Hi everyone, I would like to start a discussion on FLINK-25132: KafkaSource cannot work with object-reusing DeserializationSchema.
The implementation of FLINK-25132 moves the deserialization of kafka records from SplitFetcher thread to thread that runs KafkaRecordEmitter, namely flink's main task thread. In our experience, even without triggering this object-reusing bug in KafkaDeserializerSchema , it still caused a performance regression of 20% to 30%. I tend to add a parameter to flink conf to indicate where the deserialization process needs to be placed, that is, in the fetcher thread or the main thread, since not all deserializer will encounter the mentioned object-reusing bug. Looking forward to your reply. Best, Jun He
