Recovering from memory errors at run time is unreliable.
I should add that I have more like a romantic view of software release cycles where testing is done until the software is in a very, very sophisticated and stable state. More than usual.
Not that I want to solely rely on such an approach. It should still try to recover from failures, and recovering should be trained like fire fighting is trained. But there should be no smokers in the building, so to speak.
