The vast majority of our traffic is from devices that retry later if anything goes wrong. It is analytics information and hence there is no human affected by it.
Despite having identical settings and code across a number of appengine apps, we see variance in the number of instances they decide to run (eg they will start 4 at once even though only 1 is needed for the load).
At one point they were starting hundreds of instances due to some of our aggressive inter-app traffic (thundering herd problem). Things would have worked fine if we could limit the number of instances, but tough luck there. It required considerable engineering and tuning to make the problem go away that would have trivially been solved (in the short/medium term) by limiting instance numbers.
We aren't live yet, so I haven't needed to think about that stuff yet. However, I've talked to people that have applications live on App Engine, and that hasn't come up in conversations ... so there must be some solution ... if not the exact one you want.