Status

...

Page properties

Discussion thread	https:

...

JIRA:

Released:

//lists.apache.org/thread.html/rf09dfeeaf35da5ee98afe559b5a6e955c9f03ade0262727f6b5c4c1e%40%3Cdev.flink.apache.org%3E

Vote thread

https://lists.apache.org/thread.html/r0f71d53810c7970430e5952ec69aa59ac54353de6233844421f753d6%40%3Cdev.flink.apache.org%3E

JIRA

Jira

server	ASF JIRA
serverId	5aa69414-a9e9-3523-82ec-879b028fb15b
key	FLINK-19510

Release

1.12

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Table of Contents

exclude	Status

Motivation

What does the unified sink API mean? The short answer is that write a sink once and use it everywhere. The long answer is that: Flink allows the user to

As discussed in FLIP-131, Flink will deprecate the DataSet API in favor of DataStream API and Table API. Users should be able to use DataStream API to write jobs that support both bounded and unbounded execution modes. However Flink does not provide a sink API to guarantee the exactly once semantics in both bounded and unbounded scenarios, which blocks the unification.

So we want to introduce a new unified sink API which could let the user develop sink once and run it everywhere. Specifically Flink allows the user to

Choose the Choose the different SDK(SQL/Table/DataStream)
Choose the different execution mode(Batch/StreamingStream) according to the scenarios(bounded/unbounded) ~~Choose exactly once or at least once consistency~~

We hope these things(SDK/Execution mode) are transparent to the sink API.

Currently Flink’s sink API is not unified. There are two types of sink API. One is the batch style API `OutputFormats` and the other is the streaming style API `SinkFunction`. Streaming style sink API could not support the bounded scenarios and the batch style sink API could not support the unbounded scenarios. So the sink developer has to implement two versions of the sink for one external system. After introducing the new unified sink API we hope that there is only one implementation sink for one external system.

The document includes three parts: The first part describes the semantics the unified sink API should support. According to the first part the second part proposes two versions of unified sink API. In the last part we introduce two open questions related to the API.

Semantics

This section would talk about which semantics the unified sink API would support. This section includes two parts. The first part describes how the sink API supports the exactly once semantics. The second part mainly talks about that drain and snapshot are two general semantics, which are valid for both bounded and unbounded scenarios.

Consistency

This section tries to figure out the sink developer would be responsible for which part and Flink would be responsible for which part in terms of ensuring consistency. After this we could know what semantics the sink API should have to support the exact once semantics.

From a high-level perspective, we can always abstract a job by reading data from an external system and then writing the processed data to another external system. Sink's responsibility is to write data to external systems. In general, writing data to external systems mainly answers four questions What & How & When & Where to commit the data. Following we would give some explanation about these four questions. After that, we could answer the question asked at the beginning of this part.

What to Commit. The data generated by the job goes through two stages before it could be committed to the external system. The first stage is the preparation. Normally the data produced by the job could not be written/committed to the external system immediately thanks to some conditions that have not been met. For example, in the `StreamingFileSink` the data is first written to an in-progress-file before it could be committed to the filesystem. The second stage is a commitment. In this stage, the data could be committed to an external system. The data in the second stage is called What to commit.

How to Commit. When data is ready to be committed we need a system-specific way to commit the data to the external. Sometimes the sink needs to commit the data by modifying the file’s name. Sometimes the sink needs to submit some meta infos to let data be visible to the end users.

When to Commit. From the users’ perspective, one of the most important requirements is correctness. Normally the user wants the data generated by the job to be committed to the external system once and only once. If we commit the data at the wrong time there might be duplicated data. For example current Flink through restarting the job when there is failover. This always replays some elements.

Where to Commit. This is about how many resources that we need to guarantee exactly once when time is ready. For example If we assume that commit operation will happen at where the committable data is produced the lifecycle of the component that is responsible for producing the committable data would last until the end of the commit operation.

