According to a flow test carried out on the latest code base, scale down functionality does not work as intended.
1. Partition Max Instance Count set to 3
2. Send a sample request load to take scale up action. Autoscaler will spin up 3 instances.
3. Stop the sample request load.
4. Let the service cluster scale down.
5. One instance being terminated but two remains.
As found in the Autoscaler logs Member Stat Context of the terminated member is not being removed from the Cluster Context. As a result the same member being selected in each scale down action round.