Examples with TaskManagerLocation org.apache.flink.runtime.taskmanager.TaskManagerLocation used on opensource projects

Example 51 with TaskManagerLocation

use of org.apache.flink.runtime.taskmanager.TaskManagerLocation in project flink by apache.

the class InputChannelDeploymentDescriptor method fromEdges.

// ------------------------------------------------------------------------
/**
	 * Creates an input channel deployment descriptor for each partition.
	 */
public static InputChannelDeploymentDescriptor[] fromEdges(ExecutionEdge[] edges, SimpleSlot consumerSlot, boolean allowLazyDeployment) throws ExecutionGraphException {
    final ResourceID consumerTaskManager = consumerSlot.getTaskManagerID();
    final InputChannelDeploymentDescriptor[] icdd = new InputChannelDeploymentDescriptor[edges.length];
    // Each edge is connected to a different result partition
    for (int i = 0; i < edges.length; i++) {
        final IntermediateResultPartition consumedPartition = edges[i].getSource();
        final Execution producer = consumedPartition.getProducer().getCurrentExecutionAttempt();
        final ExecutionState producerState = producer.getState();
        final SimpleSlot producerSlot = producer.getAssignedResource();
        final ResultPartitionLocation partitionLocation;
        // The producing task needs to be RUNNING or already FINISHED
        if (consumedPartition.isConsumable() && producerSlot != null && (producerState == ExecutionState.RUNNING || producerState == ExecutionState.FINISHED || producerState == ExecutionState.SCHEDULED || producerState == ExecutionState.DEPLOYING)) {
            final TaskManagerLocation partitionTaskManagerLocation = producerSlot.getTaskManagerLocation();
            final ResourceID partitionTaskManager = partitionTaskManagerLocation.getResourceID();
            if (partitionTaskManager.equals(consumerTaskManager)) {
                // Consuming task is deployed to the same TaskManager as the partition => local
                partitionLocation = ResultPartitionLocation.createLocal();
            } else {
                // Different instances => remote
                final ConnectionID connectionId = new ConnectionID(partitionTaskManagerLocation, consumedPartition.getIntermediateResult().getConnectionIndex());
                partitionLocation = ResultPartitionLocation.createRemote(connectionId);
            }
        } else if (allowLazyDeployment) {
            // The producing task might not have registered the partition yet
            partitionLocation = ResultPartitionLocation.createUnknown();
        } else if (producerState == ExecutionState.CANCELING || producerState == ExecutionState.CANCELED || producerState == ExecutionState.FAILED) {
            String msg = "Trying to schedule a task whose inputs were canceled or failed. " + "The producer is in state " + producerState + ".";
            throw new ExecutionGraphException(msg);
        } else {
            String msg = String.format("Trying to eagerly schedule a task whose inputs " + "are not ready (partition consumable? %s, producer state: %s, producer slot: %s).", consumedPartition.isConsumable(), producerState, producerSlot);
            throw new ExecutionGraphException(msg);
        }
        final ResultPartitionID consumedPartitionId = new ResultPartitionID(consumedPartition.getPartitionId(), producer.getAttemptId());
        icdd[i] = new InputChannelDeploymentDescriptor(consumedPartitionId, partitionLocation);
    }
    return icdd;
}

Example 52 with TaskManagerLocation

use of org.apache.flink.runtime.taskmanager.TaskManagerLocation in project flink by apache.

the class PartialInputChannelDeploymentDescriptor method createInputChannelDeploymentDescriptor.

