metadata is stored in and built from the KSQL command topic. Each KSQL server has its own in-memory version of the metastore.

Kafka topics are divided into partitions and spread across brokers. Each brokers knows about all the metadata and each broker is a bootstrap broker, but only one of them is elected controller

All brokers can respond to Metadata request, so a client can connect to any broker in the cluster.

Although any Kafka Streams application is stateless as the state is stored in Kafka, it can take a while and lots of resources to recover the state from Kafka. In order to speed up recovery, it is advised to store the Kafka Streams state on a persistent volume, so that only the missing part of the state needs to be recovered.

In case the consumer has the wrong leader of a partition, it will issue a metadata request. The Metadata request can be handled by any node, so clients know afterwards which broker are the designated leader for the topic partitions. Produce and consume requests can only be sent to the node hosting partition leader.