Somewhat. Depends how good the predictors of failures are. If they're strong enough, it doesn't matter how many promising signs there are.

But you're right that this is a perilous move. We only did it because we had to, and I'd like to edge back toward our old approach now that we can scale better. We'll probably always have a tighter filter than we used to, though, if only because some of the techniques we learned for predicting failure seem very useful.