For us, any VLDB that is critical has some kind of high-availalbility option like mirroring or active-passive clustering to help stave off outages. However, no solution is perfect.

What we have started doing is warehousing data that is necessary to keep for processing, but not needed in the "day-to-day" shuffle of data.

We can conglomerate data between the OLTP and warehouse databases for reporting, when needed and we have processes that pull data out of the warehouse for day-to-day operations when needed.

One aside, how does one recover a database when it is damaged for use? How can you ensure that all of the information your application requires will be intact and has integrity?