If you take the feature switching rollout to its logical conclusion, then you can switch on a new refactored version for a small percentage of users but not actually expose the output. You will run the existing version in parallel as a control and show this to the user. However, you will collect the results of both the versions, including performance data. You can then compare them to ensure that the new version (hopefully with a higher performance) is correct and consistent with the existing behavior.
GitHub has an open source Ruby library, called Scientist, which can be used to successfully refactor permissions and for merging code. In the process, GitHub discovered existing legacy bugs, found missing features, ...
