A Connect cluster is a group of worker JVMs sharing one group.id. A connector is a configured plugin instance that splits the job among its tasks, which copy: a source task poll()s an external system, a sink task gets batches in put(). A converter turns records into bytes and back; Single Message Transforms (SMTs) adjust each record.
Workers share no disks: compacted topics hold connector configurations, source offsets (such as the last updated_at read) and task status, and sinks commit consumer offsets under the group connect-<name>. When a worker joins or leaves, the leader reassigns work with incremental cooperative rebalancing (Kafka 2.3+), stopping only tasks that move.