I read a blog post on this topic recently.

I like the approach depicted in the article - try to resolve and then rollback after having run there enough troubleshooting to hopefully identify the problem and shortcomings. In addition, I would say there needs to be a go/nogo point in the rollout where the rollback needs to be considered.

http://tjaybelt.blogspot.com/2014/02/be-tenacious.html