Overview

Our example Redpanda monitoring dashboards include a derived metric display of Partition Balance, which displays a percentage value. This value can be a source of confusion, but used correctly is a good indication of the balance of the cluster and an early-warning sign for any issues in this area.

 

Detail

The value is always expected to be in the 0-100% range, where 100% implies that each node in the cluster is hosting an identical amount of data (or more accurately technically an identical number of messages). In a running system, a value of exactly 100% is rarely possible as this depends not only on the distribution of partitions within the cluster, but also on the distribution of messages to partitions within that topic.
 
Values above ~80% are not necessarily cause for concern if you're not seeing any other issues with your cluster, but it may be worth reviewing to see if you have any partitions of particular topics that hold more data than others (i.e. there is an application level skew). It is also worth considering if you have any topics with a high message cardinality but low number of partitions (particularly where this number is not divisible by the number of Redpanda nodes), or topics with a replication factor of 1, as these are both factors that can contribute to an uneven distribution of data.