get(0);
rootStage.setOutputBuffers(rootOutputBuffers);
this.rootStageId = rootStage.getStageId();
this.stages = stages.stream().collect(toImmutableMap(SqlStageExecution::getStageId));
this.stageSchedulers = stageSchedulers.build();
this.stageLinkages = stageLinkages.build();
this.executor = executor;
rootStage.addStateChangeListener(
state -> {
if (state == FINISHED) {
queryStateMachine.transitionToFinishing();
} else if (state == CANCELED) {
// output stage was canceled
queryStateMachine.transitionToFailed(
new PrestoException(USER_CANCELED, "Query was canceled"));
}
});
for (SqlStageExecution stage : stages) {
stage.addStateChangeListener(
state -> {
if (queryStateMachine.isDone()) {
return;
}
if (state == FAILED) {
queryStateMachine.transitionToFailed(
stage.getStageInfo().getFailureCause().toException());
} else if (state == ABORTED) {
// this should never happen, since abort can only be triggered in query clean up after
// the query is finished
queryStateMachine.transitionToFailed(
new PrestoException(INTERNAL_ERROR, "Query stage was aborted"));
} else if (queryStateMachine.getQueryState() == QueryState.STARTING) {
// if the stage has at least one task, we are running
if (stage.hasTasks()) {
queryStateMachine.transitionToRunning();
}
}
});
}
}
private List<SqlStageExecution> createStages(
Optional<SqlStageExecution> parent,
AtomicInteger nextStageId,
LocationFactory locationFactory,
StageExecutionPlan plan,
NodeScheduler nodeScheduler,
RemoteTaskFactory remoteTaskFactory,
Session session,
int splitBatchSize,
Function<PartitioningHandle, NodePartitionMap> partitioningCache,
ExecutorService executor,
NodeTaskMap nodeTaskMap,
ImmutableMap.Builder<StageId, StageScheduler> stageSchedulers,
ImmutableMap.Builder<StageId, StageLinkage> stageLinkages) {
ImmutableList.Builder<SqlStageExecution> stages = ImmutableList.builder();
StageId stageId =
new StageId(queryStateMachine.getQueryId(), String.valueOf(nextStageId.getAndIncrement()));
SqlStageExecution stage =
new SqlStageExecution(
stageId,
locationFactory.createStageLocation(stageId),
plan.getFragment(),
remoteTaskFactory,
session,
summarizeTaskInfo,
nodeTaskMap,
executor);
stages.add(stage);
Optional<int[]> bucketToPartition;
PartitioningHandle partitioningHandle = plan.getFragment().getPartitioning();
if (partitioningHandle.equals(SOURCE_DISTRIBUTION)) {
// nodes are selected dynamically based on the constraints of the splits and the system load
SplitSource splitSource = plan.getDataSource().get();
NodeSelector nodeSelector = nodeScheduler.createNodeSelector(splitSource.getDataSourceName());
SplitPlacementPolicy placementPolicy =
new DynamicSplitPlacementPolicy(nodeSelector, stage::getAllTasks);
stageSchedulers.put(
stageId,
new SourcePartitionedScheduler(stage, splitSource, placementPolicy, splitBatchSize));
bucketToPartition = Optional.of(new int[1]);
} else {
// nodes are pre determined by the nodePartitionMap
NodePartitionMap nodePartitionMap =
partitioningCache.apply(plan.getFragment().getPartitioning());
if (plan.getDataSource().isPresent()) {
stageSchedulers.put(
stageId,
new FixedSourcePartitionedScheduler(
stage,
plan.getDataSource().get(),
nodePartitionMap,
splitBatchSize,
nodeScheduler.createNodeSelector(null)));
bucketToPartition = Optional.of(nodePartitionMap.getBucketToPartition());
} else {
Map<Integer, Node> 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(stage, partitionToNode));
bucketToPartition = Optional.of(nodePartitionMap.getBucketToPartition());
}
}
ImmutableSet.Builder<SqlStageExecution> childStagesBuilder = ImmutableSet.builder();
for (StageExecutionPlan subStagePlan : plan.getSubStages()) {
List<SqlStageExecution> subTree =
createStages(
Optional.of(stage),
nextStageId,
locationFactory,
subStagePlan.withBucketToPartition(bucketToPartition),
nodeScheduler,
remoteTaskFactory,
session,
splitBatchSize,
partitioningCache,
executor,
nodeTaskMap,
stageSchedulers,
stageLinkages);
stages.addAll(subTree);
SqlStageExecution childStage = subTree.get(0);
childStagesBuilder.add(childStage);
}
Set<SqlStageExecution> childStages = childStagesBuilder.build();
stage.addStateChangeListener(
newState -> {
if (newState.isDone()) {
childStages.stream().forEach(SqlStageExecution::cancel);
}
});
stageLinkages.put(stageId, new StageLinkage(plan.getFragment().getId(), parent, childStages));
return stages.build();
}