Hacker News Reader

Comment by FullyFunctional | Hacker News Reader

I'm definitely in the camp that worry about the feature/complexity creep. I saw it with Haskell. Haskell 98 is a pretty nice and simple language. GHC Haskell is a monster. This matters when you try to read other people code and the overuse of experimental cool new features becomes a major burden for comprehension.

Rust is already a very large language and for intrinsic reasons pile on a load more complexity than Haskell (just look at the iterators).

I write this as a fan and advocate of Rust.

View on HN

Rust is necessarily a comparatively large language, but I’m content that at present it’s not getting noticeably worse: additions are mostly filling things that could be expected but are missing, plus some focusing on improving ergonomics.

Take generic associated types, for example: they’re a major new feature… but are conceptually just removing a potentially surprising limitation.

Or const generics: they really are a major new feature, but in general the code that can use them was a good deal more complex and inconsistent before.

You’d have a stronger argument about complexity creep on async—there was lots of demand for it, but it’s not clear that the design that is now frozen was in fact ideal.

One of the annoying things though is lots of the fills are rather inelegant extensions of APIs. They’re not nonsensical, and TBH I don’t really have a solution, but e.g. allocators support (between custom allocators and faillible allocations) ultimately doubles the size of Vec, or near enough. This makes going through the API a major chore.

But maybe this would be better fixed through Rustdoc, by supporting topics / sections, and being able to “globally” set which standard topics you’re interested in e.g. I don’t do embedded so generally unsafe and explicit allocations I’m uninterested in, that stuff could be folded away / hidden in both the sidebar & main text.

I feel like rustdoc has been going through cycles of being useful, then growing beyond what the current design can handle well, and needing major changes again. It’s currently not good for discovery in std, because each type has far too many methods and each method has far too much text, with examples added to everything.

That's not wrong.

Wrt examples I feel like the examples are mostly obvious / uninteresting and thus have little value, but I’ve been in the field a while so that may well be an experience bias.

Alternatively some examples show the neat bits but not clearly because they’re artificial e.g. HashSet::insert demonstrates that it returns a bool via an assert_eq, you have to figure out how cool and useful that is (then gripe that other langages don’t give this info) separately.

Even uninteresting examples are absolutely crucial. I've taught Rust to many, many people, and one of the things that new users rave about is how the docs provide an executable example of every API, no matter how trivial. This especially matters for highly generic functions (which are very common in libstd), where the signature is extremely imposing but the usage is surprisingly simple.

Cannot agree hard enough. Nothing is more deeply infuriating than api documentation that just expects the reader to understand what the author means. The trivial examples of everything means I never feel like I’m guessing when it comes to rust’s standard library. I dearly love working with Python and it’s many included batteries but plenty of the most complex and by extension powerful parts of the Python standard library need the kind of “show me how to do everything this is capable of” type example documentation that Rust uses. Last notable one for me was the documentation about changing stopitetation to an explicit error, and how the behaviour differed pre and post change. It took me far to long to work out what changes I needed for some library code to be updated because the documentation demonstrates a few good examples but does not cover all permutations of the behaviour they changed. It’s a low level feature so obviously they can’t show everything it can be used to do but they should have at least tried to cover all the different fundamental behaviours of the code. This is something the community are doing better with now, in part because of the excellent example the rust community is setting for documentation of a programming language.

I'm so glad to hear this.

My guiding principle for the absolute minimum of API documentation is "If I read the documentation and want to use the thing I read, I should be able to copy/paste something from right there to get started with."

This isn't perfect (trait implementations...) but it's nice to see that it's been as helpful as I hoped it would be.

Plus in Rust, the examples are also unit tests. So you can do TDD and flesh out the documentation while you're doing it.

if you have better example suggestions, you can easily make a pull requests with them.

I feel like the pace of Rust language evolution slowed down significantly after the 2018 edition and the work on async shortly thereafter. Most of the more recent language changes feel like pretty minor ergonomic improvements, things which always should have worked. I suppose we got const functions (can be evaluated at compile time), and const generics (allow you define fixed-length vectors for size N). But those features are well established in other languages and they don't interact much with anything else. And I'm sure I'm forgetting a few things, but that's sort of my point—subjectively, not a lot has changed for me recently.

What I have seen is a lot of good work in the larger ecosystem. And the tooling continues to improve.

I often see people complaining how bloated or large the language is, but I never see specifics. As someone who started learning and using Rust at the beginning of this year I was surprised to find out how simple the language is given its reputation on forums. What makes you feel the language is becoming too large, where is the feature creep, and where is the complexity you're talking about? Are you talking about the stdlib or language features?

Some people (including me) felt how async was added was worrying.

