Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
With FLIP-143 we introduced the unified Sink API to make it easier for connector developers to support batch and streaming scenarios. While developing the unified Sink API it was already noted that the unified Sink API might not be flexible enough to support all scenarios from the beginning.
With this document, we mainly focus on the
small-file-compaction problem which is one of the not yet covered scenarios. The problem manifests when using the
FileSink and you want to guarantee a certain file size and not end up with a lot of small files (i.e. one file per parallel subtask). It becomes very important for modern data lake use cases where usually the data is written in some kind of columnar format (Parquet or ORC) and larger files are recommended to amplify the read performance. A few examples are the Iceberg sink , Delta lake , or Hive. For Hive in particular we already have a sink that supports compaction but it is not generally applicable and only available in Flink’s Table API .
The goal of this document is to extend the unified Sink API to broaden the spectrum of supported scenarios and fix the
After collecting more feedback we came up with the following design which incorporates the design of a Sink V2. Over the last few releases, we learned from our users that the abstraction of having a writer and a committer to implement a two-phase commit is very helpful but the GlobalCommitter in its current design cannot fulfill all necessary duties. Unfortunately, there is not an easy way to replace/remove the GlobalCommitter from the existing interfaces without breaking already built sinks due to the typed parameters of the interface. Therefore we propose Sink V2 that subsumes the current Sink interface and offers the same functionality.
The Sink V2 uses a mix-in approach to reflect the different setup options. We will create a separate package inside of flink-core to distinguish between Sink V1 and Sink V2.
First, we explain how the Sink V2 subsumes the Sink V1. In general, sinks that do not use the GlobalCommitter should have it very easy to migrate to the new interfaces because all methods of the Committer and SinkWriter still exist and are unchanged.
We reduced the implementation overhead in Sink V2 and split the Sink V1 into three mixin interfaces providing different functionalities. Sink developers can now decide based on their requirements which sink interfaces they need to combine to implement their sinks.
To overcome the limitations of Sink V1 we offer different hooks to insert custom topologies into the sink. In contrary, to the basic Sink we mark these interfaces as experimental because they are introducing a new concept.
We also introduce two common wrappers around the Sink Committables to ease the implementation of the topologies based on committables.
In these topologies, the committable will always be part of the CommittableWithLinage to have a notion to which checkpoint it belongs and from which subtask it was sent. In this case, it can be either the SinkWriter or Committer depending on which topology is used (pre- or post-commit). Moreover, we introduce the CommittableSummary that is sent once as the last message of this checkpoint. Down-stream operators receiving this message can be sure that the subtask the message was coming from does not send more messages in this checkpoint.
To replace the GlobalCommitter in existing sinks we offer a utility that simulates the GlobalCommitter with a post committing topology.
We envision the implementation of certain connectors can be done with the following interfaces.
Hive and Iceberg will leverage small file compaction. Though both connectors implement it differently. Hive will try to do a best-effort compaction before actually making the files visible in the metastore because it does not support updates. On the other hand, the Iceberg connector writes the files immediately and the post topology will take of compacting the already written files and updating the file log after the compaction.
Since the DataStream class is not part of the flink-core module all advanced Sink interfaces are part of the flink-streaming-java. We mitigate the problem by adding enough documentation i.e. in the doc strings that users are aware of the split.
Compatibility, Deprecation, and Migration Plan
As part of this FLIP, we will migrate all sinks that are currently part of the Flink main repository to the Sink V2 interfaces (i.e. Async, Kafka, Elasticsearch, File). We will not break the existing Sink V1 interfaces because they are already used within the community but we will make them
@PublicEvolving and immediately deprecate them. Our plan is that we keep the Sink V1 interfaces at least until the next major version (2.0) of Flink.
Alternative 1 Global Sink Coordinator:
The idea is similar to the SourceCoordinator which was implemented as part of FLIP-27 for the unified Source interface. On the contrary, to the unified source interface the Sink coordinator is optional and not all sinks must implement it.
