...

Currently, in terms of computing engines, Hudi has mainly integrated deeply with Spark. Apache Flink is a popular streaming processing engine. Integrating Hudi with Flink is a valuable work. This will enable Hudi to embrace more computing engines, and the pluggable design will also make its architecture more flexible and open.

Background

The current design of Hudi is highly dependent on Spark in four modules as shown below:

Image Removed

If we expect Hudi to be decoupled from the computing engine, then we have to do some refactoring to improve the current situation. At a very high level, there are roughly two options:

1. Keep the existing implementation and re-implement all Spark-related capabilities based on Flink (this means we may add four more Flink-related modules);
2. Refactor the current design so that the parts related to Spark are cohesive into a specific module;

We need to rethink the functional design related to Spark so that it can better follow the pluggable design.

Implementation

The implementation contains two sides.

Step 1: decouple Hudi from Spark in core design

There are two key modules that need to redesign, they are: hudi-client and hudi-utlilites.

About hudi-client, We can split hudi-client module into two new modules: hudi-writer-common and hudi-spark. hudi-writer-common will have the HoodieIndex, HoodieTable abstract classes along with IOhandle classes, metrics, exceptions. Index implementations themselves now can move to hudi-spark. HoodieWriteClient and the table classes can also put into hudi-spark module. After this refactoring, we can introduce a new hudi-flink module to package flink specific implementation of the index.

Image Removed

HoodieIndex, as a public interface, should be refactored into engine-independent classes. We can generalize Spark-related types. Like this:

/**
 * Base class for different types of indexes to determine the mapping from uuid.
 *
 * @param <T> The specific type of the {@link HoodieRecordPayload}.
 * @param <DS> The data set or data stream related computation engine.
 * @param <WS> The type of {@link org.apache.hudi.WriteStatus} set.
 * @param <HKS> The type of {@link HoodieKey} set.
 * @param <HKVS> The type of {@link HoodieKey} and value pair set.
 */
public abstract class HoodieGeneralIndex<T extends HoodieRecordPayload, DS, WS, HKS, HKVS> implements Serializable {

  public static <T extends HoodieRecordPayload, DS, WS, HKS, HKVS> HoodieGeneralIndex<T, DS, WS, HKS, HKVS> createIndex(HoodieWriteConfig config) throws HoodieIndexException {
    //...
  }

  /**
   * Checks if the given [Keys] exists in the hoodie table and returns [Key, Option[partitionPath, fileID]] If the
   * optional is empty, then the key is not found.
   */
  public abstract HKVS fetchRecordLocation(HKS hoodieKeys, HoodieTable<T> hoodieTable);

  /**
   * Looks up the index and tags each incoming record with a location of a file that contains the row (if it is actually
   * present).
   */
  public abstract DS tagLocation(DS recordRDD, HoodieTable<T> hoodieTable) throws HoodieIndexException;

  /**
   * Extracts the location of written records, and updates the index.
   * <p>
   * TODO(vc): We may need to propagate the record as well in a WriteStatus class
   */
  public abstract WS updateLocation(WS writeStatusRDD, HoodieTable<T> hoodieTable) throws HoodieIndexException;

}

For the basic index class in Spark context, we can create a HoodieSparkIndex, the definition is:

public class HoodieSparkIndex<T extends HoodieRecordPayload>
    extends HoodieGeneralIndex<
    T,
    JavaRDD<HoodieRecord<T>>,
    JavaRDD<WriteStatus>,
    JavaRDD<HoodieKey>,
    JavaPairRDD<HoodieKey, Option<Pair<String, String>>>> {

  @Override
  public JavaPairRDD<HoodieKey, Option<Pair<String, String>>> fetchRecordLocation(
      JavaRDD<HoodieKey> hoodieKeys, HoodieTable<T> hoodieTable) {
    //...
  }

  @Override
  public JavaRDD<HoodieRecord<T>> tagLocation(JavaRDD<HoodieRecord<T>> recordRDD,
      HoodieTable<T> hoodieTable) throws HoodieIndexException {
    //...
  }

