[
https://issues.apache.org/jira/browse/KAFKA-13197?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17402211#comment-17402211
]
Tommy Becker commented on KAFKA-13197:
--------------------------------------
Hey [~guozhang] thanks for the response. To be clear, I'm looking at
{{KStream.leftJoin(GlobalKTable, KeyValueMapper, ValueJoiner)}} as shown
[here|https://github.com/apache/kafka/blob/trunk/streams/src/main/java/org/apache/kafka/streams/kstream/KStream.java#L2964]
and similar signatures, which still have the (incorrect) verbiage I quoted.
It seems from the churn in this area that there is a need to allow both null
stream key/value records as well as non-null stream side records that map to
null join keys,, though my use-case and this issue are specifically about the
latter, which used to work prior to 2.7.
> KStream-GlobalKTable join semantics don't match documentation
> -------------------------------------------------------------
>
> Key: KAFKA-13197
> URL: https://issues.apache.org/jira/browse/KAFKA-13197
> Project: Kafka
> Issue Type: Bug
> Affects Versions: 2.7.0
> Reporter: Tommy Becker
> Priority: Major
>
> As part of KAFKA-10277, the behavior of KStream-GlobalKTable joins was
> changed. It appears the change was intended to merely relax a requirement but
> it actually broke backwards compatibility. Although it does allow {{null}}
> keys and values in the KStream to be joined, it now excludes {{null}} results
> of the {{KeyValueMapper}}. We have an application which can return {{null}}
> from the {{KeyValueMapper}} for non-null keys in the KStream, and relies on
> these nulls being passed to the {{ValueJoiner}}. Indeed the javadoc still
> explicitly says this is done:
> {quote}If a KStream input record key or value is null the record will not be
> included in the join operation and thus no output record will be added to the
> resulting KStream.
> If keyValueMapper returns null implying no match exists, a null value will
> be provided to ValueJoiner.
> {quote}
> Both these statements are incorrect.
> I think the new behavior is worse than the previous/documented behavior. It
> feels more reasonable to have a non-null stream record map to a null join key
> (our use-case is event-enhancement where the incoming record doesn't have the
> join field), than the reverse.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
