Search in sources :

Example 1 with OutputBuffers

use of io.trino.execution.buffer.OutputBuffers in project trino by trinodb.

the class FaultTolerantStageScheduler method schedule.

public synchronized void schedule() throws Exception {
    if (failure != null) {
        propagateIfPossible(failure, Exception.class);
        throw new RuntimeException(failure);
    }
    if (closed) {
        return;
    }
    if (isFinished()) {
        return;
    }
    if (!blocked.isDone()) {
        return;
    }
    if (taskSource == null) {
        Map<PlanFragmentId, ListenableFuture<List<ExchangeSourceHandle>>> sourceHandles = sourceExchanges.entrySet().stream().collect(toImmutableMap(Map.Entry::getKey, entry -> toListenableFuture(entry.getValue().getSourceHandles())));
        List<ListenableFuture<List<ExchangeSourceHandle>>> blockedFutures = sourceHandles.values().stream().filter(future -> !future.isDone()).collect(toImmutableList());
        if (!blockedFutures.isEmpty()) {
            blocked = asVoid(allAsList(blockedFutures));
            return;
        }
        Multimap<PlanFragmentId, ExchangeSourceHandle> exchangeSources = sourceHandles.entrySet().stream().collect(flatteningToImmutableListMultimap(Map.Entry::getKey, entry -> getFutureValue(entry.getValue()).stream()));
        taskSource = taskSourceFactory.create(session, stage.getFragment(), sourceExchanges, exchangeSources, stage::recordGetSplitTime, sourceBucketToPartitionMap, sourceBucketNodeMap);
    }
    while (!queuedPartitions.isEmpty() || !taskSource.isFinished()) {
        while (queuedPartitions.isEmpty() && !taskSource.isFinished()) {
            List<TaskDescriptor> tasks = taskSource.getMoreTasks();
            for (TaskDescriptor task : tasks) {
                queuedPartitions.add(task.getPartitionId());
                allPartitions.add(task.getPartitionId());
                taskDescriptorStorage.put(stage.getStageId(), task);
                sinkExchange.ifPresent(exchange -> {
                    ExchangeSinkHandle exchangeSinkHandle = exchange.addSink(task.getPartitionId());
                    partitionToExchangeSinkHandleMap.put(task.getPartitionId(), exchangeSinkHandle);
                });
            }
            if (taskSource.isFinished()) {
                sinkExchange.ifPresent(Exchange::noMoreSinks);
            }
        }
        if (queuedPartitions.isEmpty()) {
            break;
        }
        int partition = queuedPartitions.peek();
        Optional<TaskDescriptor> taskDescriptorOptional = taskDescriptorStorage.get(stage.getStageId(), partition);
        if (taskDescriptorOptional.isEmpty()) {
            // query has been terminated
            return;
        }
        TaskDescriptor taskDescriptor = taskDescriptorOptional.get();
        MemoryRequirements memoryRequirements = partitionMemoryRequirements.computeIfAbsent(partition, ignored -> partitionMemoryEstimator.getInitialMemoryRequirements(session, taskDescriptor.getNodeRequirements().getMemory()));
        if (nodeLease == null) {
            NodeRequirements nodeRequirements = taskDescriptor.getNodeRequirements();
            nodeRequirements = nodeRequirements.withMemory(memoryRequirements.getRequiredMemory());
            nodeLease = nodeAllocator.acquire(nodeRequirements);
        }
        if (!nodeLease.getNode().isDone()) {
            blocked = asVoid(nodeLease.getNode());
            return;
        }
        NodeInfo node = getFutureValue(nodeLease.getNode());
        queuedPartitions.poll();
        Multimap<PlanNodeId, Split> tableScanSplits = taskDescriptor.getSplits();
        Multimap<PlanNodeId, Split> remoteSplits = createRemoteSplits(taskDescriptor.getExchangeSourceHandles());
