Status

Current state: ["Under Discussion"]

Discussion thread: here

JIRA: here 

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

Pre-Kraft brokers currently register 0 for every controller metric. They do this because in the case of a controller failover, the metrics of a newly elected controller will no longer be 0. KRaft does not have this issue because processes with the "broker" role are not eligible to become controllers. There is some performance overhead associated with registering metrics, so moving forward it would be best if KRaft brokers do not expose controller metrics.

Public Interfaces

The behavior of which nodes expose controller metrics will be different in KRaft clusters than the behavior when using Kafka with Zookeeper. KRaft nodes will only expose the controller metrics on the controllers. The controller metrics are calculated as metadata records are replayed e.g. OfflinePartitionCount is calculated as PartitionRecord and PartitionChangeRecord are replayed by the active controller. If the controller is a standby, it will expose 0 for the metric since it has not replayed any metadata records yet. 

Proposed Changes

Pre-Kraft brokers expose 0 for controller metrics unless they are the active controller. KRaft brokers should not since they are not controller eligible.

Compatibility, Deprecation, and Migration Plan

• Certain tools that expect brokers to expose controller metrics may need to be extended for KRaft to no longer expect brokers to expose controller metrics.

Rejected Alternatives

• Expose 0 on KRaft brokers for controller metrics.
  • There is non-negligible performance impact associated with doing this, so it would be best to change the behavior of controller metrics moving forward instead.