About a week ago I had a problem with after_commit callbacks not being called. It turns out (after diving into the AR source) that they indiscriminately swallow errors from after_commit callbacks. If you're testing a newly-created callback it is difficult to determine whether it's actually been called or merely just raised an error.

I couldn't have figured that out without reading the source- it's not documented. From the developer level that behavior is indistinguishable from magic, and the callback code in AR tries to act like magic as well. It's extremely difficult to follow, let alone understand.