zhijiangW commented on a change in pull request #8362: [FLINK-11391] Introduce shuffle master interface URL: https://github.com/apache/flink/pull/8362#discussion_r289690010
########## File path: flink-runtime/src/main/java/org/apache/flink/runtime/deployment/TaskDeploymentDescriptorFactory.java ########## @@ -0,0 +1,268 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.flink.runtime.deployment; + +import org.apache.flink.annotation.VisibleForTesting; +import org.apache.flink.api.common.JobID; +import org.apache.flink.runtime.blob.PermanentBlobKey; +import org.apache.flink.runtime.checkpoint.JobManagerTaskRestore; +import org.apache.flink.runtime.clusterframework.types.AllocationID; +import org.apache.flink.runtime.clusterframework.types.ResourceID; +import org.apache.flink.runtime.deployment.TaskDeploymentDescriptor.MaybeOffloaded; +import org.apache.flink.runtime.execution.ExecutionState; +import org.apache.flink.runtime.executiongraph.Execution; +import org.apache.flink.runtime.executiongraph.ExecutionAttemptID; +import org.apache.flink.runtime.executiongraph.ExecutionEdge; +import org.apache.flink.runtime.executiongraph.ExecutionGraph; +import org.apache.flink.runtime.executiongraph.ExecutionGraphException; +import org.apache.flink.runtime.executiongraph.ExecutionVertex; +import org.apache.flink.runtime.executiongraph.IntermediateResult; +import org.apache.flink.runtime.executiongraph.IntermediateResultPartition; +import org.apache.flink.runtime.executiongraph.JobInformation; +import org.apache.flink.runtime.executiongraph.TaskInformation; +import org.apache.flink.runtime.io.network.partition.ResultPartitionID; +import org.apache.flink.runtime.io.network.partition.ResultPartitionType; +import org.apache.flink.runtime.jobgraph.IntermediateDataSetID; +import org.apache.flink.runtime.shuffle.ShuffleDescriptor; +import org.apache.flink.runtime.shuffle.UnknownShuffleDescriptor; +import org.apache.flink.types.Either; +import org.apache.flink.util.SerializedValue; + +import javax.annotation.Nullable; + +import java.util.ArrayList; +import java.util.Collection; +import java.util.List; +import java.util.Optional; + +/** + * Factory of {@link TaskDeploymentDescriptor} to deploy {@link org.apache.flink.runtime.taskmanager.Task} from {@link Execution}. + */ +public class TaskDeploymentDescriptorFactory { + private final ExecutionAttemptID executionId; + private final int attemptNumber; + private final MaybeOffloaded<JobInformation> serializedJobInformation; + private final MaybeOffloaded<TaskInformation> taskInfo; + private final JobID jobID; + private final boolean lazyScheduling; + private final int subtaskIndex; + private final ExecutionEdge[][] inputEdges; + + private TaskDeploymentDescriptorFactory( + ExecutionAttemptID executionId, + int attemptNumber, + MaybeOffloaded<JobInformation> serializedJobInformation, + MaybeOffloaded<TaskInformation> taskInfo, + JobID jobID, + boolean lazyScheduling, + int subtaskIndex, + ExecutionEdge[][] inputEdges) { + this.executionId = executionId; + this.attemptNumber = attemptNumber; + this.serializedJobInformation = serializedJobInformation; + this.taskInfo = taskInfo; + this.jobID = jobID; + this.lazyScheduling = lazyScheduling; + this.subtaskIndex = subtaskIndex; + this.inputEdges = inputEdges; + } + + public TaskDeploymentDescriptor createDeploymentDescriptor( + ResourceID location, + AllocationID allocationID, + int targetSlotNumber, + @Nullable JobManagerTaskRestore taskRestore, + Collection<ResultPartitionDeploymentDescriptor> producedPartitions) throws Exception { + return new TaskDeploymentDescriptor( + jobID, + serializedJobInformation, + taskInfo, + executionId, + allocationID, + subtaskIndex, + attemptNumber, + targetSlotNumber, + taskRestore, + new ArrayList<>(producedPartitions), + createInputGateDeploymentDescriptors(location, inputEdges, subtaskIndex, lazyScheduling)); + } + + public static TaskDeploymentDescriptorFactory fromExecutionVertex( + ExecutionVertex executionVertex, + int attemptNumber) throws ExecutionGraphException { + ExecutionGraph executionGraph = executionVertex.getExecutionGraph(); + return new TaskDeploymentDescriptorFactory( + executionVertex.getCurrentExecutionAttempt().getAttemptId(), + attemptNumber, + getSerializedJobInformation(executionGraph), + getSerializedTaskInformation(executionVertex.getJobVertex().getTaskInformationOrBlobKey()), + executionGraph.getJobID(), + executionGraph.getScheduleMode().allowLazyDeployment(), + executionVertex.getParallelSubtaskIndex(), + executionVertex.getAllInputEdges()); + } + + private