Actually When & Where are all about how to guarantee the exactly once semantics. Now Flink exposes the internal implementation to the sink developer through two interfaces `CheckpointedFunction` & `CheckpointListener`. All the exactly once sinks are coupled with the internal exactly once implementation through implementing the two interfaces to guarantee the exactly once semantics.

For supporting the bounded scenario we find that the mechanism of `CheckpointedFunction` & `CheckpointListener` is not inline with the bounded scenario. In a bounded scenario Flink should guarantee the exact once semantics:

Even if there is only one resource.(Where)
Even if there is no normal checkpoint at all.(When)

It means current streaming style sink implementation is not suitable for the bounded scenario. Flink must introduce some new mechanism for the bounded scenario. However, we might not couple the sink API to another new internal mechanism which guarantees consistency in the bounded scenario. These internal mechanisms should be decoupled with the sink’s API.

In summary the sink API should be responsible for producing What to commit & providing How to commit. Flink should be responsible for guaranteeing the exact semantics. Flink could “optimize” When & Where to commit according to the execution mode and these optimizations should be transparent to the sink developer.

Drain and Snapshot

It seems that

The sink has to deal with Snapshot/Drain and not to deal with the Termination in the unbounded scenario.
The sink should not care about the Snapshot/Drain and only care about the Termination in the bounded scenario

If it is true the sink API could not be unified because it implies that the sink developer has to be aware of the bounded and unbounded scenario. However, one target of the unified sink API is to make bounded/unbounded be transparent to the sink developer.

In general these three concepts(Snapshot/Drain/Termination) are all related with the progress of a job. The progress is a “position” of the dataset that a job has received. It means that the job has already received all the data before that “position”. A job would be executed by the Flink. The progress for Flink means that all the components of the Flink have received the data before the “position”.

Drain means that all the components need to flush all the data before the “position”. For the internal operator/udf this means flushing all the data before the “position” to the downstream operator. For the sink operator this means flushing all the data before the “position” to the external system.

Termination is a special Drain. The requirement is the same. The difference between Drain and Termination is that the “position” is specified by the user or the “position” is specified at the end of the source.

Snapshot is the state of a component at a “position”. Checkpoint is a collection of the snapshot of all the components at a “position”. Compared to the Termination/Drain we could see the Snapshot as a fine grained progress of a job. In general the developer might not care about when the Snapshot happens because that is an implementation of the optimizing the failover. In other words we could choose not to expose the Snapshot to the sink API in theory. However from the perspective of compatibility and performance exposing this to the sink API may be more realatics. Even if we expose the Snapshot to the sink API users should not assume how fine the granularity is. For example in the bounded scenario + streaming mode there might be no regular checkpoint at all because that the job runs too fast. In general how many times a snapshot would happen is just a runtime concept which depends on the specific implementation at the time or scenario.

From the above analysis we could see that

Both the bounded and unbounded scenario have to deal with the Drain. The Termination is not a unique concept to the bounded scenario. It's just a special case of Drain in the bounded scenario. So we should provide the Drain semantics to the sink API.
Snapshot is also available to both the bounded and unbounded scenario. But users should not assume how many snapshots would happen at runtime.

Sink API

From above analysis we could know that there are four semantics the sink API should support.

What to commit
How to commit
SnapshotState
Drain(Flush)

Following we propose two versions of the sink APIs. There is no difference on how to express the four semantics between the two versions:

`Sink` is responsible for creating the `Writer` and `Committer`
The `Writer` is responsible for writing data and handling any potential tmp area used to write yet unstaged data, e.g. in-progress files. As soon as some data is ready to commit, they (or metadata pointing to where the actual data is staged) are shipped to an operator who knows when to commit them.
The `Writer` also responds to the progress notification: Snapshot and Flush.
The `Committer` knows how to commit the data to the destination storage system.

The only difference between the two versions is how to expose the state to the user. The cons & pros are following.

