In web site management, experiments deployed to a fraction of users are a great thing for testing new features, or for comparing different UI approaches. They let you find out things like “users use the help button more often if it’s next to the search box instead of at the top right” or “people click on ads more if they have a cute pink background”.

Literally turning off the service for a fraction of users is, um, not the sort of experiment worth performing. You already know what happens when you have an outage, because you’ve had outages before. The only difference is that by doing it deliberately, you show that it’s not the tech that’s unreliable; it’s the people running it.