I'm confident that SQLAlchemy has the potential to be the best database abstraction layer out there. However, I found the docs impervious and frustrating. Some of the examples are given in a way that's quite unhelpful - they could be constructed to distinguish the way it works from other assumptions you could make and generally they're not. When things break, the error messages are unhelpful.

I eventually got to the stage where I had a product built on top of it, working, but I still didn't feel like I knew the framework fluently. I've never been in that position before. Frustating, because it means if I do more work on that codebase I'll either have to struggle through again or else throw it away and write on top of something else.

ORM is supposed to be good for nailing the right schema and then abstracting it away without a performance hit. That's what I get with Cayenne. I found the opposite with sqlalchemy - it's more difficult than it would have been for me to have just written the queries inline, or rolled my own simple ORM. I think it's because it's just too thick - unusual for a python framework - usually they're the opposite.

It's been a year since I looked at it so I'm a bit vague above, but I gave it a go for over two years on and off.