This is a bit like asking “how do you cook meat for a lot of people?” Not only does number-of-people and kind-of-meat matter a great deal, but even with that info, there’s a million different valid answers and an entire sub-field-ish of science on how to do it.
Based on what little info there is, I’m going to guess that A B testing with groups of experimental features enabled would be best for your case.
Heroin