org.apache.flink.runtime.shuffle

接口 ShuffleEnvironment<P extends ResultPartitionWriter,G extends IndexedInputGate>

类型参数:

P - type of provided result partition writers

G - type of provided input gates

所有超级接口:

AutoCloseable

所有已知实现类:

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.

方法概要

限定符和类型	方法和说明
List<G>	createInputGates(ShuffleIOOwnerContext ownerContext, PartitionProducerStateProvider partitionProducerStateProvider, List<InputGateDeploymentDescriptor> inputGateDeploymentDescriptors) Factory method for the InputGates to consume result partitions.
List<P>	createResultPartitionWriters(ShuffleIOOwnerContext ownerContext, List<ResultPartitionDeploymentDescriptor> resultPartitionDeploymentDescriptors) Factory method for the ResultPartitionWriters to produce result partitions.
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.
Collection<ResultPartitionID>	getPartitionsOccupyingLocalResources() Report partitions which still occupy some resources locally.
void	releasePartitionsLocally(Collection<ResultPartitionID> partitionIds) Release local resources occupied by the given partitions.
int	start() Start the internal related services before using the shuffle service environment.
boolean	updatePartitionInfo(ExecutionAttemptID consumerID, PartitionInfo partitionInfo) Update a gate with the newly available partition information, previously unknown.

从接口继承的方法 java.lang.AutoCloseable

close

方法详细资料

start

int start()
   throws IOException

Start the internal related services before using the shuffle service environment.

返回:

a port to connect for the shuffle data exchange, -1 if only local connection is possible.

抛出:

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.

参数:

ownerName - the owner name, used for logs

executionAttemptID - execution attempt id of the producer or consumer

parentGroup - parent of shuffle specific metric group

返回:

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.

参数:

ownerContext - the owner context relevant for partition creation

resultPartitionDeploymentDescriptors - descriptors of the partition, produced by the owner

返回:

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()).

参数:

partitionIds - identifying the partitions to be released

getPartitionsOccupyingLocalResources

Collection<ResultPartitionID> getPartitionsOccupyingLocalResources()

Report partitions which still occupy some resources locally.

返回:

collection of partitions which still occupy some resources locally and have not been released yet.

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.

参数:

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

返回:

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.

参数:

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

返回:

true if the partition has been updated or false if the partition is not available anymore.

抛出:

IOException - IO problem by the update

InterruptedException - potentially blocking operation was interrupted