From the end:

> Pithy statements such as “only use exceptions for exceptional cases” or “Let it crash”, while catchy, do not do justice to the complexities that programmers need to deal with when deciding how to define, propagate, handle and reason about errors.

I think the author is on to something here w.r.t. how current programming languages and conventions brush over what good code might need.

I searched for conditions and restarts. I did not find them.

That makes me sad because that is what my systems programming language uses.

For context, conditions and restarts come from the Lisp world. They do not unwind the stack. Instead, they run an error handler on error, and if it handles the error, you have the option of restarting the operation that failed.

I have implemented them in C; they are great for systems programming.

Very nice article. I agree with the theses and the resulting design makes sense.

Regarding context handling, one issue that I struggle with is where the context should be attached: by the caller or callee? In particular, in the context of intermediate libraries, I would like to limit needless work attaching the context if the caller already handles it (some sort of opt-out).

Usually I just attach the context in the callee as it's the simplest, but I'd like some systematic/reusable approach that would control this. It would be nice if it fits in the type system too. I use a bunch of languages like Rust, TypeScript, Kotlin, Python and it feels like a larger design issue.

To give a more concrete example. In Node, `fs.readSync(path)` includes the error path when it fails: the callee attaches the context to the error. In Rust, `std::fs::read(path)` does not attach the path to the error: the caller is responsible for the context. I would like some lightweight way to control if the file read should include or not the path. The ancestor scanning example from the article is a case where caller context is good IMO, but usually callee context is better for unrecoverable errors. Since it's contextual, having control would be nice.

> Almost all catastrophic failures (92%) are the result of incorrect handling of non-fatal errors explicitly signaled in software

So in a laps of 15 years (2010-1025), they hand picked 20 bugs from 5 open source filesystem projects (198 total), and extrapolated this result.

That is not science.

Missing piece: Common Lisp continuable condition system.

It sometimes feels like there is a strange resistance to defining ad hoc sum types. If your language has good pattern matching capabilities and an expressive syntax for declaring data types, then this scales very well.

For example in a pseudolanguage:

    data FooResult = Number | Error
    def foo(n: Number): FooResult = ...
    foo(42) match {
      case n: Number => print(n + 1)
      case e: Error => // handle error case
    }

This gets you very far most of the time. If the tag overhead matters (which should largely only matter in a hot loop), then try to refactor so that the hot loop doesn't error out.

Granted in many languages the above cannot be expressed succinctly and is often avoided for this reason.

    data FooResult = Number | Number

    ErrorA = FileNotFoundError | ParsingError
    ErrorB = FileNotFoundError | PermissionError
    ErrorC = ErrorA | ErrorB // Oops, trying to ad hoc sum FileNotFoundError twice

You know, there is a programming language based on a PhD thesis on fault tolerance. This programming language is built around a VM that implements lots of great design on error handling. And there is a functional programming language that compiles to the bytecode that this VM runs. It is a delight to use, gives you all the fault tolerance goodness for free, but the name is often cursed in HN threads for some reason.

edit: I might be giving some random episode too much weight. Seems like Elixir and Erlang are actually quite well liked around here

What a great resource.

I’ve been in the errors-as-values (like Rust and Haskell) camp for over a decade. But I think you need to go a bit beyond that philosophy (i.e. not just as a slogan) in order to truly unlock what treating errors as values can be like.

Errors as values are just values. Not any more complex than other values. But except in cases like `None | Error` (no regular value or an error) you at least double the work that you need to do. The thing about the “happy path” is that it’s just one channel of information. With errors you layer on one more channel.

I think that has lead me to be careful about treating errors in the same way I treat other kinds of values. Because simplifying errors down to just one value, like a string, simplifies the whole error channel. But how is error-as-string useful if you are using the code as a library and not just erroring out and letting the end-user/operator read it with their primate brain? Well that lead me to add variants, sum types for all the different error possibilities. Along with associated data like how this int was larger than some exected bound.

Already here I run into a sort of sub-error problem: only advertising what variants of an error that I can return from a function. It seems that this isn’t supported directly in Rust, which is what I was using at that time. So if I just advertise that I can return all the error variants (which is a soft lie) then I might have to deal with them downstream.

More expressive error values seem to both (1) burden the implementation (sub-errors) and (2) the clients who have to consume them.

So do I use less expressive errors so that they become easier to handle for client code?

