My take on this is that it's not always the best idea to abstract-out SQL. You see, the SQL itself is too valuable abstraction, and also a very "wide" one. Any attempt to hide it behind another abstraction layer will face these problems:
- need to learn secondary API which still doesn't cover the whole scope of SQL
- abstraction which is guaranteed to leak, because any time you'll need to optimize - you'll need to start reason in terms of SQL and try to force the ORM produce SQL you need.
- performance
- deceptive simplicity, when it's super-easy to start on simple examples, but it's getting increasingly hard as you go. But at the point you realize it doesn't work (well) - you already produced tons of code which business will disallow you to simply rewrite
- need to learn secondary API which still doesn't cover the whole scope of SQL
- abstraction which is guaranteed to leak, because any time you'll need to optimize - you'll need to start reason in terms of SQL and try to force the ORM produce SQL you need.
- performance
- deceptive simplicity, when it's super-easy to start on simple examples, but it's getting increasingly hard as you go. But at the point you realize it doesn't work (well) - you already produced tons of code which business will disallow you to simply rewrite
(knowledge based on my own hard experiences)