Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

Affected methodsCurrent argument typeNew argument type
(KGroupedStream|KGroupedTable).aggregate

Aggregator<K, V, T>

Initializer<T>

Serde<T>

Aggregator<? super K, ? super V, ? extends T>

Initializer<? extends T>

Serde<? extends T>

(KTable|KStream).filter*, KStream.branchPredicate<K, V> Predicate<? super K, ? super V>
(KStream|KTable).groupBy

KeyValueMapper<K, V, X>T>

Serde<X>

KeyValueMapper<? super K, ? super V, ? extends X>

Serde<? extends X>

T>

KStream.(selectKey|map|flatMap), KTable.toStreamKeyValueMapper<K, V, X>KeyValueMapper<? super K, ? super V, ? extends X>
(KStream|KTable).mapValues, KStream.flatMapValuesValueMapper<V, X>ValueMapper<? super V, ? extends X>
KStream.transformTransformerSupplier<K, V, X>TransformerSupplier<? super K, ? super V, ? extends X>

KStream.transformValues

ValueTransformerSupplier<V, X>ValueTransformerSupplier<? super V, ? extends X>
(KStream|Ktable).foreachForeachAction<K, V>ForeachAction<? super K, ? super V>

KStream.process

ProcessorSupplier<K, V>ProcessorSupplier<? super K, ? super V>
(KStream|KTable).*joinValueJoiner<K, V, R>ValueJoiner<? super K, ? super V, ? extends R>

(KStream|KTable).(to|through)

StreamPartitioner<K, V>StreamPartitioner<? super K, ? super V>
KafkaStreams.metadataForKey
StreamPartitioner<K, V>StreamPartitioner<? super K, ? super V>

...

For KGroupedStream/KGroupedTable groupBy and aggregate methods it was decided to leave the return type invariant, since the change is not as straightforward for groupBy and aggregate methods, since those methods sometimes also require passing a Serde for the corresponding . Those methods sometimes require passing a Serde<T> or Initializer<T> where T needs to be consistent with the Aggregator (for aggregate) or KeyValueMapper (for groupBy) result type.

For backwards compatibility reasons, and to avoid runtime class cast exceptions, the choice was made to not enforce type contraints across initializer, aggregate, and serde, make the result type covariant, even though that would have been more correct (see rejected alternatives).

...

  1. make no changes to the output type, i.e. keep the existing output type invariant, leaving an the inconsistent API which would require and do another API change once we can drop support for 1.7
  2. use the more correct <T, VAGG extends T> contraint, and break source compatibility for 1.7 targets, forcing those users to rely on ugly casts or intermediate variables.
  3. make the API consistent by making result types covariant using wildcards ? extends V, relaxing compile time correctness across initializer, aggregator, and serde output types. We initially explored this route because if was backward compatible, but decided to drop it because it would a) require non-trivial changes to the existing streams code, and b) introduce lots of unchecked casts that could blow up at runtime if a user is not careful to ensure consistency across output types in aggregate / groupBy

We decided to chose this last approach (3.) in order to ensure a consistent (and more correct) API while ensuring 1. at the expense of a more consistent API, to ensure backwards compatibility for 1.7 users and avoid the complexity and potential pitfalls of the last approach.

Once we drop support for 1.8 we can always decide to switch to approach 2. without breaking source compatibility, by making a proposal similar to this KIP.