/**
	 * Creates a channel deployment descriptor by completing the partition location.
	 *
	 * @see InputChannelDeploymentDescriptor
	 */
public InputChannelDeploymentDescriptor createInputChannelDeploymentDescriptor(Execution consumerExecution) {
    checkNotNull(consumerExecution, "consumerExecution");
    TaskManagerLocation consumerLocation = consumerExecution.getAssignedResourceLocation();
    checkNotNull(consumerLocation, "Consumer connection info null");
    final ResultPartitionLocation partitionLocation;
    if (consumerLocation.equals(partitionTaskManagerLocation)) {
        partitionLocation = ResultPartitionLocation.createLocal();
    } else {
        partitionLocation = ResultPartitionLocation.createRemote(new ConnectionID(partitionTaskManagerLocation, partitionConnectionIndex));
    }
    return new InputChannelDeploymentDescriptor(partitionID, partitionLocation);
}

Example 53 with TaskManagerLocation

use of org.apache.flink.runtime.taskmanager.TaskManagerLocation in project flink by apache.

the class PartialInputChannelDeploymentDescriptor method fromEdge.

// ------------------------------------------------------------------------
/**
	 * Creates a partial input channel for the given partition and producing task.
	 */
public static PartialInputChannelDeploymentDescriptor fromEdge(IntermediateResultPartition partition, Execution producer) {
    final ResultPartitionID partitionId = new ResultPartitionID(partition.getPartitionId(), producer.getAttemptId());
    final IntermediateResult result = partition.getIntermediateResult();
    final IntermediateDataSetID resultId = result.getId();
    final TaskManagerLocation partitionConnectionInfo = producer.getAssignedResourceLocation();
    final int partitionConnectionIndex = result.getConnectionIndex();
    return new PartialInputChannelDeploymentDescriptor(resultId, partitionId, partitionConnectionInfo, partitionConnectionIndex);
}

Example 54 with TaskManagerLocation

use of org.apache.flink.runtime.taskmanager.TaskManagerLocation in project flink by apache.

the class Execution method scheduleOrUpdateConsumers.

void scheduleOrUpdateConsumers(List<List<ExecutionEdge>> allConsumers) {
    final int numConsumers = allConsumers.size();
    if (numConsumers > 1) {
        fail(new IllegalStateException("Currently, only a single consumer group per partition is supported."));
    } else if (numConsumers == 0) {
        return;
    }
    for (ExecutionEdge edge : allConsumers.get(0)) {
        final ExecutionVertex consumerVertex = edge.getTarget();
        final Execution consumer = consumerVertex.getCurrentExecutionAttempt();
        final ExecutionState consumerState = consumer.getState();
        final IntermediateResultPartition partition = edge.getSource();
        // ----------------------------------------------------------------
        if (consumerState == CREATED) {
            final Execution partitionExecution = partition.getProducer().getCurrentExecutionAttempt();
            consumerVertex.cachePartitionInfo(PartialInputChannelDeploymentDescriptor.fromEdge(partition, partitionExecution));
            // When deploying a consuming task, its task deployment descriptor will contain all
            // deployment information available at the respective time. It is possible that some
            // of the partitions to be consumed have not been created yet. These are updated
            // runtime via the update messages.
            //
            // TODO The current approach may send many update messages even though the consuming
            // task has already been deployed with all necessary information. We have to check
            // whether this is a problem and fix it, if it is.
            FlinkFuture.supplyAsync(new Callable<Void>() {