        Multimap<PlanNodeId, Split> taskSplits = ImmutableListMultimap.<PlanNodeId, Split>builder().putAll(tableScanSplits).putAll(remoteSplits).build();
        int attemptId = getNextAttemptIdForPartition(partition);
        OutputBuffers outputBuffers;
        Optional<ExchangeSinkInstanceHandle> exchangeSinkInstanceHandle;
        if (sinkExchange.isPresent()) {
            ExchangeSinkHandle sinkHandle = partitionToExchangeSinkHandleMap.get(partition);
            exchangeSinkInstanceHandle = Optional.of(sinkExchange.get().instantiateSink(sinkHandle, attemptId));
            outputBuffers = createSpoolingExchangeOutputBuffers(exchangeSinkInstanceHandle.get());
        } else {
            exchangeSinkInstanceHandle = Optional.empty();
            // stage will be consumed by the coordinator using direct exchange
            outputBuffers = createInitialEmptyOutputBuffers(PARTITIONED).withBuffer(new OutputBuffers.OutputBufferId(0), 0).withNoMoreBufferIds();
        }
        Set<PlanNodeId> allSourcePlanNodeIds = ImmutableSet.<PlanNodeId>builder().addAll(stage.getFragment().getPartitionedSources()).addAll(stage.getFragment().getRemoteSourceNodes().stream().map(RemoteSourceNode::getId).iterator()).build();
        RemoteTask task = stage.createTask(node.getNode(), partition, attemptId, sinkBucketToPartitionMap, outputBuffers, taskSplits, allSourcePlanNodeIds.stream().collect(toImmutableListMultimap(Function.identity(), planNodeId -> Lifespan.taskWide())), allSourcePlanNodeIds).orElseThrow(() -> new VerifyException("stage execution is expected to be active"));
        partitionToRemoteTaskMap.put(partition, task);
        runningTasks.put(task.getTaskId(), task);
        runningNodes.put(task.getTaskId(), nodeLease);
        nodeLease = null;
        if (taskFinishedFuture == null) {
            taskFinishedFuture = SettableFuture.create();
        }
        taskLifecycleListener.taskCreated(stage.getFragment().getId(), task);
        task.addStateChangeListener(taskStatus -> updateTaskStatus(taskStatus, exchangeSinkInstanceHandle));
        task.start();
    }
    if (taskFinishedFuture != null && !taskFinishedFuture.isDone()) {
        blocked = taskFinishedFuture;
    }
}
Also used : ArrayListMultimap(com.google.common.collect.ArrayListMultimap) SettableFuture(com.google.common.util.concurrent.SettableFuture) RemoteSourceNode(io.trino.sql.planner.plan.RemoteSourceNode) Preconditions.checkArgument(com.google.common.base.Preconditions.checkArgument) Throwables.propagateIfPossible(com.google.common.base.Throwables.propagateIfPossible) ImmutableListMultimap.toImmutableListMultimap(com.google.common.collect.ImmutableListMultimap.toImmutableListMultimap) MemoryRequirements(io.trino.execution.scheduler.PartitionMemoryEstimator.MemoryRequirements) PlanNodeId(io.trino.sql.planner.plan.PlanNodeId) Map(java.util.Map) SpoolingExchangeInput(io.trino.split.RemoteSplit.SpoolingExchangeInput) REMOTE_HOST_GONE(io.trino.spi.StandardErrorCode.REMOTE_HOST_GONE) Futures.immediateVoidFuture(com.google.common.util.concurrent.Futures.immediateVoidFuture) ImmutableSet(com.google.common.collect.ImmutableSet) ExchangeSinkInstanceHandle(io.trino.spi.exchange.ExchangeSinkInstanceHandle) OutputBuffers.createSpoolingExchangeOutputBuffers(io.trino.execution.buffer.OutputBuffers.createSpoolingExchangeOutputBuffers) ImmutableMap(com.google.common.collect.ImmutableMap) ExecutionFailureInfo(io.trino.execution.ExecutionFailureInfo) Futures.allAsList(com.google.common.util.concurrent.Futures.allAsList) ImmutableList.toImmutableList(com.google.common.collect.ImmutableList.toImmutableList) MoreFutures.toListenableFuture(io.airlift.concurrent.MoreFutures.toListenableFuture) Set(java.util.Set) TrinoException(io.trino.spi.TrinoException) GONE(io.trino.failuredetector.FailureDetector.State.GONE) GuardedBy(javax.annotation.concurrent.GuardedBy) TaskId(io.trino.execution.TaskId) ExchangeSinkHandle(io.trino.spi.exchange.ExchangeSinkHandle) List(java.util.List) ImmutableMap.toImmutableMap(com.google.common.collect.ImmutableMap.toImmutableMap) Split(io.trino.metadata.Split) ImmutableListMultimap(com.google.common.collect.ImmutableListMultimap) Optional(java.util.Optional) Queue(java.util.Queue) PlanFragmentId(io.trino.sql.planner.plan.PlanFragmentId) OutputBuffers.createInitialEmptyOutputBuffers(io.trino.execution.buffer.OutputBuffers.createInitialEmptyOutputBuffers) ExchangeSourceHandle(io.trino.spi.exchange.ExchangeSourceHandle) Session(io.trino.Session) ImmutableListMultimap.flatteningToImmutableListMultimap(com.google.common.collect.ImmutableListMultimap.flatteningToImmutableListMultimap) ListenableFuture(com.google.common.util.concurrent.ListenableFuture) StageId(io.trino.execution.StageId) Logger(io.airlift.log.Logger) HashMap(java.util.HashMap) Multimap(com.google.common.collect.Multimap) ErrorCode(io.trino.spi.ErrorCode) Function(java.util.function.Function) Failures.toFailure(io.trino.util.Failures.toFailure) RemoteSplit(io.trino.split.RemoteSplit) HashSet(java.util.HashSet) ImmutableList(com.google.common.collect.ImmutableList) Verify.verify(com.google.common.base.Verify.verify) USER_ERROR(io.trino.spi.ErrorType.USER_ERROR) Objects.requireNonNull(java.util.Objects.requireNonNull) TaskState(io.trino.execution.TaskState) Lifespan(io.trino.execution.Lifespan) Exchange(io.trino.spi.exchange.Exchange) VerifyException(com.google.common.base.VerifyException) SqlStage(io.trino.execution.SqlStage) FailureDetector(io.trino.failuredetector.FailureDetector) RemoteTask(io.trino.execution.RemoteTask) TaskStatus(io.trino.execution.TaskStatus) MoreFutures.getFutureValue(io.airlift.concurrent.MoreFutures.getFutureValue) GENERIC_INTERNAL_ERROR(io.trino.spi.StandardErrorCode.GENERIC_INTERNAL_ERROR) MoreFutures.asVoid(io.airlift.concurrent.MoreFutures.asVoid) PARTITIONED(io.trino.execution.buffer.OutputBuffers.BufferType.PARTITIONED) Futures.nonCancellationPropagating(com.google.common.util.concurrent.Futures.nonCancellationPropagating) OutputBuffers(io.trino.execution.buffer.OutputBuffers) ArrayDeque(java.util.ArrayDeque) REMOTE_CONNECTOR_ID(io.trino.operator.ExchangeOperator.REMOTE_CONNECTOR_ID) MemoryRequirements(io.trino.execution.scheduler.PartitionMemoryEstimator.MemoryRequirements) ExchangeSourceHandle(io.trino.spi.exchange.ExchangeSourceHandle) PlanNodeId(io.trino.sql.planner.plan.PlanNodeId) OutputBuffers.createSpoolingExchangeOutputBuffers(io.trino.execution.buffer.OutputBuffers.createSpoolingExchangeOutputBuffers) OutputBuffers.createInitialEmptyOutputBuffers(io.trino.execution.buffer.OutputBuffers.createInitialEmptyOutputBuffers) OutputBuffers(io.trino.execution.buffer.OutputBuffers) PlanFragmentId(io.trino.sql.planner.plan.PlanFragmentId) ExchangeSinkInstanceHandle(io.trino.spi.exchange.ExchangeSinkInstanceHandle) RemoteTask(io.trino.execution.RemoteTask) Exchange(io.trino.spi.exchange.Exchange) ExchangeSinkHandle(io.trino.spi.exchange.ExchangeSinkHandle) VerifyException(com.google.common.base.VerifyException) MoreFutures.toListenableFuture(io.airlift.concurrent.MoreFutures.toListenableFuture) ListenableFuture(com.google.common.util.concurrent.ListenableFuture) Split(io.trino.metadata.Split) RemoteSplit(io.trino.split.RemoteSplit) Map(java.util.Map) ImmutableMap(com.google.common.collect.ImmutableMap) ImmutableMap.toImmutableMap(com.google.common.collect.ImmutableMap.toImmutableMap) HashMap(java.util.HashMap)

