Current state: Under Discussion
Discussion thread: TBD
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
The main motivation of this KIP is stated in the related JIRA: "Today the downstream sub-topology's parallelism (aka the number of tasks) are purely dependent on the upstream sub-topology's parallelism, which ultimately depends on the source topic's num.partitions. However this does not work perfectly with dynamic scaling scenarios". By delegating the stream topology power to create repartition topic with customized number of partitions gives user more flexibility. Also, for API like #to or #through, KStream has no access to sink topic creation, which means user has to create their own connecting Kafka topic for every single new application which is cumbersome. Thus we are proposing extending the capability of #Produced and #Grouped API to automatically create connecting topics through KStream application when the target topic has not created.
We shall expand the Produced API to contain numPartitions:
Also expand Grouped API with a numPartitions configuration：
When configured number of partitions, KStream application will first issue the topic lookup request and check whether the target topic is already up and running. If the target topic already exists, KStream application will log the error for topic creation failure, and still continue the initialization. For grouping operation, repartition topic gets a hint on how many partitions it should be created with. If repartition topic is not created yet, create one with specified numPartitions; otherwise use the upstream topic partition size as the new topic number of partitions. One open question is whether we should prevent destructive operation such as changing the repartition topic size if current topic numPartitions != configured numPartitions, since there could be potential data loss.
This is a pure KStream library change that shouldn't affect previously setup applications. If numPartitions is not configured by user, it will use a default value -1 which won't affect existing topology initialization logic.