Handling errors in compilers has become the most interesting part of my job. If you do the math, most calls to compilers are error-yielding calls. So those calls really need to produce the best errors possible! (Maybe this is less true in LLM land, but certainly it was right before.) Tracking spans, doing extra algorithmic work, and exploring why the expectation failed in order to give a user the most-informative error possible is extra work on the main compute path. But it pays dividends: easily-diagnosable bugs and quick fixes present in the code itself. Tracking spans to point back to origin definitions, etc., always feel like plumbing work until you see the error that explains exactly what went sideways. And once you do, the errors that don't feel like a frustration you are trying to cast your net over.
Handling errors in compilers has become the most interesting part of my job. If you do the math, most calls to compilers are error-yielding calls. So those calls really need to produce the best errors possible! (Maybe this is less true in LLM land, but certainly it was right before.) Tracking spans, doing extra algorithmic work, and exploring why the expectation failed in order to give a user the most-informative error possible is extra work on the main compute path. But it pays dividends: easily-diagnosable bugs and quick fixes present in the code itself. Tracking spans to point back to origin definitions, etc., always feel like plumbing work until you see the error that explains exactly what went sideways. And once you do, the errors that don't feel like a frustration you are trying to cast your net over.