[robsws]

As much as I understand that checked exceptions (the general term for the 'throws' feature) can lead to a bit of a maintenance nightmare and doesn't necessarily scale well with deeper call stacks, I have found python documentation extremely lacking in this area too.

I think what I've come to understand is that you can treat errors you don't know about as non-recoverable, because that's most likely what they are anyway if they aren't readily documented. Let them bubble up basically, and handle them at the top level if necessary to prevent crashes in prod/make sure they are logged correctly and so on.