I agree with this, and I'd add there are two modes of processing errors: fail-fast (stop on first error) and fail-last (do as much processing as possible, collecting all errors). The later is what you want to do when, for example, validating a form: validate every field and return all the errors to the user.

Nice article. I agree with the differentiation. You could also classify them as errors that should be fixed and errors that should exist. Some would argue that validation errors are not really errors on a system level. They are only errors on the user level. On the system level they are a feature

This post seems to conflate using `throw`, `raise`, etc. with crashing. The idea that 'handling' an error does not involve `throw`/`catch`, `try`/`except` is very strange to me. The exception facility is often the most elegant way to check inputs, and if I remember correctly the Python documentation says as much as well.

It makes me think that it's worth sitting down and considering what all the valid outcomes for a piece of functionality are. A user typing in a string in the wrong format is not necessarily "exceptional", whereas running out of memory while getting the input would be. I feel like programmers too often treat perfectly valid outcomes to be errors. For example, in Rust I'll see Option<Vec<Foo>> and I ask myself if we could just use the empty vector as a valid return value.

Another interesting article on error handling: https://www.dgtlgrove.com/p/the-easiest-way-to-handle-errors

I'd also add errors with third-party systems, which aren't the developer's or the user's fault, but which are probably worth handling nicely (e.g. retry with backoff).

This is 4XX versus 5XX errors in HTTP.

It's common that if you extract a function or a module with a well defined interface, that function or module can't tell whether a "bad" input indicates an expected or an unexpected error.

For example division by zero often indicates an "unexpected" error, but it wouldn't if you were implementing a spreadsheet.

So to me the approach of using different forms of error reporting for the two kinds of error doesn't seem promising: if you imagine you had to implement division yourself, which kind of error should it report? Should you have two variants of every fallible function so the caller can choose?

The expected/unexpected distinction assumes you have the capacity to anticipate failures in the first place. But that capacity varies - even among experienced developers, everyone has blind spots shaped by their specific history. What's "obviously expected" to one senior dev is a surprise to another. The article's model is useful, but there's a prerequisite it doesn't address: the ability to expect is itself unevenly distributed.

I don't really like this article. There isn't anything particularly noteworthy to noticing that some computations have outcomes that allow some form of recovery, and other outcomes do not.

But there are some obvious follow up questions that I do think need better answers:

Why is recovery made so hard in so many languages?

Error recovery really feels like an afterthought. Sometimes that's acceptable, what with "scripting" languages, but the poor ergonomics and design of recovery systems is just a baffling omission. We deserve better options for this type of control flow.

Also, why do so many languages make it so hard to enumerate the possible outcomes of a computation?

Java tried to ensure every method would have in its signature how it could either succeed or fail. That went so poorly we simply put everything under RuntimeException and gave up. Yet resilient production grade software still needs to know how things can fail, and which failures indicate a recoverable situation vs a process crash+restart.

Languages seem to want to treat all failures as categorically similar, yet they clearly are not. Recovery/retry, logging, and accumulation all appear in the code paths production code needs to express when errors occur.

Following programming language development the only major advancements I've noticed myself have been the push to put more of the outcomes into the values of a computation and then further use a type system to constrain those values. That has helped with the enumeration aspect, leaving exceptions to mainly just crash a system.

The other advancement has been in Algebraic Effects. I feel like this is the first real advancement I've observed. Yet this feature is decried as too academic and/or complex. Yes, error handling is complex and writing crappy software is easy.

Maybe AI will help us get past the crabs at the bottom of the bucket called error handling.

This sounds very neat in theory, but in practice errors are a continuum between these two extremes and there isn't really a clean dividing line.