First Alternative

PROS

This alternative has the advantage of not exposing anything about internal state handling to the user. We require the sink to provide us with the state to be persisted and it is up to the underlying runtime to handle them. For example, instead of using union state for the shared state, we could leverage a potential operator coordinator.
We do not expose complex classes such as StateDescriptors, TypeSerializer... (only a simplified version: SimpleVersionedSerializer)
The intent of the #snapshotState is cleaner. The signature tells the user already that it expects/allows some state out of this method.

CONS

It exposes more methods
We get the state to snapshot via a return value from a method, the state must fit into memory. Moreover we must put the whole state into the state backend on each checkpoint instead of putting only the updated values during elements processing

Code Block

language	java
theme	Confluence
firstline	0
title	Sink
linenumbers	true

/**
* The sink interface acts like a factory to create the {@link Writer} and {@link Committer}.
*
* @param <IN>      The type of the sink's input
* @param <CommT>   The type of the committable data
* @param <WriterS> The type of the writer's state
* @param <SharedS> The type of the writer's shared state
*/
public interface Sink<IN, CommT, WriterS, SharedS> extends Serializable {

	/**
	* Create a {@link Writer}.
	* @param context the runtime context
	* @return A new sink writer.
	*/
	Writer<IN, CommT, WriterS, SharedS> createWriter(InitContext context);

	/**
	* Restore a {@link Writer} from the state.
	* @param context the runtime context
	* @param state the previous writers' state
	* @param share the previous writers' shared state
	* @return A sink writer
	*/
	Writer<IN, CommT, WriterS, SharedS> restoreWriter(InitContext context, List<WriterS> state, List<SharedS> share);

	/**
	* Create a {@link Committer}.
	* @return A new sink committer
	*/
	Committer<CommT> createCommitter();

	/**
	* @return a committable serializer
	*/
	SimpleVersionedSerializer<CommT> getCommittableSerializer();

	/**
	* @return a writer's state serializer
	*/
	SimpleVersionedSerializer<WriterS> getStateSerializer();

	/**
	* @return a writer's shared state serializer
	*/
	SimpleVersionedSerializer<SharedS> getSharedStateSerializer();

	/**
	* Providing the runtime information.
	*/
	interface InitContext {

		int getSubtaskId();

		int getAttemptID();

		MetricGroup metricGroup();
	}
}

Code Block

language	java
theme	Confluence
firstline	0
title	Writer
linenumbers	true

/**
* The Writer is responsible for writing data and handling any potential tmp area used to write yet un-staged data, e.g. in-progress files. 
* As soon as some data is ready to commit, they (or metadata pointing to where the actual data is staged) are shipped to an operator who knows when to commit them.
*
* @param <IN> 			The type of the writer's input
* @param <CommT> 		The type of the committable data
* @param <StateT> 		The type of the writer's state
* @param <SharedStateT> The type of the writer's shared state
*/
public interface Writer<IN, CommT, StateT, SharedStateT> {

	/**
	* Add an element to the writer.
	* @param element The input record.
	* @param ctx     The additional information about the input record
	* @param output  The committable data could be shipped to the committer by this.
	* @throws Exception
	*/
	void write(IN element, Context ctx, WriterOutput<CommT> output) throws Exception;

	/**
	* Snapshot the writer's state.
	* @param output  The committable data could be shipped to the committer by this.
	* @return The writer's state.
	*/
	List<StateT> snapshotState(WriterOutput<CommT> output) throws Exception;

	/**
	* Snapshot the writer's shared state.
	* @return The writer's shared state.
	*/
	SharedStateT snapshotSharedState() throws Exception;

	/**
	* Flush all the data which is un-staged to the committer.
	* @param output
	*/
	void flush(WriterOutput<CommT> output) throws Exception;

	interface Context {

		long currentProcessingTime();

		long currentWatermark();