  @Override
  public JavaRDD<WriteStatus> updateLocation(JavaRDD<WriteStatus> writeStatusRDD,
      HoodieTable<T> hoodieTable) throws HoodieIndexException {
    //...
  }

  //...
}

For the basic index class in Flink context, we can create a HoodieFlinkIndex, the definition is:

public class HoodieFlinkIndex<T extends HoodieRecordPayload>
  extends HoodieGeneralIndex<
    T,
    DataStream<HoodieRecord<T>>,
    DataStream<WriteStatus>,
    DataStream<HoodieKey>,
    DataStream<Tuple2<HoodieKey, Option<Tuple2<String, String>>>>> {

  public HoodieFlinkIndex(HoodieWriteConfig config) {
    super(config);
  }

  @Override
  public DataStream<Tuple2<HoodieKey, Option<Tuple2<String, String>>>> fetchRecordLocation(
      DataStream<HoodieKey> hoodieKeys, HoodieTable<T> hoodieTable) {
    return null;
  }

  @Override
  public DataStream<HoodieRecord<T>> tagLocation(DataStream<HoodieRecord<T>> recordRDD,
      HoodieTable<T> hoodieTable) throws HoodieIndexException {
    return null;
  }

  @Override
  public DataStream<WriteStatus> updateLocation(DataStream<WriteStatus> writeStatusRDD,
      HoodieTable<T> hoodieTable) throws HoodieIndexException {
    return null;
  }

  //...
}

About HoodieTable, we can do the same refactor via Java Generic. For example, we can define a HoodieGenericTable like this:

/**
 * Abstract implementation of a HoodieTable.
 * @param <T> The specific type of the {@link HoodieRecordPayload}.
 * @param <EC> The specific context type of the computation engine.
 * @param <WSDS> The specific data set type of the {@link WriteStatus}.
 * @param <P> The specific partition type.
 */
public abstract class HoodieGenericTable<T extends HoodieRecordPayload, EC, WSDS, P> implements Serializable {
    //...
}

Then, we will also introduce HoodieSparkTable and HoodieFlinkTable just like HoodieSparkIndex and HoodieFlinkIndex.

About hudi-utlilites, we use some specific Spark data sources there. So we can either split the core deltastreamer logic as a hudi-deltastreamer-core or hudi-utilities-core and have the Sources themselves live in a separate module as hudi-utilities-spark, hudi-utilities-flink:

Image Removed

Step 2: integrate Hudi with Flink

After step 1, we have decoupled Hudi and spark. Now, we need to implement some functions just like Spark did, e.g. Index.

The implementation of the index feature is one of the parts in the Flink Job DAG. Flink Stateful API can provide the ability of state management. We can store the index via Flink stateful API. From a low-level abstraction, in unbounded streaming, window is a mechanism that split the unbounded stream into bounded stream. We can use the window in Flink to mapping the micro-batch(RDD) in Spark.

Indexing is one of the steps in the writing process which exists in the context of computation and is closely related to our computation engine. Therefore, the implementation of the existing indexes also needs to give corresponding implementations for different computation engines. The whole class diagram is shown below:

Motivation

Currently, Delta Streamer can support “streaming” data ingestion. The "streaming" here is actually a continuous mode, which is a continuous batch processing cycle. The advantage of this feature is that we can make sure about each small-batch, then the small batches are processed continuously at small time intervals to form a stream processing. This design is a bit like the Spark RDD model. But in essence, this is still not regarded as a pure stream ingestion. As streaming computing becomes more and more popular, people are more and more eager to expect data to be processed with lower latency.

When we consider using Flink as Hudi's ingestion framework, the design concepts of Spark RDD and Flink DataStream API are obviously different. Therefore, at some points, we can not fully copy the existing Delta Streamer design.

Design

Next, we will elaborate on the design of Flink-based streaming ingest and writing to Hudi.

