Spikes of 2000/second are not unrealistic, and part of the problem is that once the spike hits, inadequately designed applications start to wedge on lock conten

Spikes of 2000/second are not unrealistic, and part of the problem is that once the spike hits, inadequately designed applications start to wedge on lock contentions and just plain inadequate CPU, the VM layer starts swapping, people trying to get to your site start hitting reload (humans do not have an exponential back-off function), and the whole thing comes crashing down until traffic drops off to more sustainable levels.