...

Current state: Under Discussion

Discussion thread: here

JIRA: here

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

Motivation

Some use-cases require streaming recursionalgorithms are best expressed recursively, which involves would involve piping the output of a pipeline back to the input of an earlier point in the same pipeline. An example of this is graph/tree-traversal, where it can be useful to recursively traverse up a tree as new leaf nodes arrive.Currently, streaming recursion is possible via an explicit topic

This document introduces the concept of streaming recursion, a new DSL operator to succinctly express it, and optimizations that Kafka Streams can make to recursive algorithms that aren't currently possible.

Public Interfaces

The following new method will be introduced:

interface KStream<K, V> {
	KStream<K, V> recursively(UnaryOperator<KStream<K, V>> op);
}

Note: UnaryOperator is java.util.function.UnaryOperator

Proposed Changes

The new recursively method enables users to express recursive algorithms. Consider an example where we count all descendants of each node in a graph:

// <NodeId, ParentId>
KStream<String, String> nodes = builder.stream("nodes");

// <NodeId, ParentId>
KTable<String, String> parents = nodes.toTable();

// <NodeId, Descendants>
KStream<String, Long> updates = builder.stream("node-updates"); // EXPLICIT TOPIC

// update parent count
nodes
    .map((node, parent) -> { KeyValue(parent, 1L) })
    .to("updates")

// increment parents recursively
updates
    .join(parents, (countcount descendants by recursively producing parent records
// 1L is used as a dummy value below, since we will be discarding values when we count the records by key
KTable<String, Long> descendants = nodes
	.map((child, parent) -> { KeyValue(parent, count1L) })   // theemit root"update" nodefor has no parent, soof recursion halts at the root
    .to("updates") // recursively send updates back to explicit topic

// sum updates to produce total descendants for graph nodes
// <NodeId, Descendants>
KTable<String, Long> nodeDescendants = updates
    .groupByKey()
    .reduce((acc, next) -> acc + next)

This approach works (and can be extended to handle out-of-order graph updates), but has the following drawbacks:

1. The node-updates  topic is entirely internal to the Topology, yet it has to be managed explicitly by the user.
   1. This is functionally a repartition  topic, however, because it's explicitly managed, Streams can't automatically delete consumed records.
   2. Consequently, to prevent the topic growing unbounded, users would need to set retention criteria, which risks possible data loss.
2. In scenarios where repartitioning is not required, the explicit recursive topic adds overhead.
3. It also adds some complexity to the user's program, making it more difficult to grok than it needs to be.

A simple, yet ideal, solution would be to extend the DSL API to permit providing a KStream  to the to method. This addresses the above problems:

1. A repartitioning topic, used for recursion, would be automatically created and managed by Kafka Streams.
   1. As this is a repartitioning topic, records would be automatically deleted once consumed, without retention criteria, preventing data loss and reducing storage overhead.
2. In cases where repartitioning is not required, no topic would be created, and the stream would recurse directly from processor to processor, reducing overhead.
3. Less code would be required to implement streaming recursive algorithms, reducing cognitive burden.

Public Interfaces

The following new methods would be introduced:

interface KStream<K, V> {
    void to(KStream<K, V> other);
    void to(KStream<K, V> other, Produced<K, V> produced);
}

Proposed Changes

The above algorithm would instead look like this:

// <NodeId, ParentId>
KStream<String, String> nodes = builder.stream("nodes");

// <NodeId, ParentId>
KTable<String, String> parents = nodes.toTable();

// update parent count
KStream<String, Long> updates = nodes
    .map((nodenew node
	.recursively((updates) -> {                        // recursively emit "updates" for each ancestor of the parent
		// emit a new update for the parent of this node
	 	// the root node has no parent, so recursion terminates at the root node
		updates
			.join(parents, (count, parent) -> { KeyValue(parent, 1L) })

// increment parents recursively
updates
    .join(parents, (count			.map((child, parent) -> { KeyValue(parent, count1L) }) // the root node has no parent, so recursion halts at the root
    .to(updates)

// sum updates to produce total descendants for graph nodes
// <NodeId, Descendants>
KTable<String, Long> nodeDescendants = updates
    
	})
	.groupByKey()
    .reduce((acc, next) -> acc + next)

This method can be used for recursion, as above, but can also be used for wiring up the process graph "in reverse", if desired.

The optional Produced argument controls data written to/read from the underlying repartitioning topic if one is automatically created.

	.count()

Note: for simplicity, this example assumes that graph nodes arrive and are processed in-order; i.e. parent nodes are always processed before children.

The recursively method applies input records to its op argument. The results are then both emitted as a result of recursively and also fed back in to the op KStream.

Restrictions:

• op cannot be UnaryOperator.identity, or an equivalent function that simply returns its argument unmodified - this would produce an infinite recursive loop, since there's no opportunity refine the output to break out of the loop.
• op MUST "terminate"; that is, it must have some condition which eventually prevents further recursion of a record. In our example here, the terminating condition is the join, since the root node of our graph will have no parent, so the join will produce no output for the root node.
  • We can attempt to detect "definitely non-terminating" arguments by failing to detect operations that can cause the stream to terminate (e.g. filter, join, flatMap, etc.) in the process graph produced by the function.
  • We cannot guarantee that a function that includes terminating operations (filter, join, flatMap, etc.) actually terminates
  We use Produced  instead of Repartitioned, because when operating recursively, it would be an error to modify the number of partitions, since the topic MUST have the same number of partitions as the current Task
  • .

Implementation

In KStreamImpl, implementation is fairly simple:

1. We first determine if repartitioning is required, and if it is, we include a repartition node.
2. Unlike other methods, we then wire up the current graphNode as a parent of the current graphNode in the other KStream.

Automatic Repartitioning

Repartitioning is required only when the current KStream requires repartitioning and the other KStream does not require repartitioning. The reason for this is that if both streams require repartitioning, then the other stream is guaranteed to automatically repartition records before any operation that requires repartitioning. In the above example, to would not include a repartition topic, because join already includes one.

Restrictions:

...

1. call op, passing our current KStream as its argument. This produces our output KStream.
2. We wire up the graphNode  from the output KStream  as a parent of the current KStream. This takes care of the recursion.
3. Finally, we return the output  KStream. This enables users to operate on the records that are being recursively produced, as above.

Compatibility, Deprecation, and Migration Plan

...

The following tests will be added:

• Wire up processors "in reverse"
• Streaming recursion:
  • No repartitioning required
  • Repartitioning handled by other stream.
  • Repartitioning handled by to call.

• Counting descendants of graph nodes arriving in-order (as above)
• Counting descendants of graph nodes arriving in any order

Rejected Alternatives

It's currently possible to implement streaming recursion via explicit topics, albeit with a number of disadvantages. See "Motivation" section for details:

1. The explicit topic is entirely internal to the Topology, yet it has to be managed explicitly by the user.
   1. This is functionally a repartition  topic, however, because it's explicitly managed, Streams can't automatically delete consumed records.
   2. Consequently, to prevent the topic growing unbounded, users would need to set retention criteria, which risks possible data loss.
2. In scenarios where repartitioning is not required, the explicit recursive topic adds overhead.
3. It also adds some complexity to the user's program, making it more difficult to reason about than it needs to be.