Over time, though, my initially rosy feelings towards ORMs have begun to sour. I gradually realised I was spending a disproportionate amount of time trying to coax the ORM into doing my bidding - and when I succeeded, the results were often ugly, slow and needlessly opaque. Analysing the performance of some of the more complicated portions of my data access layer was often painful, and I spent cumulative hours poring over generated SQL, trying to figure out what the ORM was doing and why. Usually, improving performance involved side-stepping the ORM altogether. Recently, a particularly gnarly performance issue prompted me to ditch the ORM from a project altogether, with surprisingly pleasant results.