For really deep, structural (and large datasets), then it's pretty difficult to avoid downtime. That said, those circumstances are hopefully pretty rare. We ge

For really deep, structural (and large datasets), then it's pretty difficult to avoid downtime. That said, those circumstances are hopefully pretty rare. We generally write scripts that bridge between the two formats. One to copy/transform and one to cleanup after. Run the transform, test and verify, then move to cleanup. It's a huge pain, but if you keep on top of if --especially the cleanup-- then it gets easier. It tends to get really hard once you get a lot of cruft. You end up with the code vs data dichotomy. So keeping on top of that really helps longer term.