Package org.apache.flink.runtime.shuffle

Interface ShuffleEnvironment<P extends ResultPartitionWriter,G extends IndexedInputGate>

Type Parameters:
P - type of provided result partition writers
G - type of provided input gates
All Superinterfaces:
AutoCloseable
All Known Implementing Classes:
NettyShuffleEnvironment
public interface ShuffleEnvironment<P extends ResultPartitionWriter,G extends IndexedInputGate> extends AutoCloseable
Interface for the implementation of shuffle service local environment.

Input/Output interface of local shuffle service environment is based on memory Buffers. A producer can write shuffle data into the buffers, obtained from the created ResultPartitionWriters and a consumer reads the buffers from the created InputGates.

Lifecycle management.

The interface contains method's to manage the lifecycle of the local shuffle service environment:

  1. start() must be called before using the shuffle service environment.
  2. AutoCloseable.close() is called to release the shuffle service environment.

Shuffle Input/Output management.

Result partition management.

The interface implements a factory of result partition writers to produce shuffle data: createResultPartitionWriters(org.apache.flink.runtime.shuffle.ShuffleIOOwnerContext, java.util.List<org.apache.flink.runtime.deployment.ResultPartitionDeploymentDescriptor>). The created writers are grouped per owner. The owner is responsible for the writers' lifecycle from the moment of creation.

Partitions are fully released in the following cases:

  1. ResultPartitionWriter.fail(Throwable) and ResultPartitionWriter.close() are called if the production has failed.
  2. for PIPELINED partitions if there was a detected consumption attempt and it either failed or finished after the bounded production has been done (ResultPartitionWriter.finish() and ResultPartitionWriter.close() have been called). Only one consumption attempt is ever expected for the PIPELINED partition at the moment so it can be released afterwards.
  3. if the following methods are called outside of the producer thread:
    1. ShuffleMaster.releasePartitionExternally(ShuffleDescriptor)
    2. and if it occupies any producer local resources (ShuffleDescriptor.storesLocalResourcesOn()) then also releasePartitionsLocally(Collection)
    e.g. to manage the lifecycle of BLOCKING result partitions which can outlive their producers. The BLOCKING partitions can be consumed multiple times.

The partitions, which currently still occupy local resources, can be queried with getPartitionsOccupyingLocalResources().

Input gate management.

The interface implements a factory for the input gates: createInputGates(org.apache.flink.runtime.shuffle.ShuffleIOOwnerContext, org.apache.flink.runtime.io.network.partition.PartitionProducerStateProvider, java.util.List<org.apache.flink.runtime.deployment.InputGateDeploymentDescriptor>). The created gates are grouped per owner. The owner is responsible for the gates' lifecycle from the moment of creation.

When the input gates are created, it can happen that not all consumed partitions are known at that moment e.g. because their producers have not been started yet. Therefore, the ShuffleEnvironment provides a method updatePartitionInfo(org.apache.flink.runtime.executiongraph.ExecutionAttemptID, org.apache.flink.runtime.executiongraph.PartitionInfo) to update them externally, when the producer becomes known. The update mechanism has to be threadsafe because the updated gate can be read concurrently from a different thread.

  • Method Details

    • start

      int start() throws IOException
      Start the internal related services before using the shuffle service environment.
      Returns:
      a port to connect for the shuffle data exchange, -1 if only local connection is possible.
      Throws:
      IOException

    • createShuffleIOOwnerContext

      ShuffleIOOwnerContext createShuffleIOOwnerContext(String ownerName, ExecutionAttemptID executionAttemptID, org.apache.flink.metrics.MetricGroup parentGroup)
      Create a context of the shuffle input/output owner used to create partitions or gates belonging to the owner.

      This method has to be called only once to avoid duplicated internal metric group registration.

      Parameters:
      ownerName - the owner name, used for logs
      executionAttemptID - execution attempt id of the producer or consumer
      parentGroup - parent of shuffle specific metric group
      Returns:
      context of the shuffle input/output owner used to create partitions or gates belonging to the owner

    • createResultPartitionWriters

      List<P> createResultPartitionWriters(ShuffleIOOwnerContext ownerContext, List<ResultPartitionDeploymentDescriptor> resultPartitionDeploymentDescriptors)
      Factory method for the ResultPartitionWriters to produce result partitions.

      The order of the ResultPartitionWriters in the returned collection should be the same as the iteration order of the passed resultPartitionDeploymentDescriptors.

      Parameters:
      ownerContext - the owner context relevant for partition creation
      resultPartitionDeploymentDescriptors - descriptors of the partition, produced by the owner
      Returns:
      list of the ResultPartitionWriters

    • releasePartitionsLocally

      void releasePartitionsLocally(Collection<ResultPartitionID> partitionIds)
      Release local resources occupied by the given partitions.

      This is called for partitions which occupy resources locally (can be checked by ShuffleDescriptor.storesLocalResourcesOn()).

      Parameters:
      partitionIds - identifying the partitions to be released

    • getPartitionsOccupyingLocalResources

      Collection<ResultPartitionID> getPartitionsOccupyingLocalResources()
      Report partitions which still occupy some resources locally.
      Returns:
      collection of partitions which still occupy some resources locally and have not been released yet.

    • getMetricsIfPartitionOccupyingLocalResource

      default Optional<ShuffleMetrics> getMetricsIfPartitionOccupyingLocalResource(ResultPartitionID partitionId)
      Get metrics of the partition if it still occupies some resources locally and have not been released yet.
      Parameters:
      partitionId - the partition id
      Returns:
      An Optional of ShuffleMetrics, if found, of the given partition

    • createInputGates

      List<G> createInputGates(ShuffleIOOwnerContext ownerContext, PartitionProducerStateProvider partitionProducerStateProvider, List<InputGateDeploymentDescriptor> inputGateDeploymentDescriptors)
      Factory method for the InputGates to consume result partitions.

      The order of the InputGates in the returned collection should be the same as the iteration order of the passed inputGateDeploymentDescriptors.

      Parameters:
      ownerContext - the owner context relevant for gate creation
      partitionProducerStateProvider - producer state provider to query whether the producer is ready for consumption
      inputGateDeploymentDescriptors - descriptors of the input gates to consume
      Returns:
      list of the InputGates

    • updatePartitionInfo

      boolean updatePartitionInfo(ExecutionAttemptID consumerID, PartitionInfo partitionInfo) throws IOException, InterruptedException
      Update a gate with the newly available partition information, previously unknown.
      Parameters:
      consumerID - execution id to distinguish gates with the same id from the different consumer executions
      partitionInfo - information needed to consume the updated partition, e.g. network location
      Returns:
      true if the partition has been updated or false if the partition is not available anymore.
      Throws:
      IOException - IO problem by the update
      InterruptedException - potentially blocking operation was interrupted