[josevalim]

Hard to say, as it really depends on what goes wrong. In practice we would likely need to make different trade-offs (for example, do less inference).