There's some olllld writeup on this here: http://wiki.mediagoblin.org/DesignDecisions
The short of it is: I knew I liked Jinja2 better than django's templates, I knew I wanted to swap out a number of things, and originally we used Mongo (we're now using SQL... longish story for another time), and we were swapping out enough components that eventually you hit that point where you're replacing all the bits of your framework and you spend a bunch of time fighting the framework because you aren't using it quite in all the ways it's opinionated about.
Anyway, I think not using django and using a very django-like system pieced together from various libraries has been a mixed bag. I actually think it is paying off for us: the way that we're doing plugins likely would be a bit harder if we were using Django (long story again), and it's super, super fast and easy to adjust the core "unframework" of MediaGoblin if you know what you're doing (I recommend reading http://pythonpaste.org/webob/do-it-yourself.html ). Anyway, our system looks increasingly like the application structure of Django, but using components that pretty much look like Flask. I wouldn't recommend all applications do this... (and also, I think Django is really great... there's a reason MediaGoblin looks so similar) but I think it's worked out well for us.
I'd like to do a better writeup now that some time has passed. I might give a talk on the pros/cons this at next year's PyCon if my proposal is accepted.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds