- Kafka Stream join are not deterministic in sub-topology context (ie. repartitioning before a join)
.through()create a new stream within the current topology thus can't solve the above point- Whenever there is a repartition operation (
selectKey,transform, etc.) you shouldn't chain ajoinoperation. (ie. create dedicated topologies) - Timestamp synchronizationis dependant of the order Kafka subscriptions polling order thus the timestamp comparison should be strict (ie.
<strict not<=).
A suite of KStream-KTable join edge cases spotted during the Stock Booster (SEB).
Have a look to the test KStreamKTableJoinWithTimestampExtractorTest.
The project runs with Java 11 and build with Maven project. Uses HAP config files to configure Michelin Artifactory repositories if you can't access to the Internet.
./mvnw clean test