Staged rollout in the traditional sense wouldn't have helped here because the skanky kernel driver worked under all test conditions. It just didn't work when ot got fed bad data. This could have been mitigated by staging the data propagation, or by fully testing the driver with bad data (unlikely to ever have been done by any commercial organization). Perhaps some static analysis tool could have found the potential to crash (or the isomorphic "safe language" that doesn't yet exist for NT kernel drivers).