Ask HN: What error-handling patterns have you found useful for large code-bases?

It seems like a black art how to properly (1) identify, (2) classify, (3) handle (at the appropriate level in the call stack), and maybe (4) log the different kinds of errors that can occur when writing code. Then there is the question of whether to handle a certain error at all. These are not purely technical challenges. What ways of thinking about errors have you found useful? Do you think the exception-throw-catch model is always better, or have you found other approaches more useful?