The issue being that availability zones don't get throttled when they start to overload the API. The control system should have automatically throttled the misbehaving AV.

Which is probably far more difficult to do properly than I can imagine.