We observed that when we have high number of partitions, instances or stream-threads, assignment-data size grows too fast and we start getting below RecordTooLargeException at kafka-broker.
Workaround of this issue is commented at: https://issues.apache.org/jira/browse/KAFKA-6976
Still it limits the scalability of kafka streams as moving around 100MBs of assignment data for each rebalancing affects performance & reliability (timeout exceptions starts appearing) as well. Also this limits kafka streams scale even with high max.message.bytes setting as data size increases pretty quickly with number of partitions, instances or stream-threads.
To address this issue in our cluster, we are sending the compressed assignment-data. We saw assignment-data size reduced by 8X-10X. This improved the kafka streams scalability drastically for us and we could now run it with more than 8,000 partitions.