When things go wrong (and they still occasionally do - on every platform), I'm one of the more vocal critics on the GAE mailing list. However, the zero-ops promise has held up pretty well. There's always something to complain about, but isn't it.
Also, that last statement is a little puzzling. GAE already handles tens of millions of users (probably considerably more) in shared infrastructure. The typical bottleneck (the database) isn't an issue because the database is already federated across every app in the system; you don't have to scale dedicated resources.
The real issue is that people don understand that scaling challenges really aren't that bad until you have tons of users; in the few millions of clients range you can very easily handle the traffic with one or two centralized RDBMS systems (and in fact, this is how many sites actually work in practice, even large ones you use daily).
* You must be careful of entity group boundaries. This ends up being fairly easy to do, especially with cross-group transactions (added last year). It's definitely a design consideration, but one you internalize very quickly and as long as you don't screw up your database design, isn't a scaling bottleneck.
* Tablet splitting. If you write monotonically increasing values to one index at high rates (1k/s is what I've heard, but I haven't hit it myself) you may have to work around this issue by federating the index. It's not pretty, but it's not a common problem.
I agree that traditional RDBMS architectures can handle quite a lot of traffic before they hit a wall. It's just a lot more ops work. What I like about GAE-land is that you just have to be a little more careful in your data modeling and you can stop worrying about all the usual problems like monitoring the performance of your database server. I've built rather large sites on RDBMSes in the past, and getting the ops right was always painful.
Thanks for mentioning this: I didn't know they had added this feature; that definitely makes some of the data modeling problems easier. Without this, I found it often led to situations where you had to make horribly painful decisions that affected both functionality and scalability very early on in your project lest you end up with tons of data that you now have to programmatically and mostly manually migrate (which is why, to me, I keep being surprised that you keep mentioning ops challenges for normal databases: with the ability to do global schema and index changes concurrently and safely inside of transactions, I found the amount of time and thought I need to put into managing something like PostgreSQL a fraction of most alternatives that supposedly scale better).
I really do not get the impression that the people on the App Engine product mailing lists actually feel like they are able to spend less time on operations, and there are numerous alternatives that let you use more standardized stacks to get similar benefits, especially if you are serving less than a few tens of millions of users (which, honestly, it is not clear App Engine would handle in practice anyway, even though in theory it would work great).