Confluent vs Sanity

Confluent Platform 8.2 ships on Apache Kafka 4.2, and its headline is KIP-932 Queues for Kafka reaching general availability: share groups and share consumers let multiple consumers process the same topic-partition concurrently without manual offset management. Supporting KIPs bound the behaviour — strict max fetch records for predictable consumer memory, acquisition-lock renewal so long-running processing does not trigger premature redelivery, and share-partition lag metrics. Kafka Streams gains a native dead-letter queue and anchored punctuation, and Schema Registry can carry schema IDs in message headers instead of the payload.

The release is aimed at the reasons a Kafka shop keeps other infrastructure alongside it. Queueing covers the competing-consumer pattern that sent teams to RabbitMQ or SQS, the Streams dead-letter queue removes a common reason to hand-roll error handling, and schema IDs in headers tidies the payload contract. Packaging is moving the same way: Control Center now ships independently of the platform, from its own repository and release train.

With the queueing mechanism generally available, the next work is likely operational — share-group tooling and metrics in the console, since the KIPs so far have delivered the mechanism and only the beginnings of its observability.