You have to define health checks on your instances that reflect the availability of all services they host.
And you have to allow there to be more instances than your target number in each autoscaling group.