Status

Current state: Under Discussion

Discussion thread: here

JIRA: KAFKA-5488

Motivation

KStream#branch method uses varargs to supply predicates and returns array of streams ('Each stream in the result array corresponds position-wise (index) to the predicate in the supplied predicates').

This is poor API design that makes building branches very inconvenient because of 'impedance mismatch' between arrays and generics in Java language.

• In general, the code have poor cohesion: we need to define predicates in one place, and respective stream processors in another place of code. In case of change we must remember to edit two pieces of code.
• If the number of predicates is predefined, this method forces us to use 'magic numbers' to extract the right branch from the result (see examples here).
• If we need to build branches dynamically (e. g. one branch per enum value) we inevitably have to deal with 'generic arrays' and 'unchecked typecasts'.

Public Interfaces

In accordance with KStreams DSL Grammar, we introduce the following new elements:

• split DSLOperation
• BranchedKStream DSLObject with following DSLOperations:
  • branch
  • defaultBranch
  • noDefaultBranch
• Branched DSLParameter.

Description

1. The split(Named named) operation returns BranchedKStream<K,V>. Named parameter is needed so one can name the branch operator itself, and then all the branches might get index-suffixed names built from the branch operator name.

The overloaded parameterless alternative split() is also available.

2. BranchedKStream has the following methods:

• BranchedKStream<K,V> branch(Predicate<? super K, ? super V> predicate, Branched<K,V> branched) -- creates a branch for messages that match the predicate and returns this in order to facilitate method chaining.
• Map<String, KStream<K,V>> defaultBranch(Branched<K,V> branched) -- creates a default branch (for messages not intercepted by other branches) and returns the dictionary of named KStreams.
• Map<String, KStream<K,V>> noDefaultBranch() -- returns the dictionary of named KStreams.

Both branch and defaultBranch operations also have overloaded alternatives without the Branched parameter.

3. Branched parameter extends NamedOperation and has the following static methods:

• as(String name) -- sets the name of the branch (auto-generated by default, when split operation is named, then the names are index-suffixed).
• with(Function<? super KStream<? super K, ? super V>, ? extends KStream<? extends K, ? extends V>> chain) -- sets an operation with a given branch. By default, it is an s->s identity function. Can be complex, like s->s.mapValues...., a composition of functions etc.
• with(Function<? super KStream<? super K, ? super V>, ? extends KStream<? extends K, ? extends V>> chain, String name) -- sets both operation and a name.

The Map returned by defaultBranch/noDefaultBranch allows us to collect all the KStream branch objects in single scope. The branches collected are the results of transformations defined by `chain` functions. If a function returns `null`, its result is omitted.

Simple Example: Direct Branch Consuming

In many cases we do not need to have a single scope for all the branches, each branch being processed completely independently from others. Then we can use 'consuming' lambdas or method references in Branched parameter:

source.split()
.branch((key, value) -> value.contains("A"), Branched.with(ks->{ks.to("A"); return ks;}))
.branch((key, value) -> value.contains("B"), Branched.with(ks->{ks.to("B"); return ks;}))
.defaultBranch(Branched.withJavaConsumer(ks->{ks.to("C"); return ks;}));

More Complex Example: Merging Branches

In other cases we want to combine branches again after splitting. The map returned by defaultBranch /noDefaultBranch methods provides  access to the branches in the same scope:

Map<String, KStream<String, String>> branches = source.split()
  .branch((key, value) -> value == null, 
          Branched.with(s->s.mapValues(v->"NULL"), "null")
  .defaultBranch(
          Branched.as("non-null"));  
  
branches.get("non-null")
 .merge(branches.get("null"));

Dynamic Branching

There is also a case when one might need to create branches dynamically, e. g. one per enum value. This can be implemented the following way:

BranchedKStream branched = stream.split();
for (RecordType recordType : RecordType.values())
  branched.branch((k, v) -> v.getRecType() == recordType,
    Branched.with(recordType::processRecords));

This is why 'starting' split() operation is necessary and it is better to have it rather than add new branch method to KStream directly.

Otherwise we should treat the first iteration separately, and the code for dynamic branching becomes cluttered:

RecordType[] recordTypes = RecordType.values();
if (recordTypes.length != 0) {
  BranchedKStream branched = stream.
    branch((k, v) -> v.getRecType() == recordTypes[0],
       Branched.with(recordType::processRecords));

  for (int i = 1; i < recordTypes.length; i++) 
    branched.branch((k, v) -> v.getRecType() == recordTypes[i],
      Branched.with(recordType::processRecords));
}

Proposed Changes

1. Add the following methods to KStream:

BranchedKStream<K,V> split();
BranchedKStream<K,V> split(Named n);

2. Deprecate the existing KStream#branch method.

3. Add and implement the following Branched class:

class Branched<K, V> implements Named<Branched<K,V>> {
    static Branched<K, V> as(String name);
    static Branched<K, V> with(Function<? super KStream<? super K, ? super V>, ? extends KStream<? extends K, ? extends V>> chain);
    static Branched<K, V> with(Function<? super KStream<? super K, ? super V>, ? extends KStream<? extends K, ? extends V>> chain, String name);
}

Add and implement the following BranchedKStream interface:

interface BranchedKStream<K, V> {
    BranchedKStream<K, V> branch(Predicate<? super K, ? super V> predicate);
    BranchedKStream<K, V> branch(Predicate<? super K, ? super V> predicate, Branched<K, V> branched);
    Map<String, KStream<K, V>> defaultBranch(Branched<K, V> branched);
    Map<String, KStream<K, V>> defaultBranch();
    Map<String, KStream<K, V>> noDefaultBranch();
}

(See https://github.com/apache/kafka/pull/6512 for a very rough draft).

Compatibility, Deprecation, and Migration Plan

The proposed change is backwards compatible.

The old KStreams#branch method should be deprecated.

Rejected Alternatives

1. A KStreamsBrancher class that works the same way, but does not require KStream interface modification:

new KafkaStreamsBrancher<String, String>()
   .branch((key, value) -> value.contains("A"), ks->ks.to("A"))
   .branch((key, value) -> value.contains("B"), ks->ks.to("B"))
   //default branch should not necessarily be defined in the end!
   .defaultBranch(ks->ks.to("C"))
   .onTopOf(builder.stream("source"));

Rejected because of violation of method-chaining (new auxiliary object is needed).

2.

source
   .split()
   .branch((key, value) -> value.contains("A"), ks->ks.to("A"))
   .branch((key, value) -> value.contains("B"), ks->ks.to("B"))
   .defaultBranch(ks->ks.to("C"));

Here the new KStream#branch() method returns KBranchedStream<K, V> object, which, in turn, contains `branch` and `defaultBranch` methods. This is critical that KStream consumers in .branch methods should be invoked immediately during the `branch` methods invocation. This is necessary for the case when we need to gather the streams that were defined in separate scopes back into one scope using auxiliary object:

@Setter
class CouponIssuer{
   private KStream<....> coffePurchases;
   private KStream<....> electronicsPurchases;

   KStream<...> coupons(){
       return coffePurchases.join(electronicsPurchases...)...
  }
}

CouponIssuer couponIssuer = new CouponIssuer();

transactionStream.branch()
     .branch(predicate1, couponIssuer::setCoffePurchases)
     .branch(predicate2, couponIssuer::setElectronicsPurchases);

KStream<..> coupons = couponIssuer.coupons();

This was rejected because of the difficulty of having branches in the same scope.