Dan O'Reilly

At Salesforce, we orchestrate one of the industry's largest data streaming footprints: over 500 Kafka clusters processing 3 trillion events per day and consuming 60 PB of data monthly. At this extreme scale, our Kafka Connect clusters often operate with consumer groups exceeding 1,000 members. With group sizes this large, rebalances of consumer groups using standard dynamic membership pose a serious availability risk. When hundreds of tasks stop, rejoin, and sync simultaneously, this inevitable disruption can cause massive latency spikes and significant throughput degradation, making routine maintenance a high-risk operation. Even with Incremental Cooperative Rebalancing, a worker restart can trigger an extended rebalance for the underlying Task-level consumer group. To solve this, we extended Static Membership to the Kafka Connect task level. By assigning persistent identities (group.instance.id) to ephemeral tasks, we prevent unnecessary rebalances during transient failures and rolling restarts. However, bringing KIP-345 principles to the Connect framework wasn't seamless. This session provides a candid technical deep dive into that journey. We will detail the implementation…