Example 2 with OutputBuffers

use of io.trino.execution.buffer.OutputBuffers in project trino by trinodb.

the class TestSqlTask method testBufferCloseOnFinish.

@Test
public void testBufferCloseOnFinish() throws Exception {
    SqlTask sqlTask = createInitialTask();
    OutputBuffers outputBuffers = createInitialEmptyOutputBuffers(PARTITIONED).withBuffer(OUT, 0).withNoMoreBufferIds();
    updateTask(sqlTask, EMPTY_SPLIT_ASSIGNMENTS, outputBuffers);
    ListenableFuture<BufferResult> bufferResult = sqlTask.getTaskResults(OUT, 0, DataSize.of(1, MEGABYTE));
    assertFalse(bufferResult.isDone());
    // close the sources (no splits will ever be added)
    updateTask(sqlTask, ImmutableList.of(new SplitAssignment(TABLE_SCAN_NODE_ID, ImmutableSet.of(), true)), outputBuffers);
    // finish the task by calling abort on it
    sqlTask.destroyTaskResults(OUT);
    // buffer will be closed by cancel event (wait for event to fire)
    bufferResult.get(1, SECONDS);
    // verify the buffer is closed
    bufferResult = sqlTask.getTaskResults(OUT, 0, DataSize.of(1, MEGABYTE));
    assertTrue(bufferResult.isDone());
    assertTrue(bufferResult.get().isBufferComplete());
}
Also used : SqlTask.createSqlTask(io.trino.execution.SqlTask.createSqlTask) OutputBuffers.createInitialEmptyOutputBuffers(io.trino.execution.buffer.OutputBuffers.createInitialEmptyOutputBuffers) OutputBuffers(io.trino.execution.buffer.OutputBuffers) BufferResult(io.trino.execution.buffer.BufferResult) Test(org.testng.annotations.Test)

Example 3 with OutputBuffers

use of io.trino.execution.buffer.OutputBuffers in project trino by trinodb.

the class TestBroadcastOutputBufferManager method test.

@Test
public void test() {
    BroadcastOutputBufferManager hashOutputBufferManager = new BroadcastOutputBufferManager();
    assertEquals(hashOutputBufferManager.getOutputBuffers(), createInitialEmptyOutputBuffers(BROADCAST));
    hashOutputBufferManager.addOutputBuffer(new OutputBufferId(0));
    OutputBuffers expectedOutputBuffers = createInitialEmptyOutputBuffers(BROADCAST).withBuffer(new OutputBufferId(0), BROADCAST_PARTITION_ID);
    assertEquals(hashOutputBufferManager.getOutputBuffers(), expectedOutputBuffers);
    hashOutputBufferManager.addOutputBuffer(new OutputBufferId(1));
    hashOutputBufferManager.addOutputBuffer(new OutputBufferId(2));
    expectedOutputBuffers = expectedOutputBuffers.withBuffer(new OutputBufferId(1), BROADCAST_PARTITION_ID);
    expectedOutputBuffers = expectedOutputBuffers.withBuffer(new OutputBufferId(2), BROADCAST_PARTITION_ID);
    assertEquals(hashOutputBufferManager.getOutputBuffers(), expectedOutputBuffers);
    // set no more buffers
    hashOutputBufferManager.addOutputBuffer(new OutputBufferId(3));
    hashOutputBufferManager.noMoreBuffers();
    expectedOutputBuffers = expectedOutputBuffers.withBuffer(new OutputBufferId(3), BROADCAST_PARTITION_ID);
    expectedOutputBuffers = expectedOutputBuffers.withNoMoreBufferIds();
    assertEquals(hashOutputBufferManager.getOutputBuffers(), expectedOutputBuffers);
    // try to add another buffer, which should not result in an error
    // and output buffers should not change
    hashOutputBufferManager.addOutputBuffer(new OutputBufferId(5));
    assertEquals(hashOutputBufferManager.getOutputBuffers(), expectedOutputBuffers);
    // try to set no more buffers again, which should not result in an error
    // and output buffers should not change
    hashOutputBufferManager.addOutputBuffer(new OutputBufferId(6));
    assertEquals(hashOutputBufferManager.getOutputBuffers(), expectedOutputBuffers);
}
Also used : OutputBuffers(io.trino.execution.buffer.OutputBuffers) OutputBuffers.createInitialEmptyOutputBuffers(io.trino.execution.buffer.OutputBuffers.createInitialEmptyOutputBuffers) OutputBufferId(io.trino.execution.buffer.OutputBuffers.OutputBufferId) Test(org.testng.annotations.Test)

Example 4 with OutputBuffers

use of io.trino.execution.buffer.OutputBuffers in project trino by trinodb.

the class PipelinedStageExecution method scheduleTask.