static MaybeOffloaded<JobInformation> getSerializedJobInformation(ExecutionGraph executionGraph) { + Either<SerializedValue<JobInformation>, PermanentBlobKey> jobInformationOrBlobKey = + executionGraph.getJobInformationOrBlobKey(); + if (jobInformationOrBlobKey.isLeft()) { + return new TaskDeploymentDescriptor.NonOffloaded<>(jobInformationOrBlobKey.left()); + } else { + return new TaskDeploymentDescriptor.Offloaded<>(jobInformationOrBlobKey.right()); + } + } + + private static MaybeOffloaded<TaskInformation> getSerializedTaskInformation( + Either<SerializedValue<TaskInformation>, + PermanentBlobKey> taskInfo) { + return taskInfo.isLeft() ? + new TaskDeploymentDescriptor.NonOffloaded<>(taskInfo.left()) : + new TaskDeploymentDescriptor.Offloaded<>(taskInfo.right()); + } + + private static List<InputGateDeploymentDescriptor> createInputGateDeploymentDescriptors( + ResourceID location, + ExecutionEdge[][] inputEdges, + int subtaskIndex, + boolean allowLazyDeployment) throws ExecutionGraphException { + List<InputGateDeploymentDescriptor> inputGates = new ArrayList<>(inputEdges.length); + + for (ExecutionEdge[] edges : inputEdges) { + ShuffleDescriptor[] consumedPartitions = getConsumedPartitionShuffleDescriptors(edges, allowLazyDeployment); + // If the produced partition has multiple consumers registered, we + // need to request the one matching our sub task index. + // TODO Refactor after removing the consumers from the intermediate result partitions + int numConsumerEdges = edges[0].getSource().getConsumers().get(0).size(); + + int queueToRequest = subtaskIndex % numConsumerEdges; + + IntermediateResult consumedIntermediateResult = edges[0].getSource().getIntermediateResult(); + IntermediateDataSetID resultId = consumedIntermediateResult.getId(); + ResultPartitionType partitionType = consumedIntermediateResult.getResultType(); + + inputGates.add(new InputGateDeploymentDescriptor( + resultId, + partitionType, + queueToRequest, + consumedPartitions, + location)); + } + + return inputGates; + } + + private static ShuffleDescriptor[] getConsumedPartitionShuffleDescriptors( + ExecutionEdge[] edges, + boolean allowLazyDeployment) throws ExecutionGraphException { + ShuffleDescriptor[] shuffleDescriptors = new ShuffleDescriptor[edges.length]; + // Each edge is connected to a different result partition + for (int i = 0; i < edges.length; i++) { + shuffleDescriptors[i] = + getConsumedPartitionShuffleDescriptor(edges[i], allowLazyDeployment); + } + return shuffleDescriptors; + } + + private static ShuffleDescriptor getConsumedPartitionShuffleDescriptor( + ExecutionEdge edge, + boolean allowLazyDeployment) throws ExecutionGraphException { + IntermediateResultPartition consumedPartition = edge.getSource(); + Execution producer = consumedPartition.getProducer().getCurrentExecutionAttempt(); + + ExecutionState producerState = producer.getState(); + Optional<ResultPartitionDeploymentDescriptor> consumedPartitionDescriptor = + producer.getProducedPartition(consumedPartition.getPartitionId()); + + ResultPartitionID consumedPartitionId = new ResultPartitionID( + consumedPartition.getPartitionId(), + producer.getAttemptId()); + + return getConsumedPartitionShuffleDescriptor( + consumedPartitionId, + consumedPartition.getResultType(), + consumedPartition.isConsumable(), + producerState, + allowLazyDeployment, + consumedPartitionDescriptor.orElse(null)); + } + + @VisibleForTesting + static ShuffleDescriptor getConsumedPartitionShuffleDescriptor( + ResultPartitionID consumedPartitionId, + ResultPartitionType resultPartitionType, + boolean isConsumable, + ExecutionState producerState, + boolean allowLazyDeployment, + @Nullable ResultPartitionDeploymentDescriptor consumedPartitionDescriptor) throws ExecutionGraphException { + // The producing task needs to be RUNNING or already FINISHED + if ((resultPartitionType.isPipelined() || isConsumable) && Review comment: We have three cases here. I am not sure whether we could change the sequence of checking the conditions. My concern is that let the short easy case happen early, and focus on the complicated common case in the end. That means if the last two `else if` `else` clauses could be checked earlier, we could handle the general logic of getting `ShuffleDescriptor` from `ExecutionEdge`. Then the logic of `PartitionInfo#getKnownConsumedPartitionShuffleDescriptor` could be reused. I have not thought through it yet. Just an initial thought. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
