Upstart's event-based model was semi-broken and AFAIR the original author admitted as much. (I believe he posted a message stating this to a maining list, but I cannot find it ATM.)

Exposing the event model directly to the jobs also made reasoning about the current system state (which is what actually matters, not how we got there) tended to make upstart jobs a lot more complicated than necessary. (This is from memory, it's been ages since I actually looked at upstart jobs.)