How does leader election work in Apache Kafka®? For the past 2 ½ years, Adithya Chandra, Staff Software Engineer at Confluent, has been working on Kafka scalability and performance, specifically partition leader election. In this episode, he gives Kris Jenkins a deep dive into the power of leader election in Kafka replication, why we need it, how it works, what can go wrong, and how it's being improved.
Adithya explains that you can configure a certain number of replicas to be distributed across Kafka brokers and then set one of them as the elected leader - the others become followers. This leader-based model proves efficient because clients only have to write to the leader, who handles the replication process internally.
But what happens when a broker goes offline, when a replica reassignment occurs, or when a broker shuts down? Adithya explains that when these triggers occur, one of the followers becomes the elected leader, and all the other replicas take their cue from the new leader. This failover reassignment ensures that messages are replicated effectively and efficiently with multiple copies across different brokers.
Adithya explains how you can select a broker as the preferred election leader. The preferred leader then becomes the new leader in failure events. This reduces latency and ensures messages consistently write to the same broker for easier tracking and debugging.
Leader failover cannot cover all failures, Adithya says. If a broker can’t be reached externally but can talk to other brokers in the cluster, leader failover won’t be triggered. If a broker experiences transient disk or network issues, the leader election process might fail, and the broker will not be elected as a leader. In both cases, manual intervention is required.
Leadership priority is an important feature of Confluent Cloud that allows you to prioritize certain brokers over others and specify which broker is most likely to become the leader in case of a failover. This way, we can prioritize certain brokers to ensure that the most reliable broker handles more important and sensitive replication tasks. Additionally, this feature ensures that replication remains consistent and available even in an unexpected failure event.
Improvements to this component of Kafka will enable it to be applied to a wide variety of scenarios. On-call engineers can use it to mitigate single-broker performance issues while debugging. Network and storage health solutions can use it to prioritize brokers. Adithya explains that preferred leader election and leadership failover ensure data is available and consistent during failure scenarios so that Kafka replication can run smoothly and efficiently.