A setup with an ASG per AZ makes it a lot harder to do real autoscaling based on load/mem/connections.
If it was purely for running a fixed amount of instances equally spread acros AZ's this would probably work, but not in our setup where we have unpredictable traffic and load patterns.
[edit] I know it can be done with combined metrics etc, but it would make it a lot more complicated ;-)
It is most certainly more complicated. We were ok to be put up with that additional complication because service reliability (especially tolerating single AZ outages with ease) was higher on the requirements than avoiding complication. :)
[edit] I know it can be done with combined metrics etc, but it would make it a lot more complicated ;-)