The current service clean up logic in the nova-operator deletes a nova-conductor or nova-scheduler service via the Nova REST API if the service is down. We can improve on that as we can predict the name of the service pod that was deleted by the StatefulSet controller when reconciled a Replica count decrease.

The new logic is: When the StatfulSet is Ready the scheduler (and conductor) controller can look at the current replica count and delete all the services which has a bigger index in its name than the current replica count.

This will have the same performance as both the current and the new logic requires a service list API call and one delete call for each service. We cannot avoid the service list command as we only know the name of the service but delete works via uuid.

However the new logic would remove two potential issues:

• deleting a service because it is down for other reasons than scale down(like pod crash loop)
• deleting the scaled down service happens independently, in a different reconcile loop, than the actual scale-down event.