Add a sensor to identify a cluster containing partitions with replication factor > number of racks

To enhance the fault tolerance of clusters with insufficient number of racks to provide a full rack-awareness (i.e. each replica of each partition resides on a different rack), we added a new hard goal, called RackAwareDistributionGoal (see https://github.com/linkedin/cruise-control/pull/1345). This goal is a relaxed version of RackAwareGoal, and evenly distributes replicas over racks. Contrary to RackAwareGoal, if replicas of each partition can be evenly distributed across the racks, this goal lets placement of multiple replicas of a partition into a single rack.

To help identification of clusters containing partitions with RF > number of racks, add a sensor that monitoring systems can alert on if needed.

[mgrubent]

For our own purposes, I think the boolean datatype (or its 0/1 integer representation) is sufficient.

I would prefer not to have to de-noise this sensor by comparison to whether there is an ongoing partition reassignment.

The ideal monitoring-ready sensor for me would be:

• 0: No static partitions have RF > the number of racks in the kafka cluster
• 1: At least one static partition has RF > the number of racks in the kafka cluster

Where a “static” partition is one for which there is not an ongoing partition reassignment.
Ideally, this would include partition reassignments from non-cruise-control sources as well, but that may be infeasible to provide in kafka 2.4+

[efeg]

@mgrubent Other than the special value (see https://github.com/linkedin/cruise-control/issues/1365#issuecomment-717438456), it would probably be similar to the sensor that signals whether the cluster has unfixable goals – i.e. 0: no unfixable goal, 1: has unfixable goal.

I would be happy to discuss suggestions that intend to provide a universal solution applicable to all downstream monitoring / alerting frameworks.