                @Override
                public Void call() throws Exception {
                    try {
                        consumerVertex.scheduleForExecution(consumerVertex.getExecutionGraph().getSlotProvider(), consumerVertex.getExecutionGraph().isQueuedSchedulingAllowed());
                    } catch (Throwable t) {
                        consumerVertex.fail(new IllegalStateException("Could not schedule consumer " + "vertex " + consumerVertex, t));
                    }
                    return null;
                }
            }, executor);
            // double check to resolve race conditions
            if (consumerVertex.getExecutionState() == RUNNING) {
                consumerVertex.sendPartitionInfos();
            }
        } else // ----------------------------------------------------------------
        // Consumer is running => send update message now
        // ----------------------------------------------------------------
        {
            if (consumerState == RUNNING) {
                final SimpleSlot consumerSlot = consumer.getAssignedResource();
                if (consumerSlot == null) {
                    // The consumer has been reset concurrently
                    continue;
                }
                final TaskManagerLocation partitionTaskManagerLocation = partition.getProducer().getCurrentAssignedResource().getTaskManagerLocation();
                final ResourceID partitionTaskManager = partitionTaskManagerLocation.getResourceID();
                final ResourceID consumerTaskManager = consumerSlot.getTaskManagerID();
                final ResultPartitionID partitionId = new ResultPartitionID(partition.getPartitionId(), attemptId);
                final ResultPartitionLocation partitionLocation;
                if (consumerTaskManager.equals(partitionTaskManager)) {
                    // Consuming task is deployed to the same instance as the partition => local
                    partitionLocation = ResultPartitionLocation.createLocal();
                } else {
                    // Different instances => remote
                    final ConnectionID connectionId = new ConnectionID(partitionTaskManagerLocation, partition.getIntermediateResult().getConnectionIndex());
                    partitionLocation = ResultPartitionLocation.createRemote(connectionId);
                }
                final InputChannelDeploymentDescriptor descriptor = new InputChannelDeploymentDescriptor(partitionId, partitionLocation);
                consumer.sendUpdatePartitionInfoRpcCall(Collections.singleton(new PartitionInfo(partition.getIntermediateResult().getId(), descriptor)));
            } else // ----------------------------------------------------------------
            if (consumerState == SCHEDULED || consumerState == DEPLOYING) {
                final Execution partitionExecution = partition.getProducer().getCurrentExecutionAttempt();
                consumerVertex.cachePartitionInfo(PartialInputChannelDeploymentDescriptor.fromEdge(partition, partitionExecution));
                // double check to resolve race conditions
                if (consumerVertex.getExecutionState() == RUNNING) {
                    consumerVertex.sendPartitionInfos();
                }
            }
        }
    }
}

Example 55 with TaskManagerLocation

use of org.apache.flink.runtime.taskmanager.TaskManagerLocation in project flink by apache.

the class SlotSharingGroupAssignment method getSlotForTaskInternal.

private Tuple2<SharedSlot, Locality> getSlotForTaskInternal(AbstractID groupId, Iterable<TaskManagerLocation> preferredLocations, boolean localOnly) {
    // check if there is anything at all in this group assignment
    if (allSlots.isEmpty()) {
        return null;
    }
    // get the available slots for the group
    Map<ResourceID, List<SharedSlot>> slotsForGroup = availableSlotsPerJid.get(groupId);
    if (slotsForGroup == null) {
        // we have a new group, so all slots are available
        slotsForGroup = new LinkedHashMap<>();
        availableSlotsPerJid.put(groupId, slotsForGroup);
        for (SharedSlot availableSlot : allSlots) {
            putIntoMultiMap(slotsForGroup, availableSlot.getTaskManagerID(), availableSlot);
        }
    } else if (slotsForGroup.isEmpty()) {
        // the group exists, but nothing is available for that group
        return null;
    }
    // check whether we can schedule the task to a preferred location
    boolean didNotGetPreferred = false;
    if (preferredLocations != null) {
        for (TaskManagerLocation location : preferredLocations) {
            // set the flag that we failed a preferred location. If one will be found,
            // we return early anyways and skip the flag evaluation
            didNotGetPreferred = true;
            SharedSlot slot = removeFromMultiMap(slotsForGroup, location.getResourceID());
            if (slot != null && slot.isAlive()) {
                return new Tuple2<>(slot, Locality.LOCAL);
            }
        }
    }
    // if we want only local assignments, exit now with a "not found" result
    if (didNotGetPreferred && localOnly) {
        return null;
    }
    Locality locality = didNotGetPreferred ? Locality.NON_LOCAL : Locality.UNCONSTRAINED;
    // schedule the task to any available location
    SharedSlot slot;
    while ((slot = pollFromMultiMap(slotsForGroup)) != null) {
        if (slot.isAlive()) {
            return new Tuple2<>(slot, locality);
        }
    }
    // nothing available after all, all slots were dead
    return null;
}