		Long timestamp();
	}
}

Code Block

language	java
theme	Confluence
firstline	0
title	Committer
linenumbers	true

/**
* This interface knows how to commit the data to the external system.
* 
* @param <CommT> The type of the committable data.
*/
public interface Committer<CommT> {
	void commit(CommT committable) throws Exception;
}

Code Block

language	java
theme	Confluence
firstline	0
title	WriterOutput
linenumbers	true

/**
* The {@link Writer} uses this interface to send the committable data to the operator who knows when to commit to the external system.
*
* @param <CommT> The type of the committable data.
*/
public interface WriterOutput<CommT> {

	/**
	* Send the committable data to the operator who knows when to commit.
	*
	* @param committable The data that is ready for committing.
	*/
	void sendToCommit(CommT committable) throws IOException;
}

Second Alternative

PROS

It is more consistent with development of other udf. Whether using the state or how many states it depends on the sink developer himself.
We can e.g. append a single value to the state.

CONS

All the sinks depend on the operator state backend. All the improvements depend on the state backend.
Exposes more complex classes (see approach 1)
Less clean contracts on the restore operations, number of supported states, potential for dangling states when updating the writers implementation

Code Block

language	java
theme	Confluence
firstline	0
title	Sink
linenumbers	true

/**
* The sink interface acts like a factory to create the {@link Writer} and {@link Committer}.
*
* @param <T>       The type of the sink's input.
* @param <CommT>   The type of the committable data.
*/
public interface Sink<T, CommT> extends Serializable {

    /**
	* Create a {@link Writer}.
	* @param context the runtime context
	* @return A new sink writer.
	*/
	Writer<T, CommT> createWriter(InitialContext context) throws Exception;

	/**
	* Create a {@link Committer}.
	* @return A new sink committer
	*/
	Committer<CommT> createCommitter();

	/**
	* @return a committable serializer
	*/
	SimpleVersionedSerializer<CommT> getCommittableSerializer();

    /**
	* Providing the runtime information.
	*/
	interface InitialContext {

		boolean isRestored();

		int getSubtaskIndex();

		OperatorStateStore getOperatorStateStore();

		MetricGroup metricGroup();

	}
}

Code Block

language	java
theme	Confluence
firstline	0
title	Writer
linenumbers	true

/**
* The Writer is responsible for writing data and handling any potential tmp area used to write yet un-staged data, e.g. in-progress files.
* As soon as some data is ready to commit, they (or metadata pointing to where the actual data is staged) are shipped to an operator who knows when to commit them.
*
* @param <T> 		The type of writer's input
* @param <CommT> 	The type of the committable data.
*/
public interface Writer<T, CommT> extends AutoCloseable {

	/**
	* Add an element to the writer.
	* @param t          The input record
	* @param context    The additional information about the input record
	* @param output     The committable data could be shipped to the committer by this
	* @throws Exception
	*/
	void write(T t, Context context, WriterOutput<CommT> output) throws Exception;

	/**
	* Snapshot the state of the writer.
	* @param output The committable data could be shipped to the committer by this.
	*/
	void snapshotState(WriterOutput<CommT> output) throws Exception;

	/**
	* Flush all the data which is un-staged to the committer.
	*/
	void flush(WriterOutput<CommT> output) throws IOException;

    interface Context {

		long currentProcessingTime();

		long currentWatermark();
		
		Long timestamp();
	}
}

Open Questions

There are still two open questions related to the unified sink API:

How does the sink API support to write to the Hive?
Is the sink an operator or a topology?

Compatibility, Deprecation, and Migration Plan

We does not change the current streaming and batch style sink API. So there is no compatibility issue to the exsits sink.
In the long run we might deprecate the old streaming and batch style sink API.
At first we plan to migrate the StreamingFileSink to this new api.

Test Plan

Rejected Alternatives

...

The document includes three parts: The first part describes the semantics the unified sink API should support. According to the first part the second part proposes a new unified sink API. In the last part we introduce two open questions related to the API.