Go Data Structures (2009)
research.swtch.com
research.swtch.com
The classic slice mistake is to forget that append doesn't necessarily return a completely new slice backed by a disjoint piece of memory, especially easy if you've been using a language where variables are immutable by default.
The problem is that it will often accidentally work; if you over-run the backing store, you will get a new piece of memory:
a := make([]int, 1, 1) // a is backed by a one-element-long piece of memory
a[0] = 42
b := append(a, 17) // this must allocate a new piece of memory
c := append(a, 37) // this must also allocate a new piece of memory
// At this point a == [42], b == [42, 17], c == [42, 37]
a := make([]int, 1, 10) // a is backed by a ten-element-long piece of memory
a[0] = 42
b := append(a, 17) // no need to allocate: just use the existing memory
c := append(a, 37) // no need to allocate: just (re)use the existing memory
// At this point a == [42], b == [42, 37], c == [42, 37]
// ^^^^^^^^^^^^^https://build-your-own.org/blog/20241125_go_slice_surprise/ goes a bit further, but this is really the point where I decided I wouldn't be able to stand Go.
I think you want Rust.
In Go, this garbage collector frees people from thinking about freeing memory. But it doesn't free people from thinking about owned vs borrowed memory. And that's where bugs can happen.
If you created the slice, you control it and you can modify its elements, append to it, etc.
If you didn't create the slice, you don't control it. If you want to modify it or append to it, you should copy it first.
This has reflected how I've seen Go used professionally by people experienced in the language. The language could offer more help (but probably never will), but it's not hardly a regression from C. The real regression from C is the lack of const pointers.
type Colours (Red, Blue, Green)
//....
c := Colours.Blue
a := pred(c)
b := succ(c)
for x := range Colours {
//....
}
Unfortunately afterwards they decided to follow Wirth's quest in Oberon-07 minimalism, instead of Active Oberon.
People have to look at Go alternative languages like Goplus, V (Vlang), Borogo, etc... for wanted features or greater responsiveness to user demands.
You can introduce data structures and types that utilize enums, with some languages taking that idea further than others, but that's well beyond enums themselves.
You may want to meet a modern language one day. Or hell not even a modern language, even a shit one like Java.
As it turns out there are languages where you can define an
enum foo { bar, baz }
Like what? CrabLang? Its enums are identical to Go, unsurprisingly. Above the enum rests a discriminated union that ends up hiding the details of the enum. That is where things begin to differ greatly. You have to do some real trickery to actually get at the underlying enum result in that language. But, when you do, you see that it is exactly the same.
> and when you type a parameter as a `foo` you're not at risk of getting 69 or 4328.
That's thanks to the type system, though, not enums. Enums are not a type. Enums are a number generator. Hence the name.
What's happened here is that you've mistaken "Things I believe" for "What everybody else believes" but you're talking to other people, not yourself, so, this makes you seem like a blithering idiot.
The particular version of this trap you've fallen into is a variation of the "It's all just the machine integers" mental model from C. It's just a model and the problem arises when you mistake that model for reality.
Now, technically this model isn't even correct for C's abstract machine, but it's close enough for many programmers and it tends to match how they think the hardware works, which is even more confusing for the hardware people who know how it actually works but that's another conversation.
This model is completely useless for languages which don't have the same type system, and so it's no surprise that it immediately led you astray.
No, I am clearly talking to a computer program. It is possible that the program is forwarding the discussion on to people. Perhaps that is what you are trying to allude to? The details of how the software works behind the scenes is beyond my concern. There is no intention to talk to other people, even if the software has somehow created that situation incidentally. If I wanted to talk to people, I would go out and talk to people, not type away at my computer into a box given to me by the software.
> The particular version of this trap you've fallen into is a variation of the "It's all just the machine integers" mental model from C.
As much as I enjoy your pet definition that you've arbitrarily made up on the spot here, the particular trap I have fallen into is the dictionary. It literally states what an enumeration is according to the prevailing usage. It does not describe it as a type, it describes it as the action of mentioning a number of things one by one. Which is exactly what an enum does.
