We treat a stack as mostly a configuration boundary, such that everything in the stack gets configured the same. We wanted to be able to update/reconfigure clusters independently, so that contributed to our stack per cluster approach.