Adding ".await" means when I teach Rust we now need to say "x.y is member access. Unless y is 'await'. Then it's something totally different". That's the type of strange rule that C++ has, and if you gather enough of them, languages become very hard to teach.

However, async seems to be the only big language feature which effected many things which has been added.

.await is a postfix keyword - it effectively means "crazy control-flow (monadic) magic is happening here". "?" for failure handling is another bit of postfix syntax, that means something very similar. A good IDE will show it with a different highlight, so that it will never be confused for a struct member or method call.

This stuff was discussed by the community for a long time - bikeshedding about syntax is a well-known "trap" in programming language design. It has turned out to be (IMHO) reasonably intuitive, and preferable to the alternatives.

My biggest problem with await is it's very non-obvious if you don't know to expect.

Seeing a '?' is obviously a "new thing", so people go look up what it is (in practice, I would expect any Rust programmer to learn '?' really early, but if they somehow miss learning it, they will know to go look it up).

I've seen beginners assume they were "missing the await member". Of course, once you know it's something new you can learn it. I would have prefered something like .!await, just to make clear it's a "new thing".

Isn't "?" just a sugar for something like the following? That's how I think about it in my mind at least.

  // with "?"
  let v = x?;

  // without "?"
  let v = match x {
    Ok(o) => Ok(o),
    Err(e) => return Err(e), // exit current scope due to the return
  }.unwrap();

That feels like something that could be done with a macro.

It's more like

  let v = match x {
    Ok(o) => o,
    Err(e) => return Err(e.into()),
  };

> That feels like something that could be done with a macro.

Indeed. It used to be a macro called try!.

Ah, thanks for the correction, that's better yes

Yes, that's control flow magic. It was prototyped with a macro, but failure handling comes up all the time in practical programming, and it's important to make it as ergonomic as possible.

I see. I think reading your initial comment I took issue with the hyperbole "crazy control-flow (monadic) magic is happening here".

But yes it is a hidden control flow, and taking in account hidden control flows is inherently more complicated (or at least has a steeper learning curve).

That's a fair point. I haven't touched async yet but I've seen lot of discussions regarding "await" and that seems to create a good amount of confusion. Still we are multiple magnitude away from the level of complexity C++ has.

Yeah, the await syntax is just bizarre. One of the worst eyesores of any languages I've personally encountered.

An await macro or function would have sufficed but because Rust is still mainly a hipster language with a hipster community, they had to choose a hipster syntax after years of bikeshedding about how to make it absolutely perfect.

Prefix syntax has drawbacks that were constantly pointed out in that long discussion. In particular, it doesn't compose well in practical scenarios. .await just 'flows' better and in a more intuituve way.

Completely agree. I've used it at work now for about a year and the postfix notation flows perfectly with more intricate usages of futures. Especially with the sense that you construct one, then decide when to await it, or await a collection or whatever. Following the rust style of immutable variables and chaining function calls.

You can even await the same future multiple times in a select statement keeping it executing a bit further each time until it returns, if you do it by mutable reference. Really quite magic.

Now we just need async closures and some nice async style iterators. Something like FuturesOrdered or even FuturesUnordered in an inline iterator style allowing efficient nice composability. Without any of the pitfalls and gotchas those two currently have.

I think that might be what throws many people off? The regular style of method chaining and having the cold futures just be a dead piece of memory you can pass around until actually passed to the runtime?

> An await macro or function would have sufficed

We tried, and it literally did not suffice.

Works in other languages fine.

Different languages have different semantics. The challenges here are related to the ways Rust works specifically, as well as its goals.

Complexity is a death by a thousand cuts issue. I suspect most people complaining are going to have a hard time enumerating the many small sources of complexity they encountered.

Rust isn't a complex language by the standards of languages like C++. Complexity arises by features that work together in surprising ways. Rust has a medium-large amount of features (similar to Python (though obviously with a more imposing learning curve than Python)), but those features tend to compose very well, which keeps complexity from ballooning.

It is true that Rust is competing most directly with C++, but "less complex than C++" might be a true statement about every single language in existence that has more than a dozen professional users.

Kind of, Python, C#, Java have enough stuff on language + standard library to compete with C++.

A pub quizz with them would be just as fun.

I personally think those are all nowhere close. There are like ten different ways to initialize values in C++ and they all have slight differences.

Believe me, C# 10, Java 17 and Python 3.7, including standard library and runtimes across 25 - 30 years, have plenty of question material.

Rust doesn't reuse concepts well. The different capabilities and subsystems are all build from disjoint building blocks, which you could call modular in a sense, but it doesn't make them easy to use, learn, or joyful.

The module system, the macro system(s), the core language, the borrow checker. It's all different micro-languages without a shared foundation.

