Kafka brokers in a ZooKeeper-based cluster maintain an active session with the ZooKeeper ensemble by sending regular heartbeats. A long "stop-the-world" JVM garbage collection (GC) pause can prevent a broker from sending these heartbeats. If the duration of the GC pause exceeds the configured ZooKeeper session timeout (zookeeper.session.timeout.ms), ZooKeeper will assume the broker has failed. It will close the session and remove the broker's ephemeral znode, signaling to the rest of the cluster (specifically the controller) that the broker is dead. This initiates leader re-elections for partitions hosted on that broker.

A. The log cleaner thread is triggered based on log size and time configurations for compacted topics, not by a GC pause. A long pause would halt this thread, not trigger it.

C. While clients will experience backpressure because the broker is unresponsive, this is a secondary effect. The primary, cluster-level event is the failure detection by ZooKeeper.

D. Loading log files into the page cache is a standard operating system memory management function related to file I/O, which would be suspended during a GC pause.

1. Confluent Documentation, Broker Configurations: The documentation for zookeeper.session.timeout.ms explains its function: "Zookeeper session timeout. If the broker fails to heartbeat to Zookeeper for this period of time it is considered dead." A long GC pause directly causes this failure to heartbeat.

Source: Confluent Documentation, "Broker Configurations", section on zookeeper.session.timeout.ms.

2. Shapira, G., Narkhede, N., & Palino, T. (2021). Kafka: The Definitive Guide (2nd ed.). O'Reilly Media, Inc.

Reference: In Chapter 8, "Reliability," the section "Cluster Membership" details how brokers use ZooKeeper sessions to maintain their status in the cluster. It explicitly mentions that "long garbage-collection pauses" are a common reason for a broker's session to time out, causing ZooKeeper to declare it dead.

3. Confluent Developer, "Apache Kafka and Garbage Collection":

Reference: This official resource explains the impact of GC on Kafka brokers: "A long GC pause can cause a broker to miss its heartbeat to ZooKeeper, which will cause ZooKeeper to declare the broker as dead. This will in turn cause all the leaders on the dead broker to be moved to other replicas." This directly confirms that a long GC pause leads to the broker being considered dead by ZooKeeper.