We try to recover from some of the failures, sometimes at the expense of potential glitches but the risk is to end up with an unusable browser if things are too glitchy. On the compositor side it is very easy to count the number of api calls that fail and impose a certain threshold above which we should force the browser to crash so that repeated failures show up in crash stats and to give the user a chance to restart in safe mode if the browser is not usable. I think that the count should be periodically reset in order to avoid a benign error happening rarely to cause the crash if the browser was up for a while.