@Override
public synchronized Optional<RemoteTask> scheduleTask(InternalNode node, int partition, Multimap<PlanNodeId, Split> initialSplits, Multimap<PlanNodeId, Lifespan> noMoreSplitsForLifespan) {
    if (stateMachine.getState().isDone()) {
        return Optional.empty();
    }
    checkArgument(!tasks.containsKey(partition), "A task for partition %s already exists", partition);
    OutputBuffers outputBuffers = outputBufferManagers.get(stage.getFragment().getId()).getOutputBuffers();
    Optional<RemoteTask> optionalTask = stage.createTask(node, partition, attempt, bucketToPartition, outputBuffers, initialSplits, ImmutableMultimap.of(), ImmutableSet.of());
    if (optionalTask.isEmpty()) {
        return Optional.empty();
    }
    RemoteTask task = optionalTask.get();
    tasks.put(partition, task);
    ImmutableMultimap.Builder<PlanNodeId, Split> exchangeSplits = ImmutableMultimap.builder();
    sourceTasks.forEach((fragmentId, sourceTask) -> {
        TaskStatus status = sourceTask.getTaskStatus();
        if (status.getState() != TaskState.FINISHED) {
            PlanNodeId planNodeId = exchangeSources.get(fragmentId).getId();
            exchangeSplits.put(planNodeId, createExchangeSplit(sourceTask, task));
        }
    });
    allTasks.add(task.getTaskId());
    task.addSplits(exchangeSplits.build());
    noMoreSplitsForLifespan.forEach(task::noMoreSplits);
    completeSources.forEach(task::noMoreSplits);
    task.addStateChangeListener(this::updateTaskStatus);
    task.addStateChangeListener(this::updateCompletedDriverGroups);
    task.start();
    taskLifecycleListener.taskCreated(stage.getFragment().getId(), task);
    // update output buffers
    OutputBufferId outputBufferId = new OutputBufferId(task.getTaskId().getPartitionId());
    updateSourceTasksOutputBuffers(outputBufferManager -> outputBufferManager.addOutputBuffer(outputBufferId));
    return Optional.of(task);
}
Also used : PlanNodeId(io.trino.sql.planner.plan.PlanNodeId) OutputBuffers(io.trino.execution.buffer.OutputBuffers) RemoteTask(io.trino.execution.RemoteTask) OutputBufferId(io.trino.execution.buffer.OutputBuffers.OutputBufferId) ImmutableMultimap(com.google.common.collect.ImmutableMultimap) Split(io.trino.metadata.Split) RemoteSplit(io.trino.split.RemoteSplit) TaskStatus(io.trino.execution.TaskStatus)

Aggregations

OutputBuffers (io.trino.execution.buffer.OutputBuffers)4 OutputBuffers.createInitialEmptyOutputBuffers (io.trino.execution.buffer.OutputBuffers.createInitialEmptyOutputBuffers)3 RemoteTask (io.trino.execution.RemoteTask)2 TaskStatus (io.trino.execution.TaskStatus)2 OutputBufferId (io.trino.execution.buffer.OutputBuffers.OutputBufferId)2 Test (org.testng.annotations.Test)2 Preconditions.checkArgument (com.google.common.base.Preconditions.checkArgument)1 Throwables.propagateIfPossible (com.google.common.base.Throwables.propagateIfPossible)1 Verify.verify (com.google.common.base.Verify.verify)1 VerifyException (com.google.common.base.VerifyException)1 ArrayListMultimap (com.google.common.collect.ArrayListMultimap)1 ImmutableList (com.google.common.collect.ImmutableList)1 ImmutableList.toImmutableList (com.google.common.collect.ImmutableList.toImmutableList)1 ImmutableListMultimap (com.google.common.collect.ImmutableListMultimap)1 ImmutableListMultimap.flatteningToImmutableListMultimap (com.google.common.collect.ImmutableListMultimap.flatteningToImmutableListMultimap)1 ImmutableListMultimap.toImmutableListMultimap (com.google.common.collect.ImmutableListMultimap.toImmutableListMultimap)1 ImmutableMap (com.google.common.collect.ImmutableMap)1 ImmutableMap.toImmutableMap (com.google.common.collect.ImmutableMap.toImmutableMap)1 ImmutableMultimap (com.google.common.collect.ImmutableMultimap)1 ImmutableSet (com.google.common.collect.ImmutableSet)1