...

This page is meant as a template for writing a KIP. To create a KIP choose Tools->Copy on this page and modify with your content and replace the heading with the next KIP number and a description of your issue. Replace anything in italics with your own description.

Status

Current state: In DiscussionDRAFT

Discussion thread: here [Change the link from the KIP proposal email archive to your own email thread]

JIRA: here

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

Motivation

Kafka Stream supports different ways to window a stream for aggregation, but all of them are extremely sensitive to out of order timestamps and late arriving data. To handle this, we introduced the concept of a "grace period" by which a window can be retained for longer than the window end to allow for updating the result of a window with a late arriving record. This concept is powerful, but doesn't help users that want to aggregate based on offset order alone, not discarding any late arriving record.

A common use case for such a windowed aggregation is to collect records in a very small window (usually a few seconds) into a small batch before being passed to a consumer that can handle batched messages more effectively than individual messages. An example here is collecting records for a few seconds and suppressing them before sending them all together as a single RPC to another service for processing. A naive implementation of that using existing mechanics would look likeOften users want to reduce the cardinality of a stream (aggregate) independent of time semantics so that downstream topologies are not overwhelmed by the throughput of upstream producers. A classic example is to aggregate together N seconds of events for a single key into a single one so that a downstream consumer that does something heavy per-event is not overloaded (e.g. it makes a remote call):

stream
	.windowBy(TimeWindows.ofSizeAndNoGrace(Duration.ofSeconds(10))
	.aggregate((k, v, kvList) -> kvList.add((k, v)))
	.suppress(Suppressed.untilWindowCloses(...))

...

Typically the behavior that users want with this type of operation is to group keys together until a certain amount of time passes and emit them as a single record, always accepting late arriving records and just putting them into whatever the “current” window is. Today the only way to implement this semantic is using a punctuation and manually storing events in an aggregation (which is inefficient for many reasons and can cause severe bottlenecks).

Public Interfaces

Briefly list any new interfaces that will be introduced as part of this proposal or any existing interfaces that will be removed or changed. The purpose of this section is to concisely call out the public contract that will come along with this feature.

We will be introducing a new type of Windows that always computes the current window based on stream time instead of event time.

...

And all events that come within 10 seconds of stream time (independent of their event time) will be aggregated together, suppressed and emitted when the stream time elapses by 10 seconds.

Proposed Changes

Describe the new thing you want to do in appropriate detail. This may be fairly extensive and have large subsections of its own. Or it may be a few sentences. Use judgement based on the scope of the change.

First, we need to expand the StreamTimeWindows class to take in the current stream time when computing the window for a given event:

public abstract class Windows<W extends Window> {
  ...
  public abstract Map<Long, W> windowsFor(
    final long timestamp, 
    final long observedStreamTime
  );
}

The implementation of StreamTimeWindows will then use the observedStreamTime when computing which window the current event should fall into:

@Override
public Map<Long, TimeWindow> windowsFor(
  final long timestamp, 
  final long observedStreamTime
) {
	long windowStart = (observedStreamTime / sizeMs) * sizeMs;
  return Map.of(windowStart, new TimeWindow(windowStart, windowStart + sizeMs));
}

The gracePeriodMs for StreamTimeWindows is always zero, which means there will only ever be one open window per key and events will always fall within this window.

Compatibility, Deprecation, and Migration Plan

N/A

Test Plan

Describe in few sentences how the KIP will be tested. We are mostly interested in system tests (since unit-tests are specific to implementation details). How will we know that the implementation works as expected? How will we know nothing broke?

Nothing special needs to be tested beyond unit and integration tests to make sure we cover situations like out of order and late arriving events.

Rejected Alternatives

If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.

...