We consider strategies for full backups from the viewpoint of competitive analysis of online problems. We concentrate upon the re- alistic case that faults are rare, i.e. the cost of work between two faults is typically large compared to the cost of one backup. Instead of the (worst- case) competitive ratio we use a refined and more expressive quality measure, in terms of the average fault frequency. This is not standard in the online algorithm literature. The interesting matter is, roughly speak- ing, to adapt the backup frequency to the fault frequency, while future faults are unpredictable. We give an asymptotically optimal determin- istic strategy and propose a randomized strategy whose expected cost beats the deterministic bound.