Make autoscale work in practice for big deployments
Autoscale doesn't work for big deployments. The design is wrong.
Autoscale detects that it should scale up or down, and sends the message to RDFE, but if RDFE sees any error condition (e.g. broken disk, something that will happen more often on big deployments), it will not scale the instance.
Autoscale developers claim that autoscale works because RDFE gets the signal. They are wrong. Autoscale works when we get enough instances to handle the load. As a customer I don't care about internal signals in Azure. I use autoscale to get the instances I need when I need them and to get rid of them when I don't need them.