Well it seems here that I’m still stuck treating errors as special values. Because I still seem hesitant to treat errors in their full generality as values. Think of a value:

1. It’s not just a single “atom”/scalar like an integer

2. It can be in a list or a dictionary/map or on a stack or some other data structure

3. It can be transformed with a map or reduce operation since you either want to transform the error or leave out some details

Maybe the library should provide a list of errors. But the client only cares about the last one. So it takes the head of the list. Maybe the library should provide a detailed error containing associated data about the bounds, expected and actual. But the client just cares about the error type and elides the rest.

Maybe this doesn’t seem useful for errors. But take the `Hurdle` type in the article. These are effectively a mandatory (product type) list of warnings. Maybe the client code only cares about five out of eight of them so he filters out the rest. Or he doesn’t care about them at all so he just ignores them.

And this seems to lead into iterators and lazy evaluation. In high level languages at least we’ve come to a point where we value pretty functional, declarative, and coarse transformation of data. “Coarse” in the sense that we use a few datastructures and operations to process the data without worrying about special datastructures and micromanaging control flow. But crucially these constructs don’t have to allocate N intermediary collections. They can be lazy. So maybe errors (and warnings (hurdles)) should be as well. Maybe you can have a very rich error/warning declaration while only generating them when the client code wants it. And what happens when you make a nod (at least) that all the error computations can be elided if only, say, you just kill the application on the first sight of any problems (like in a script)? People are incentivized to treat errors as any other value without fearing that it is wasted effort (computation).

Now on the other hand I do program in Java. And personally, philosophically, I am fine with exceptions for the type of application advocated in the article: report errors way up the stack. But there I’ve been running into the same problem: treating exceptions as not-quite-values. For any other custom class I would create however many fields and methods I need. But for custom exceptions I just give them a name. And maybe pass a string to the constructor. Which I could do for a built-in or third-party error. (So effectively I just give them a new name.)

But why? There must have been a block in my mind. Because Java exceptions are just classes. And they can have fields. And ultimately we log errors that bubble all the way up. So why not record all the associated data as fields on the exception and use the error handler (with `instanceof`) to gather up all the data into a structured log? That should be a ton better than just creating format strings everywhere and sticking the associated data into them. And of course you can do whatever else with the exception objects in the error handler; once it’s structured you can use the structure directly, without any parsing or indirect access.

    (X = A + B, Y = X + C, Y = A + B + C is false)

Is this an "epic treatise" ?

Really nice article that covers quite well the complexities involved in dealing with errors. "They're just another kind of value" and "Just crash and restart" are both wrong, error handling is much more complex than that.

Different kinds of errors need to be handled in different ways in various contexts as the article points out.

I won't comment on the error model proposed in the article as I find that without actually trying out stuff like this it's very hard to get a "feel" for it, but at first glance it looks a bit over-engineered.

I think the gist of what makes a good error handling system is as follows:

Certain errors are contract violations/bugs that should not occur in normal program operation, and should not "taint" the code with explicit error handling. The language should provide a mechanism that allows to abort the current "task" when such an error occurs and trigger some error handler. Exceptions are good for this, but it falls on the programmer to define a "task boundary" appropriately. Defensive try/catch all over the place is not that, and is usually done because...

Certain errors are part of normal program operation and the programmer should be reminded to deal with them. Exceptions are terrible for this but Result<T,E> types are great. Having to use Result<T,E> in situations where the error is actually not normal part of program operation (e.g., you're accessing a key on a map that can only be missing due to a bug) is annoying but is easily solved by providing two separate APIs (like Python's __getitem__() and get() methods on dict).

Often times an expected error has no good way to be dealt with at the point it occurs, so there should be an easy way to just abort the current "task" when that happens. This is your ".expect()" or just a giant pile of ? propagation in Rust but Rust panics don't carry enough information by default like you can with an Exception type. Note that is different from the two API version above. You use the "panicking on error" API only in cases where the error is not expected, whereas you use some .expect() equivalent as a way to abort the current task for an expected error.

Sometimes you also want to collect multiple errors and abort only after every error is collected, like in a typechecker. This is often done with "NullObject" patterns or "ZII" (most of the time handled very poorly) but there are ways to do it right where the error case "poisons" the data up to the task boundary but cannot propagate further than that (turning into a Result<T,E> equivalent at that point).

The most important part is understanding task boundaries. I find most programs (including my own) do a very poor job of this.

The information contained in the error (sometimes called the context) is very tricky to deal with as what constitutes relevant information for a library or for an app using that library can be very different and there's a performance cost involved with its collection.