Apache Kafka Hot Partition

Question

Which metrics in Kafka should reveal if there is any hot partition?

I have checked this https://docs.confluent.io/platform/current/kafka/monitoring.html but it does not mention anything about hot partition, may be this is not a very common jargon to be used to specify this scenario.

By hot partition what I meant is, due to certain design issue / data, if a wrong partition-key-selection-mechanism is used that leads to more messages falling into one partition vs rest of the partitions.

Please share if there is a definitive way to know if there is hot partition, by looking at Kafka metrics so that I can also add alerts to it as well.

Answer 1

I do not believe the broker has knowledge of partition size. Considering each replica could be different sizes, then the cluster as a whole would not show this in a centralized location.

In a previous job, we wrote an agent that when installed on the brokers, reported disk usage for all directories under log.dirs . Then our database was able to order the sizes and compute standard deviations for all partitions. It also allowed us to alert if there was one replica that didn't match the other brokers.

---

Your other option would be to track byte/record send rate, but I think this is per topic, not down to the partition.