Here, we need to change the setting max.poll.interval.ms (default 300000) to its double in order to tell Kafka a consumer should be considered dead if the consumer only if it hasn't called the .poll() method in 10 minutes instead of 5.

One consumer per partition assignment will keep 5 consumers idle.

Only one of the replicas is elected as partition leader. And a broker can definitely hold many partitions from the same topic on its disk, try creating a topic with 12 partitions on one broker!

Result is aggregated into a table with key as the unique word and value its frequency. We have to enable log compaction for this topic to align the topic's cleanup policy with KTable semantics.

Controller is a broker that in addition to usual broker functions is responsible for partition leader election. The election of that broker happens thanks to Zookeeper and at any time only one broker can be a controller

Here offset is committed before processing the message. If consumer crashes before processing the message, message will be lost when it comes back up.

Once the client connects to any broker, it is connected to the entire cluster and in case of leadership changes, the clients automatically do a Metadata Request to an available broker to find out who is the new leader for the topic. Hence the producer will automatically keep on producing to the correct Kafka Broker

Schema registry is a separate application that provides RESTful interface for storing and retrieving Avro schemas.

See: 

https://kafka.apache.org/21/javadoc/org/apache/kafka/clients/producer/KafkaProducer.html

Sending a message with the null value is called a tombstone in Kafka and will ensure the log compacted topic does not contain any messages with the key K upon compaction