I still haven't seen a methodology that is more thorough and efficient then when I was a noob mainframe programmer 20 years ago at a large insurance company. Basically, the routine is: unit and system/integeration testing uses a sampled subset of your production data while pre-production testing uses all of it. Usually that ideal can be achieved.

Today, though, on the most massive systems I would sample and make a respectable effort to maintain a decent regression test plan. Management has to be convinced of it's value though. Maybe a few more flash trading crashes on Wall Street will get the message through to a new generation...