[groestl]

You basically said so in your last paragraph, and I agree it's a tradeoff, I just want to reiterate that in my experience, bubbling up and retrying/degrading/failing at the top level (system boundary, client side) makes for a robust (distributed) system you can reason about. Having hundreds of easter eggs in the project where somebody tried to do something smart when encountering an error seems more like a nightmare scenario to me. Old Java enterprise applications are filled with this, and they are rightfully frowned upon. Checked exceptions are the culprit.