> It serves about 35,000 concurrently running A/B experiments,

With that amount of optimizations, you would expect them to have the best user experience in the world. Do they? I think there is something severely wrong with the current A/B testing methodology.