The previous comment is talking about type constraints. You can definitely constrain a type such that it is invalid to use it outside of the numbers generated by the enum, just as you can constrain a type to only accept certain strings. e.g. from Typescript: `type Email = "{string}@{string}"` This idea is not limited to enums.
That's definitely a thing, and definitely a thing that could be used in conjunction with an enum, but is not an enum itself. Not as enum is normally used. Of course you can arbitrarily define it however you please. But, if you are accepting of each holding their own pet definition, your comment doesn't work. You can't have it both ways.
I agree with your point about Go enums.
But in defense of Java, modern Java is actually pretty pleasant.
Virtual threads, records and sealed classes, pattern matching, state-of-the-art garbage collectors, a great standard library etc. (and obviously well-behaved enums).
Not to mention the other languages you get for free with the JVM ecosystem.
It might not be as expressive as Rust, but certainly Java/JVM > Go.
encoding/json vs
Python: 2008, json package added in version 2.6
.NET: 2019, System.Text.Json came with .NET Core 3.0
JVM: still nothing
Python: gave up, "just use requests"
.NET: also 2012, HttpClient came with .NET Framework 4.5
JVM: 2018, HttpClient came with Java 11
What about actually counting the set of features across a full stack application?
I can pick countless of examples of features in Java and .NET standard library that Go 1.24 still doesn't have any answer for.
Of course, plenty of high-quality third-party libraries exist in the other languages. I've used commons-httpclient and jackson in Java, and everybody and their brother knows about requests in Python (though I prefer aiohttp nowadays). But they are odd omissions from the standard libraries of those languages. At least .NET caught up.
Which, I guess, is why we have this desperate attempt to redefine what an enum is, having become bored with a term that has no invitation potential. But, unfortunately, we have not found shared consensus on what "enum" should become. Some think it should be a constraint, others think it should be a discriminated union, others think a type declaration, so on and so forth.
All of which already have names, which makes the whole thing particularly bizarre. You'd think if someone really feels the need to make their mark on the world by coining "enum" under new usage in the popular lexicon they would at least pick a concept that isn't already otherwise named.
It is true that some type system features built upon enums. Like a previous commenter mentioned, Pascal offers a type that constrains allowable values of that type to be within the values generated by the enumerator. Likewise, I mentioned in another discussion that in CrabLang the enumerator value is used as the discriminant in its discriminated union types, which achieves a similar effect. I expect that confuses some people into thinking types and enums are the same thing, which may be what you are trying to get at, although doesn't really apply here. The difference is known to those reading this discussion.
The biggest problem with this desperate attempt to find new meaning for "enum" is: What are we going to call what has traditionally been known as an enum? It does not seem to have another word to describe it.
It's not like you could have appended twice to the same List in Java and expected to get two disjoint arrays, unless you copy first, and `slices.Clone()` makes that easy in Go now too.
If all I need is a slice of memory, I'd rather only have to pass around the two pieces of data (pointer and length) than all three.
This type isn't either of those things, it's a hybrid, and Go tries to get away with that because it's a garbage collected language so people don't need to care as much about ownership.
The growable array type (what you call "vector") is a venerable data type, although it does still have parameters you might reasonable tweak e.g. what's the correct growth factor? Doubling is popular but e.g. Zig chooses new_size = (old_size * 1.5) + 8 and you'll see disagreement about what API should be offered and why - e.g. C++ std::vector has the wrong reservation API
But this thing clearly isn't a mistake, it's intentionally a hybrid, and if it had been a huge success maybe everybody else would scramble to copy it.
A mutable String or Vec has all three fields, but the borrowed slice or str has just pointer and length.
Go chooses to be simpler by not having those kinds of separations, which also makes it a little less efficient under some circumstances.
a := make([]int, 1, 10)
a[0] = 42
b := append(a, 17)
c := append(a, 37)
// Surprise! b and c both equal [42, 37]
While I understand the performance benefits of reusing memory, language fundamentals shouldn't require developers to constantly think about backing array capacity to avoid bugs. This is especially problematic because it "usually works" until it doesn't - the worst kind of bug.