Batch abstraction on Flink streams: windows and checkpoints

It needs to be clear that in Hudi’s concept we need to ensure that a batch of records must be atomically written in a table, which also must be guaranteed when implemented via Flink. So, this involves how we define batches in Flink (obviously, considering the performance and the problem of small files that HDFS has been criticized for a long time, we still hope to use the batch-write mechanism). 

In Flink, for a single key, we can use the window as a batch to cut out the stream, which is well known. In a sense, checkpoints are also divided into batches by triggering via time interval, but this is a more obscure concept relative to windows because checkpoints emphasize "points", and the "snapshot" corresponding to the "point". So, in the face of two mechanisms for cutting batches on the stream, how should we choose?

Here we choose checkpoint, reasons will be given later.

Checkpoint mechanism and commit/rollback

A major feature of Hudi is that it provides the semantics of typical database-like transactions on HDFS. This maps the related APIs it provides, such as: commit/rollback. In order to support rollback, we need to save the starting offset corresponding to our batch. This seems to naturally have a mapping relationship with Flink's checkpoint mechanism, because Flink's checkpoint is also for rollback during restore. Another problem is that in Hudi's Spark write path, it is a distributed "batch" rather than a single-partition batch. So, in a sense, we think the concept of applying checkpoints is more appropriate here.

Looking back, it is guaranteed that the characteristics of "atomic" writing are provided by the commit operation. But to ensure performance, our write operations are distributed in parallel. Therefore, in a distributed scenario, we need a coordination mechanism to commit after writing.

This can be achieved by count the number of results from subtasks(e.g. 12 parallelisms, 12 results). when we implement it based on windows, if some subtasks have no input data, it will not emit results to sink. But Relying on Checkpoint, We can introduce a WriteProcessOperator to mock results to send to the sink. In this way, The sink will receive exactly the num of parallelism results from WriteProcessOperator, regardless of data skew.

The design of Flink’s job DAG about writing into Hudi

The above illustrates some of the design considerations in the process of integrating Flink with Hudi. Next, let's describe our design. The DAG of the job can be represented by the following diagram:

Image Added

From the above picture, we can see that it is a pipeline without branches. Let's introduce the purpose of each operator:

• source: used to connect Kafka's message stream. Kafka data will be transformed into HoodieRecord here;
• instant generate operator (customized): used to generate a globally unique instant (each batch of hudi needs to correspond to an instant), its parallelism is 1. Before emitting a new instant, it will check the state of last instant, if it exists and not completed, it will wait until timeout.
• keyBy: partition the data with partitionPath as the key to avoid concurrent write operations to the same partition;
• keyed Process: It carries the main logic of the write path, including index search and file writing operations. If some subtasks have no data flow in, they will send an empty result to the sink;
• sink: a global commit sink, its parallelism is also 1. This wink will count the num of results it received, it will not commit until the num equal the parallelism.

The generation of instant

As mentioned above, the current Flink-based write implementation is very different from the existing Delta Streamer's Spark RDD-based write implementation. It is real streaming processing, not a circulating small batch processing. Therefore, in addition to the different ways of defining "batch" here, we also face the problem of how to generate instants for one batch. For the implementation of Delta Streamer, because the loop is sequential, so it can generate a unique instant, but in real streaming, we must find a way to generate a globally unique instant for each checkpoint.

Here we can introduce an operator with a parallelism of 1 in the pipeline, which will generate an instant when the last one is completed(if not we can should wait, make sure will be only one instant inflight). so that there will be no consistency problems caused by concurrency. Since the timing of our offset saving, writing operations, and commits are all tied to Flink's checkpoint mechanism, then the timing of our instant generation should also be done. For this, we need to extend Flink's operator and rewrite its prepareSnapshotPreBarrier method. This method will be executed firstly, Then the barrier is sent to the downstream, At last, snapshotState method is executed. This ensures that when the snapshotState method is executed downstream, the upstream instant must have been generated.

...

Execution Plan 

draw.io Diagram
border true diagramName execution-plan-rfc13 simpleViewer false width links auto tbstyle top lbox true diagramWidth

...

171 revision

...

1

Rollout/Adoption Plan

None

...