[
https://issues.apache.org/jira/browse/KAFKA-10847?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17273206#comment-17273206
]
Matthias J. Sax commented on KAFKA-10847:
-----------------------------------------
Thanks [~spena] – overall using a second store might be the simples solution
and if we can get some perf results we can make a better decision if the
performance is acceptable or not.
The only thing I tend to object is the usage of _wall-clock_ time punctuation,
because it would introduce non-determinism. And if we use stream-time
punctuations, we could even avoid punctuations at all, an "piggy-back" emitting
left/outer join result, each time we process an input record.
Also take into account the grace period, ie, we should only emit left/outer
join result after a window closes (not when a window end): window close =
window end + grace period.
> Avoid spurious left/outer join results in stream-stream join
> -------------------------------------------------------------
>
> Key: KAFKA-10847
> URL: https://issues.apache.org/jira/browse/KAFKA-10847
> Project: Kafka
> Issue Type: Improvement
> Components: streams
> Reporter: Matthias J. Sax
> Assignee: Sergio Peña
> Priority: Major
>
> KafkaStreams follows an eager execution model, ie, it never buffers input
> records but processes them right away. For left/outer stream-stream join,
> this implies that left/outer join result might be emitted before the window
> end (or window close) time is reached. Thus, a record what will be an
> inner-join result, might produce a eager (and spurious) left/outer join
> result.
> We should change the implementation of the join, to not emit eager left/outer
> join result, but instead delay the emission of such result after the window
> grace period passed.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
