Examples with SerializedPrestoSparkTaskDescriptor com.facebook.presto.spark.classloader_interface.SerializedPrestoSparkTaskDescriptor used on opensource projects

Example 1 with SerializedPrestoSparkTaskDescriptor

use of com.facebook.presto.spark.classloader_interface.SerializedPrestoSparkTaskDescriptor 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 2 with SerializedPrestoSparkTaskDescriptor

use of com.facebook.presto.spark.classloader_interface.SerializedPrestoSparkTaskDescriptor in project presto by prestodb.

the class PrestoSparkRddFactory method createRdd.

private <T extends PrestoSparkTaskOutput> JavaPairRDD<MutablePartitionId, T> createRdd(JavaSparkContext sparkContext, Session session, PlanFragment fragment, PrestoSparkTaskExecutorFactoryProvider executorFactoryProvider, CollectionAccumulator<SerializedTaskInfo> taskInfoCollector, CollectionAccumulator<PrestoSparkShuffleStats> shuffleStatsCollector, TableWriteInfo tableWriteInfo, Map<PlanFragmentId, JavaPairRDD<MutablePartitionId, PrestoSparkMutableRow>> rddInputs, Map<PlanFragmentId, Broadcast<?>> broadcastInputs, Class<T> outputType) {
    checkInputs(fragment.getRemoteSourceNodes(), rddInputs, broadcastInputs);
    PrestoSparkTaskDescriptor taskDescriptor = new PrestoSparkTaskDescriptor(session.toSessionRepresentation(), session.getIdentity().getExtraCredentials(), fragment, tableWriteInfo);
    SerializedPrestoSparkTaskDescriptor serializedTaskDescriptor = new SerializedPrestoSparkTaskDescriptor(taskDescriptorJsonCodec.toJsonBytes(taskDescriptor));
    Optional<Integer> numberOfShufflePartitions = Optional.empty();
    Map<String, RDD<Tuple2<MutablePartitionId, PrestoSparkMutableRow>>> shuffleInputRddMap = new HashMap<>();
    for (Map.Entry<PlanFragmentId, JavaPairRDD<MutablePartitionId, PrestoSparkMutableRow>> input : rddInputs.entrySet()) {
        RDD<Tuple2<MutablePartitionId, PrestoSparkMutableRow>> rdd = input.getValue().rdd();
        shuffleInputRddMap.put(input.getKey().toString(), rdd);
        if (!numberOfShufflePartitions.isPresent()) {
            numberOfShufflePartitions = Optional.of(rdd.getNumPartitions());
        } else {
            checkArgument(numberOfShufflePartitions.get() == rdd.getNumPartitions(), "Incompatible number of input partitions: %s != %s", numberOfShufflePartitions.get(), rdd.getNumPartitions());
        }
    }
    PrestoSparkTaskProcessor<T> taskProcessor = new PrestoSparkTaskProcessor<>(executorFactoryProvider, serializedTaskDescriptor, taskInfoCollector, shuffleStatsCollector, toTaskProcessorBroadcastInputs(broadcastInputs), outputType);
    Optional<PrestoSparkTaskSourceRdd> taskSourceRdd;
    List<TableScanNode> tableScans = findTableScanNodes(fragment.getRoot());
    if (!tableScans.isEmpty()) {
        try (CloseableSplitSourceProvider splitSourceProvider = new CloseableSplitSourceProvider(splitManager::getSplits)) {
            SplitSourceFactory splitSourceFactory = new SplitSourceFactory(splitSourceProvider, WarningCollector.NOOP);
            Map<PlanNodeId, SplitSource> splitSources = splitSourceFactory.createSplitSources(fragment, session, tableWriteInfo);
            taskSourceRdd = Optional.of(createTaskSourcesRdd(fragment.getId(), sparkContext, session, fragment.getPartitioning(), tableScans, splitSources, numberOfShufflePartitions));
        }
    } else if (rddInputs.size() == 0) {
        checkArgument(fragment.getPartitioning().equals(SINGLE_DISTRIBUTION), "SINGLE_DISTRIBUTION partitioning is expected: %s", fragment.getPartitioning());
        // In case of no inputs we still need to schedule a task.
        // Task with no inputs may produce results (e.g.: ValuesNode).
        // To force the task to be scheduled we create a PrestoSparkTaskSourceRdd that contains exactly one partition.
        // Since there's also no table scans in the fragment, the list of TaskSource's for this partition is empty.
        taskSourceRdd = Optional.of(new PrestoSparkTaskSourceRdd(sparkContext.sc(), ImmutableList.of(ImmutableList.of())));
    } else {
        taskSourceRdd = Optional.empty();
    }
    return JavaPairRDD.fromRDD(PrestoSparkTaskRdd.create(sparkContext.sc(), taskSourceRdd, shuffleInputRddMap, taskProcessor), classTag(MutablePartitionId.class), classTag(outputType));
}