KafkaStreamsWordCountApplication failing with java.lang.ClassCastException: kafka.streams.word.count.KafkaStreamsWordCountApplication$WordCount cannot be cast to java.lang.String

After the latest upgrade of the libraries to Spring Boot 2.2.0-SNAPSHOT and <spring-cloud-stream.version>Horsham.BUILD-SNAPSHOT</spring-cloud-stream.version> the KafkaStreamsWordCountApplication fails, as can be seen in the tests:

Caused by: java.lang.ClassCastException: kafka.streams.word.count.KafkaStreamsWordCountApplication$WordCount cannot be cast to java.lang.String
        at org.apache.kafka.common.serialization.StringSerializer.serialize(StringSerializer.java:28)
        at org.apache.kafka.common.serialization.Serializer.serialize(Serializer.java:60)
        at org.apache.kafka.streams.processor.internals.RecordCollectorImpl.send(RecordCollectorImpl.java:162)
        at org.apache.kafka.streams.processor.internals.RecordCollectorImpl.send(RecordCollectorImpl.java:102)
        at org.apache.kafka.streams.processor.internals.SinkNode.process(SinkNode.java:89)
        ... 32 more

The error happens after this map operation, before writing to the sink:

.map((key, value) -> new KeyValue<>(null, new WordCount(key.key(), value, new Date(key.window().start()), new Date(key.window().end()))));

I’ve tried with a previous version that worked and the topology included a mapvalues node that converted the WordCount object value to a byte array,:

The current topology doesn’t show that mapvalues processor anymore:

[sobychacko]

@codependent New changes went into core binder for native Serde changes and inferring Serdes on input/output as much as possible. I cleaned up all the Kafka Streams samples to reflect the changes. I ran your Kotlin sample locally and it runs as expected.

Here is the gist of the Serde changes that went into the master. These changes will only be available on the 3.0 (Horsham) line of Spring Cloud Stream and won’t be backported.

1. Binder now uses native Serde as the default serialization and deserialization strategy for input and output bindings.
2. Applications have to explicitly set useNativeDecoding and useNativeEncoding to false to force the conversion done by the framework.
3. When using the default native conversion by Kafka Streams, applications don’t need to set Serdes used on key/value types if those types map to one of the following:

• Integer
• Float
• Double
• Long
• Short
• byte[]
• String

If none of these matches, it defaults to JsonSerde provided by spring-kafka. If the application explicitly provides a Serde on the binding, then that gets precedence always. If the application uses Serde objects like Confluent’s Avro Serde, then that has to be set explicitly on the configuration.

These changes will also significantly reduce the topology depth of Spring Cloud Stream/Kafka Streams applications that you once reported on SO.

Once you confirm that things are working as expected, we can close this issue.

[sobychacko]

That sounds like a good plan. Feel free to close this and open another one.