I'd expect they have a reasonably effective approach for A/B testing. Essentially, you want a uniformly random sample of your users, but you also want to enforce some properties about "connectedness" of the users in the sample, since user-user interaction might be an interesting metric to track. I think the buzzwords related to this sort of thing is "causal inference."

Based on the performance of some sample, you might decide to tweak the feature and release it to a new sample. The buzzwords for this would be "bandit algorithms."

Facebook does gradual releases. One can guess this is available only in the US (or even just a part of the US) for now.