For KStream.join / aggregateByKey operations that requires the streams to be partitioned on the record key, today users should repartition themselves through the "through" call:
This pattern can actually be done by the Streams DSL itself instead of requiring users to specify themselves, i.e. users can just set the right key like (see
KAFKA-3430) and then call join, which will be translated by adding the "internal topic for repartition".
Another thing is that today if user do not call "through" after setting a new key, the aggregation result would not be correct as the aggregation is based on key B while the source partitions is partitioned by key A and hence each task will only get a partial aggregation for all keys. But this is not validated in the DSL today. We should do both the auto-translation and validation.