We're on Django; there is a relational DB in our architecture, but when you're browsing the site you're nearly always hitting either Solr or a key-value store.

But why wouldn't you use a relational database? It works, the ORM's good, it's fast enough, you're probably not going to hit scaling problems. If you do hit scaling problems, then that's a really really good problem to have.

Maybe you should look at web2py then. The database abstraction layer is ready to use gae out of the box.

Non relational support is coming soon. Hopefully in 1.3. I am sure there will be "backends" for mongodb, Appengine immediately.

start with web.py

then incrementally add code as you need it, how/where you need it, including how to persist

How have your experiences with web.py been? The disadvantages I saw with it are that it just doesn't have a really large community, plus there are features I just don't want to build myself. Django's built-in admin/ app, for example, or Pylon's error page.

I like starting web apps with web.py, and I like building what may end up being merely demoware or a narrow tech prototype using web.py. I think it excels at whipping up a dynamic Hello World web app in just a few lines of code, with no boilerplate directory tree or scriptage, and no messing around with external servers, and no external code dependencies. Add sqlite when need persistence. Add memcached when need caching. Grow and refactor as you go.

But if I'm very confident it's going to become part of a long-lived dynamic web service with enough of the features Django provides anyway -- especially that awesome built-in admin UI -- then yes I start with Django. The mindshare/community factor is helpful too. I know there are other web frameworks which are decent but going with a big framework other than Django in the Python space today is a bit like going with FreeBSD rather than Linux when all you really need is some sort of free Unix. To pick Linux is to get a bigger ecosystem today and therefore probably also a longer supported future.