org.apache.flink.streaming.api.connector.sink2

Class GlobalCommitterOperator<CommT,GlobalCommT>

java.lang.Object

org.apache.flink.streaming.api.operators.AbstractStreamOperator<Void>

org.apache.flink.streaming.api.connector.sink2.GlobalCommitterOperator<CommT,GlobalCommT>

All Implemented Interfaces:

Serializable, org.apache.flink.api.common.state.CheckpointListener, Input<CommittableMessage<CommT>>, KeyContext, KeyContextHandler, OneInputStreamOperator<CommittableMessage<CommT>,Void>, SetupableStreamOperator<Void>, StreamOperator<Void>, StreamOperatorStateHandler.CheckpointedStreamOperator, YieldingOperator<Void>

@Internal
public class GlobalCommitterOperator<CommT,GlobalCommT>
extends AbstractStreamOperator<Void>
implements OneInputStreamOperator<CommittableMessage<CommT>,Void>

Implements the GlobalCommitter.

This operator usually trails behind a CommitterOperator. In this case, the global committer will receive committables from the committer operator through processElement(StreamRecord). Once all committables from all subtasks have been received, the global committer will commit them. This approach also works for any number of intermediate custom operators between the committer and the global committer in a custom post-commit topology.

That means that the global committer will not wait for notifyCheckpointComplete(long). In many cases, it receives the callback before the actual committables anyway. So it would effectively globally commit one checkpoint later.

However, we can leverage the following observation: the global committer will only receive committables iff the respective checkpoint was completed and upstream committers received the notifyCheckpointComplete(long). So by waiting for all committables of a given checkpoint, we implicitly know that the checkpoint was successful and the global committer is supposed to globally commit.

Note that committables of checkpoint X are not checkpointed in X because the global committer is trailing behind the checkpoint. They are replayed from the committer state in case of an error. The state only includes incomplete checkpoints coming from upstream committers not receiving notifyCheckpointComplete(long). All committables received are successful.

In rare cases, the GlobalCommitterOperator may not be connected (in)directly to a committer but instead is connected (in)directly to a writer. In this case, the global committer needs to perform the 2PC protocol instead of the committer. Thus, we absolutely need to use notifyCheckpointComplete(long) similarly to the CommitterOperator. Hence, commitOnInput is set to false in this case. In particular, the following three prerequisites must be met:

• No committer is upstream of which we could implicitly infer notifyCheckpointComplete(long) as sketched above.
• The application runs in streaming mode.
• Checkpointing is enabled.

In all other cases (batch or upstream committer or checkpointing is disabled), the global committer commits on input.

See Also:

Serialized Form

Field Summary

Fields inherited from class org.apache.flink.streaming.api.operators.AbstractStreamOperator

chainingStrategy, config, lastRecordAttributes1, lastRecordAttributes2, latencyStats, LOG, metrics, output, processingTimeService, stateHandler, stateKeySelector1, stateKeySelector2, timeServiceManager

Constructor Summary

Constructors

Constructor and Description
GlobalCommitterOperator(org.apache.flink.util.function.SerializableSupplier<org.apache.flink.api.connector.sink2.Committer<CommT>> committerFactory, org.apache.flink.util.function.SerializableSupplier<org.apache.flink.core.io.SimpleVersionedSerializer<CommT>> committableSerializerFactory, boolean commitOnInput)

Method Summary

Modifier and Type	Method and Description
void	initializeState(org.apache.flink.runtime.state.StateInitializationContext context) Stream operators with state which can be restored need to override this hook method.
void	notifyCheckpointComplete(long checkpointId)
void	processElement(StreamRecord<CommittableMessage<CommT>> element) Processes one element that arrived on this input of the MultipleInputStreamOperator.
void	setup(StreamTask<?,?> containingTask, StreamConfig config, Output<StreamRecord<Void>> output) Initializes the operator.
void	snapshotState(org.apache.flink.runtime.state.StateSnapshotContext context) Stream operators with state, which want to participate in a snapshot need to override this hook method.

