[ndriscoll]

In the overwhelming majority of cases, you want to trivially send the error up the stack to be handled, so you don't want to incur the noise in the middle of your business logic (and indeed doing so obscures the times when error handling is non-trivial!). e.g. Scala handles all error cases with compile time checking (in fact it tends to take error handling much more seriously), but you don't have to write (or read) the trivial case.

[kbolino]

> In the overwhelming majority of cases, you want to trivially send the error up the stack to be handled

Maybe so, but for JVM languages, this "trivially" includes a stack trace. Go doesn't offer that for normal errors and there was no clear path to getting it. The need for contextual clues when debugging, and the community and then stdlib settling on error-wrapping as the way to do it, makes "simply return err" not a sane default anymore.

[ndriscoll]

That's just pointing out another layer of noise Go introduces. Here again in Scala, you can get stack traces for free with no extra syntax with value-based errors. There are already solutions for doing these things that have been used in anger for many years.

In a language that takes error handling seriously, you have domain types for errors, type checking for exhaustive error handling, and you get info like a stack trace out of the box.

[kbolino]

Ignoring the stack trace matter, which is explored in the sibling thread, it seems to me you're really taking issue with two of Go's language design choices that aren't specifically about errors: structural typing and the lack of proper discriminated unions.

There is nothing special about an error in Go, and there is no type hierarchy to anchor them against anyway. An error can be a struct{} carrying no value at all, or simple integer with no way to carry context, etc. Error-wrapping is the solution that has been devised to work within these constraints, but there's no obvious way to compose that with the various ideas around reducing boilerplate. For a point of comparison, Rust's anyhow crate was able to square this circle by adding a blanket trait impl for Context, but Go has no equivalent for that. (Also, Rust's ? operator is implemented using unstable components, allowing the details to evolve even while the feature itself is usable in stable Rust, something which also has no equivalent in Go).

As to the lack of discriminated unions, I agree wholeheartedly. There have been some proposals recently to get them into the language, with various levels of conservatism and cohesion with the rest of the language, but a lot of them are faltering on the "every type must have a well defined zero value" issue (yet another design decision that doesn't necessarily have anything to do with errors).

[ndriscoll]

As an aside, the "every type must have a zero value" is another level of insanity akin to pi must be 3: tempting to want it to be true but disastrously wrong. You're basically making known good practice (RAII/make invalid states unrepresentable) impossible, especially when you combine it with the lack of if-expressions. It also breaks parametric reasoning since now a function `() => a` exists.

I had to argue about this years ago with a coding standard pushing to define local variables at the top of functions and zero initialize them in C. Why default to null pointers or worse (invalid ints/structures) when you could make use of uninitialized variables be a compiler error instead!?

[kbolino]

Technically, Java has the same rule, it's just that all user-defined types are reference types, and so the zero value for them is always spelled "null". I'm not sure how Scala pretends otherwise, but I have worked with Kotlin, and non-nullable object fields and method arguments are always a bit of a lie, since the underlying JVM machinery permits, and defaults to, null. (Despite the original authors of Go being unfamiliar with Java, they do seem to have settled on a number of the same decisions.)

Personally, I'm fine with discriminated unions always allowing "nil" in Go, since IMO the most sensible way to implement them anyway is as a special case of interfaces. Not everyone feels the same way, though.

[9rx]

Every value must have a zero value was congruent with the dynamic languages of the time, which is signifiant as Go was designed to "feel like a dynamic language" that was able to handle production scale (i.e. something that could handle Google-sized loads). There is a good possibility that Go wouldn't have even had a static type system if they had figured out how to make dynamic typing fast.

That may seem out of place in the world you live in, but Go was created in a world where Python, Ruby, and Javascript were the languages all the cool kids were using. Go was intended to be the language for C++ programmers who had to use C++ for performance reasons but wished they could use Python instead.

Insanity, perhaps, but Go was designed to solve a problem for the world it lived in. You don't get to choose your circumstances. For those living in paradise, they can use the languages designed for their world.

[win311fwg]

> Go doesn't offer that for normal errors and there was no clear path to getting it.

