(knowing of course that we're more or less tied to Google's infrastructure and the GAE way of doing things)
If you know what you're building and you want to scale it using the sort of tools AE provides -- i.e. the datastore, taskqueue, etc. all fits your app well -- it's quite good. And quite cheap; it's not really fair to compare per-resource pricing to AWS because on AE you'll only pay for what you use instead of paying for idle time on a server instance. But it definitely constrains what you can do, and also locks you to a single hosting provider. If you're still exploring what you want to do, that level of design and vendor lock-in can be a pretty severe liability.
In the real world however, requirements change: * You come up with a new idea that requires a certain library. Chances are, the library won't work on GAE out of the box. * You find out that you need to change your schema. It is pretty hard to update to the new schema while keeping everything in sync.
Finally, you pay the Google cost. When Google implements a new feature, they spend enormous time making sure that it scales well. They need to do so since they could be looking at millions of users on day 1. Most of us however, are looking to build something as cheap as we can, not knowing whether anyone is going to bother to look at it. However, you have to do the same performance optimizations that Google has to do so that your app scales. Chances are, it will be wasted effort - unless your objective is to just learn. I find it funny that GAE goes completely against the rule that "Premature Optimization is the root of all evil". Yes, you should think about your application's scalability. But your bigger problem should be about finding traction, and being able to react fast, not optimize for millions of views.
At a high load (50+ requests/sec), we saw database timeouts tens of times per minute, and often more (every request in a 10-60 sec period would fail).
That's not a terribly high load by any modern standard, you can easily service that on one or two fairly commodity machines these days.
Assuming the point of GAE is scalability, this issue didn't make any sense to me. One would assume that Google would simply load balance requests out to replicated instances of your app and that those apps would in turn access Google's vaunted super scalable data services. Thousands of requests a second shouldn't even phase it.
I've noticed that response times can be highly variable, but so far (we're pre-launch) I haven't noticed anything just simply not getting returned or the kinds of slowness he's describing. Without more detail, something else must be going on here.