Methods inherited from class org.apache.flink.streaming.api.operators.AbstractStreamOperator

close, finish, getChainingStrategy, getContainingTask, getCurrentKey, getExecutionConfig, getInternalTimerService, getKeyedStateBackend, getKeyedStateStore, getMetricGroup, getOperatorConfig, getOperatorID, getOperatorName, getOperatorStateBackend, getOrCreateKeyedState, getPartitionedState, getPartitionedState, getProcessingTimeService, getRuntimeContext, getStateKeySelector1, getStateKeySelector2, getTimeServiceManager, getUserCodeClassloader, hasKeyContext1, hasKeyContext2, initializeState, isUsingCustomRawKeyedState, notifyCheckpointAborted, open, prepareSnapshotPreBarrier, processLatencyMarker, processLatencyMarker1, processLatencyMarker2, processRecordAttributes, processRecordAttributes1, processRecordAttributes2, processWatermark, processWatermark1, processWatermark2, processWatermarkStatus, processWatermarkStatus1, processWatermarkStatus2, reportOrForwardLatencyMarker, setChainingStrategy, setCurrentKey, setKeyContextElement1, setKeyContextElement2, setMailboxExecutor, setProcessingTimeService, snapshotState, useSplittableTimers

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Methods inherited from interface org.apache.flink.streaming.api.operators.OneInputStreamOperator

setKeyContextElement

Methods inherited from interface org.apache.flink.streaming.api.operators.StreamOperator

close, finish, getMetricGroup, getOperatorAttributes, getOperatorID, initializeState, open, prepareSnapshotPreBarrier, setKeyContextElement1, setKeyContextElement2, snapshotState

Methods inherited from interface org.apache.flink.api.common.state.CheckpointListener

notifyCheckpointAborted

Methods inherited from interface org.apache.flink.streaming.api.operators.KeyContext

getCurrentKey, setCurrentKey

Methods inherited from interface org.apache.flink.streaming.api.operators.Input

processLatencyMarker, processRecordAttributes, processWatermark, processWatermarkStatus

Methods inherited from interface org.apache.flink.streaming.api.operators.KeyContextHandler

hasKeyContext

Constructor Detail

GlobalCommitterOperator

public GlobalCommitterOperator(org.apache.flink.util.function.SerializableSupplier<org.apache.flink.api.connector.sink2.Committer<CommT>> committerFactory,
                               org.apache.flink.util.function.SerializableSupplier<org.apache.flink.core.io.SimpleVersionedSerializer<CommT>> committableSerializerFactory,
                               boolean commitOnInput)

Method Detail

setup

public void setup(StreamTask<?,?> containingTask,
                  StreamConfig config,
                  Output<StreamRecord<Void>> output)

Description copied from interface: SetupableStreamOperator

Initializes the operator. Sets access to the context and the output.

Specified by:

setup in interface SetupableStreamOperator<Void>

Overrides:

setup in class AbstractStreamOperator<Void>

snapshotState

public void snapshotState(org.apache.flink.runtime.state.StateSnapshotContext context)
                   throws Exception

Description copied from class: AbstractStreamOperator

Stream operators with state, which want to participate in a snapshot need to override this hook method.

Specified by:

snapshotState in interface StreamOperatorStateHandler.CheckpointedStreamOperator

Overrides:

snapshotState in class AbstractStreamOperator<Void>

Parameters:

context - context that provides information and means required for taking a snapshot

Throws:

Exception

initializeState

public void initializeState(org.apache.flink.runtime.state.StateInitializationContext context)
                     throws Exception

Description copied from class: AbstractStreamOperator

Stream operators with state which can be restored need to override this hook method.

Specified by:

initializeState in interface StreamOperatorStateHandler.CheckpointedStreamOperator

Overrides:

initializeState in class AbstractStreamOperator<Void>

Parameters:

context - context that allows to register different states.

Throws:

Exception

notifyCheckpointComplete

public void notifyCheckpointComplete(long checkpointId)
                              throws Exception

Specified by:

notifyCheckpointComplete in interface org.apache.flink.api.common.state.CheckpointListener

Overrides:

notifyCheckpointComplete in class AbstractStreamOperator<Void>

Throws:

Exception

processElement

public void processElement(StreamRecord<CommittableMessage<CommT>> element)
                    throws Exception

Description copied from interface: Input

Processes one element that arrived on this input of the MultipleInputStreamOperator. This method is guaranteed to not be called concurrently with other methods of the operator.

Specified by:

processElement in interface Input<CommittableMessage<CommT>>

Throws:

Exception