Examples with NodePartitionMap io.prestosql.sql.planner.NodePartitionMap used on opensource projects

Example 1 with NodePartitionMap

use of io.prestosql.sql.planner.NodePartitionMap 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();
}