I recently ran partition assignment on some topics. This made the replicas of some partitions move around the cluster. It was seamless. However, when it came to purging old logs following the retention.ms property of the topic, the replica partitions were not clear. The leader partition, however, was purged correctly. As our consumers have caught up, a quick fix was to reduce retention period to 1 day, let Kafka purge those replicas and reset the retention policy.
This looks like a bug. We had almost run out of disk space on few of our brokers due to this. -- Anish Samir Mashankar