It does offer that for normal errors if you pass those errors via exceptions. Although you're right that using exceptions for error handling is pretty weird and generally avoided in Go, much to the chagrin of Java developers.

The Upspin experiment, which eventually shaped the error additions that did make it into Go, showed a clear path to getting it in errors themselves, but it also revealed that nobody actually uses them when available. The research found that even with stack traces attached, developers didn't bother looking at them. Thus why that bit didn't make it into Go.

[kbolino]

Yes, I don't think stack traces are a silver bullet. They tell you the path taken through the code, but rarely do they tell you why that path was taken. I don't know what Scala stack traces look like specifically, but I've done my fair share of parsing Java stack traces, and they usually just get you in the vicinity of the problem without really telling you the cause. This was best illustrated with the standard exception that was thrown when a hostname failed to resolve, where the message did not contain the hostname. This one was solvable (and did get solved AFAIK), but in a lot of other cases, the relevant context is not available at the depth where the exception is thrown, necessitating exception-wrapping.

[win311fwg]

> In the overwhelming majority of cases, you want to trivially send the error up the stack to be handled

There is no case where you would ever trivially send the error up the stack.

First, if your function doesn't have any meaningful information to add, you need to seriously question why the function exists. It is almost certainly not adding anything of value and the function that originally produces the error can just as well be called directly. No need to trivially pass an error when the useless "middleman" function doesn't exist.

Second, even if you still decide that a function that doesn't have worthwhile information is still worthwhile having, if you pass someone else's error straight through then you become forever coupled to the implementation of someone else. If you, for example, pass a SQLite error and then later decide you want to use Postgres, now you have to emulate SQLite errors every time Postgres produces an error, bringing you back to square one while also having one hell of a confusing API. Unless you hate other programmers and are purposely trying to make their life a living hell, the only sensible approach is to produce new error values at each step so that you can maintain stable sentinels/types.

A good language solution for errors might make it look like you are trivially passing an error up the stack while using some other method to attach the necessary metadata, like we see in some other languages, but nobody has yet figured out how to make that fit into Go. It is generally agreed upon that Go would greatly benefit from this, but until someone rolls up their sleeves and solves it...

[ndriscoll]

Functions exist first to give names to conceptual blocks, and should generally be defined by "level of abstraction." i.e. you give a name to some task that then perhaps calls some low level data validation or manipulation functions which can return errors. Your business process might not care which task had an error, so it just bubbles up to the higher level coordinator (e.g. a handler just returning errors to a router).

You do not need to have new wrappers at every level. That's what interfaces are for. e.g. if all you're going to do is retry up to N times and log, and you don't need special handling, you just call String() or whatever. But you still want libraries to return domain errors so that you can handle them if you need to, or for tests, etc. And you can have some layer in the middle to make like SQLError or DatabaseError.

Right I'm saying Scala does exactly that, and has libraries that e.g. propagate stack traces across fibres automatically and invisibly. There's prior art to observe. It's not some mysterious unsolved problem.

[win311fwg]

> Functions exist first to give names to conceptual tasks

Right, and conceptual tasks will almost always have information that is usefully attached to an error. There are exceptions, but even in those exceptions you still have to ensure you don't leak implementation details, so you still cannot trivially pass the error through.

> Your business process might not care which task had an error

That's true, but you would never write upstream functions with the business process in mind. That is for the downstream functions to determine. Doing so would couple an upstream function to the business process, making it difficult (maybe even impossible) to use again when the next business process wants to reuse it, which would be complete insanity. A function is written to provide what makes the function useful for the sake of the function itself. It is up to the caller to decide what facets of that are useful for the business process being implemented downstream.

> You do not need to have new wrappers at every level.

You absolutely do need them. The language can work to hide that they are being created, but that does not preclude the necessity of them. This is most definitely a solved problem in other languages, but it remains that nobody has figured out how to take those ideas and apply them in Go. What was done in another language cannot be lifted wholesale and simply dropped into Go. The world is not quite so simple. It remains an unsolved problem in Go.

And understandably so. Who has time to work on solving the problem when there are comments on HN complaining that nobody has solved it yet to be made?