The global sink coordinator solves the
small-file-compaction problem by combining committables (files with sizing information) together until a certain threshold is reached. Afterward the combined committable is forwarded to the committers that do the actual file merging and write it to the final location.
Extend the unified Sink interface by allowing to create sink coordinator similar to the `SourceEnumerator` . The creation is optional so the forwarding of the committable to the Job Manager will also only happen if the connector specifies one. In general, the behaviour of sending the committables does not change (sending them on preSnapshot) and the sink coordinator needs to write them into the checkpoint.
Unfortunately, checkpoint for the sink coordinator is not easily possible because the coordinators are checkpointed before the operators in the pipeline. In the case of the sources, this works well because the source coordinator is basically the first operator of the pipeline. For sinks, it means that the sink coordinator would be snapshotting its state before the sink writers leading to potential data loss. Making the sink coordinator stateful would require first changes to the coordinator framework around how they are doing checkpoints.
Another limitation is that committables to the sink coordinators are sent via RPC calls which restrict the payload size. If the committables contain data this will probably not be possible anymore with this approach.
Alternative 2 Committable Aggregator Operator:
This approach is comparable to the current implementation in Flink’s Table API to solve the
small-file-compaction problem  but we make it more extensible.
The committable aggregators are optional operators between the sink writers and committers. The user can define the key for the exchange to the comittable aggregator, the parallelism of the committable aggregator, and the key for exchange from the committable aggregators to the committers.
We propose to extend the unified Sink interface with the following changes. The changes to the unified sink should be backward compatible because only a new method with a default implementation is added to the sink interface.
In the case of the
small-file-compaction the problem, we would choose a constant key for the exchange between sink writers and the committable aggregator. This transfers all committables to exactly one committable aggregator that is responsible for only forwarding a new committable, which contains the locations of the files to merge if configured file size is reached, or for batch execution when endOfInput is triggered. The committer operators will merge the files and write the final files.
The committable aggregator can be checkpointed similarly to the current SinkWriter operator. The combined committables are emitted when receiving the snapshot barrier and the outstanding commitables that haven't reached the expected size are written to state.
In addition, the approach also supports compaction across multiple checkpoints if the committables in on checkpoint do not meet the required size.
A first example POC can be found here.
Although the parallelism and the exchanges between the committable aggregator and the sink writer/committer are configurable new use cases for the unified sink might require more flexibility. In the future, we may want to add a new operator supporting a new use case. Over time this can lead to a fragmented interface with a lot of different options which may or may not be compatible.
Alternative 3 Custom sink topology:
In general, the most important part of a sink is how the commit mechanism works because usually data which is already committed is hard to revert and may lead to problems on the consumer side. The custom sink topology isolated the committer and global committer as the only components which the Sink offers and the former sink writers are only syntactic sugar. Connector developers can implement their own custom topologies which generate committables and are passed to the committers.
By providing the full flexibility of building the topology which is executed as part of the sink it should also be easy to build a topology to support small file compaction.
The plan will be to expose the DataStream incoming to the sink directly by extending the Sink interface to create a custom topology. The default is still using the sink writer topology to not break existing sinks.
During the instantiation of the DataStreamSink, the correct topology is created by configuring a different translator. The DataStreamTopology gives connector developers direct access to the incoming datastream and they only need to ensure that they return a stream with committables.
This approach probably allows implementing all sink scenarios because of its high flexibility. Unfortunately, it also goes a step back considering the batch and stream unification. If users start implementing stateful topologies they need to be familiar with both concepts to fully support them.
Another problem is that currently,
flink-core does not know the
DataStream classes currently which are in
flink-streaming-java to circumvent this issue we could have the implementation in
flink-streaming-java and only provide some base interface in
flink-core. This has a downside in that it adds unnecessary complexity and might confuse users.