...

Discussion thread: here

JIRA:

...

Each key changing operation in Kafka Streams (selectKey, map, transform, etc.) now leads to automatic repartition before an a stateful aggregation.

But there is quite a large class of key mappings that do not require repartition: every mapping that gives different result for different source partition does not require repartitioning to keep the data locally (see the picture below):

Image Removed

However, repartition is not always necessary if the input stream is already correctly partitioned. In these cases, the automatic repartition brings in additional overhead. As an example, if an input stream comes in partitioned by key1, calling the function selectKey( ... => (key1, metric)).groupByKey  will trigger a repartition today. 

In tickets KAFKA-4835 and KAFKA-10844 the option for canceling the unneeded repartition is requested. Repartition canceling is also needed for the efficient usage of distinct() operator operators proposed in KIP-655: groupBy(...).windowedBy(...).distinct() will always repartition by default, while this is not always needed in practice.

This KIP proposes a new interface for users to optimize the key changing operations (selectKey, map, transform, etc.) in Kafka Streams. 

Public Interfaces

In accordance with KStreams DSL Grammar, we introduce a new parameterless DSLOperation markAsPartitioned() on KStream, which resets keyChangingOperation flag for the upstream node.

Introducing a new parameterless method on KStream rather than adding parameters to all the key changing operations will lead to minimum changes in public APIs.

Concerns

public interface KStream<K, V> {    
	/**
     * Marking the {@code KStream} as partitioned signals the stream is partitioned as intended,
     * and does not require further repartitioning in downstream key changing operations.
     *
     * <p><em>
     *     Note that {@link KStream#markAsPartitioned()} SHOULD NOT be used with interactive query(IQ) or {@link KStream#join}.
     *     For reasons that when repartitions happen, records are physically shuffled by a composite key defined in the stateful operation.
     *     However, if the repartitions were cancelled, records stayed in their original partition by its original key. IQ or joins
     *     assumes and uses the composite key instead of the original key.
     * </p></em>
     *
     * @return a {@code KStream} that will not repartition in subsequent operations.
     */
    KStream<K, V> markAsPartitioned();

}

Proposed Change

In KStreamImpl , set the boolean flag repartitionRequired  to be false when the method is invoked. 

Usage

Example: canceling repartition in a streams aggregation would look like:

stream
	.selectKey( ... => (key1, metric))
	.markAsPartitioned()
	.groupByKey()
	.aggregate()

Concerns

• The
• This operation is unsafe: the wrong usage can lead to undefined behavior. Thus it should be used with care for performance tuning only. This warning should be given in the documentation.
•  The usage of this operation complicates the usage of IQ(Interactive Query) and joins. For reasons that when repartitions happen, records are physically shuffled by a composite key defined in the stateful operation. However, when the repartitions are canceled, records stayed in their original partition by their original key. IQ assumes and uses the composite key instead of the original key. That's when IQ can break downstream. The same applies to joins. 
• In the documentation, we specifically advise against using the interface with IQ or joins. 
• However, a potential solution to support IQ is to provide a . Without repartitioning, the mapped key does not contain the information of the correct partition anymore, since the hash of the mapped key might compute to a wrong instance. However, if we use markAsPartitioned correctly, it is always possible to provide the 'reverse mapping' for the composite key that restores the original key, which can then be used for obtaining the metadata. We can follow up with a change when there is request. 

Compatibility, Deprecation, and Migration Plan

No impact on existing users, no migration is needed.

Possible Alternatives

Option 1: Composite Key

If we don't want to introduce an unsafe operation, we might discuss introducing composite keys as an alternative.

• CompositeKey<H, P> consists of a head and a postfix, and the partition of a composite key is always defined by its 'head' only.
• Also, k and CompositeKey(k, v) must have the same partition for each k.
• We will need to introduce selectCompositeKey operations that will not lead to repartition.

CompositeKey the design will be safe both from the pov of data locality and IQ but adds complexity to the usage. 

Option 2: Optional configuration in Named Operations( Joined  , Grouped , etc)

It would allow us to hit only the relevant parts of the DSL and be less prone to undesired behaviors when it comes to IQ or joins. 

• More generic, can be applied to KTable as well. In comparison, the markAsPartitioned()  approach is targeting the KStreams interface only where it focuses on a specific set of overhead/pain points introduced by repartitionRequired 

• It touches on a larger surface area of the API