Modern languages have shown us better ways. Rust makes ownership explicit. Immutable-by-default languages prevent such surprises entirely. Even if Go wanted to maintain performance, they could have made this behavior more explicit - perhaps with separate "append_safe" and "append_reuse" functions.
Programming language fundamentals should be like good UI design - they should do what you expect them to do. When core operations have surprising edge cases, it creates cognitive overhead that ripples through all code written in that language.
``` a = append(a, elem) ```
If you assign it to anything else, `a` might still remain with the older slice memory which will cause worse problems than equality comparison.
So the kind of code you wrote is hardly ever written in production.
The advantage of having this is avoiding one more heap allocation for slice, which is a good tradeoff if you ask me.
`foo, err := callSomeFn()`
is followed by
``` if err != nil { return err } ```
too; it’s the ones that get missed that are the problem because convention will do nothing to prevent silly mistakes.
Like many things in Go, there are steps the language could take but chooses not to. For example, Swift implements copy-on-write for variable-length arrays. I’m not qualified to comment on whether the Go team is right or wrong in their decisions, but these aren’t unsolved problems.
While most people highlight the difficulty of picking up the syntax, I find Rust to be an incredibly tedious language overall. Zig has a less expressive type system, but it compiles much faster (though not as fast as Go). I like what Zig and Odin folks are doing over there.
I like the balance Go strikes between developer productivity and power, though I dearly miss union types in Go.
Don't make the fallacy of conflating Rust's slow compile time with its "advanced" (not really, it's 80's tech) type system. Rust compilation is slow for unrelated reasons.
Personally, I write java at my day job and the type system there makes me loooong for rust.
But, it's just a tool, and the tools I choose reflect the type of stuff I want to build. The JVM is extremely impressive in its own right. You're just not going to to find any one runtime or ecosystem that hits every niche. I'm happy to leave the language favoritism to the junior devs—for the vast majority of situations, what you're building dictates which language makes the most sense, not vice versa.
I've been involved in a few successful large scale projects and never felt like the type system of Go is holding me back too much. Sure the error handling could be better, union type would make interface munging easier, but the current state after generics isn't too bad.
Last time I checked, C# had clean and focused syntax for working with collection types. Could you provide an example?
It's just that compile times and DevEx haven't been a priority for most projects.
But sure, LLVM and interfacing with it is quite possibly a big contributor to it.
However, it is actually a good example regarding tooling, as the Haskell ecosystem has interpreters and REPL environments available, for quick development and prototyping, something that is yet to be common among Rustaceans.
Ideally we would be having the F# REPL/JIT, plus Native AOT for deployment, as comparable development workflow experience.
Naturally F# was chosen as example, because that's your area. :)
Not being negative per se, I also would like to have something like Haskell GHCi, or OCaml bytecode compiler, as options on rustup, so naturally something like this might eventually come.
That can be addressed by passing the slice as a pointer: https://go.dev/play/p/h9Cg8qL9kNL
Slices are passed partly by value (the length), partly by reference (the data).
func takeSlice(s []int) {
slices.Sort(s)
}
This can have other quite confusing gotchas, like:
func f(s []int) {
_ = append(s, 1)
}
func main() {
s := []int{1, 2, 3}
f(s[0:2])
fmt.Printf("%v\n", s)
}
I can see why it trips up newcomers, but it feels pretty basic otherwise.
The fact that I can pass a slice to a func 'by value' and mutate the source slice outside the func is already surprising behavior to most people. The fact that it MIGHT mutate the source slice depending on the slice capacity is the part that really drives it home as bad ergonomics for me.
Overall I enjoy working with go, but there are a few aspects that drive me up the wall, this is one of them.
Make it so you can create copy-on-write slices of a larger slice, and a huge number of bugs go away.
Or do what rust did, except at runtime, and keep track of ownership
s := []int{1, 2, 3}
s[0] = 0 // fine, s owns data
s1 := s[0:2] // ownership transferred to s1, s is now read-only
s1[0] = 1 // fine, s1 owns data
s[0] = 1 // panic or compiler error, s1 owns data, not s
It's true that little has changed, but very little is changing in the data representation of any language, really. Even ones that are evolving rapidly.