I think it's easier to when you compare Rust to simpler languages like Carp or Zig.

Carp due to its homoiconicity and lisp heritage allows you to write macros in regular old Carp. Zig doesn't really have Macros, but comptime (compile time) code evaluation, which results in something with a lot of the benefits of a macro system without the drawback of unreadable DSLs, because it too is plain old Zig annotated with `comptime`. Zigs module system is build on Struct name-spacing, so if you know Structs, you know the module system. Its build system is also build on Zig code.

My guess is that Rusts origins are partially to be blamed for this, the web is build the same way, disjoint standards that are developed by different groups of people. It's just that just like the web, the sum of the parts is less elegant and usable than each constituent.

Serious question about Zig’s comptime - isn’t that just like Rust’s const fn, except every function is implicitly opted in to being const?

Where this could be a problem - I write a comptime function, only using other functions that I’ve verified can be executed at compile time. But now the implementation details of those functions (that they’re comptime) has leaked to their definition. Now a change to the impl of those functions could break my code, without the authors of those functions realising it.

Perhaps this is not a problem in practice?

Kinda, but much more than that and the ergonomics are quite different as the entire language is basically available. For example in Zigs comptime, types are first class citizens which can be queried and operated on. This gives you generics without a lot of type system magic, and even const generics over arbitrary types. All without a macro system or Turing-Complete type checking (similar in flavour to the DeBruijn criterion: the comptime code/higher order proof system is turing complete but the type/core checking is straight-forward).

And yes the problem you describe could exists, but I've not encountered anything like it in practice. But the idea you had there captures the difference between Zigs and Rusts type-system/generics quite nicely:

> Rust tries to proof universally qualified correctness and behaviour of code, i.e. regardless of its context, while zig only proofs you the correctness of specific instantiations.

While this may sound like a bug, I'd argue that it's a feature, as it's YAGNI applied at a type level, and for example allows you to do things like:

  switch (builtin.target.cpu.arch) {
    .x86_64 => //do intel stuff
    .aarch64 => //do arm stuff
    else => @panic("unknown arch!");
  }

With the switch arms that don't apply simply being ignored and never type checked. This allows you to use a basic language construct where C requires a preprocessor and Rust requires macros or compiler magic.

The Rust solution for your arch variation specifically is hardly onerous, and does just the same, ignoring the code in not-matching branches:

  #[cfg(target_arch = "x86_64")]
  …

  #[cfg(target_arch = "aarch64")]
  …

  #[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
  compile_error!("unknown arch!");

(If you got to much more complex cfg branches, cfg-if could be worthwhile, which lets you write `cfg_if! { if #[cfg(…)] { … } else if #[cfg(…)] { … } else { … } }`. But for only a few simple ones, I prefer to keep it out.)

Yes, it can happen. In practice is not a big deal because the kind of stuff that you would want to run at comptime is pretty much all logic that manipulates its inputs without other side-effects, but one could for example add logging to a function and break its ability to be run at comptime (Zig might add in the future ways of solving this problem).

I would say that more in general the problem between public interface vs implementation details is much bigger than any type system and requires humans to negotiate what should be considered part of the public interface through other means (eg: tests, comments). Comptime is one example, another is ABI stability (eg the layout of a struct, or just its size), another could be speed or memory consumption.

Zig doesn't have an official package manager yet, we'll see what happens once we get one and people in the community have to start communicating stability guarantees to their users.

The problem with that is in large parts that Haskell 98 is very outdated, and so relatively basic and absolutely useful extensions are in the same pot as all the experimental stuff.

Haskell Prime, essentially the Haskell 98 successor, is supposed to fix that. The first thing I do in almost any Haskell project is enable a bunch of extensions that I consider absolutely essential[1]. Most or all of those should likely be incorporated into the new standard, and then the truly experimental stuff stands out as experimental again.

Unfortunately it seems that Haskell Prime efforts have slowed down, despite a very active Haskell community in total. I have not looked into details there, though.

[1] For example ScopedTypeVariables, where it's hard to imagine for me at least why this would not be the default.

I think the GHC2021 group of blessed extensions provides what you want.

https://ghc.gitlab.haskell.org/ghc/doc/users_guide/exts/cont...

That's indeed useful and something that I didn't know about, thank you!

ScopedTypeVariables by default! Excellent.

Out of curiosity, why do you find ScopedTypeVariables essential? I've written a fair amount of Haskell code and never once needed it.

Here's some explanation (not by me, found through search): https://blog.ocharles.org.uk/guest-posts/2014-12-20-scoped-t...

And among other things, it also enables pattern signatures.

The official docs at https://ghc.gitlab.haskell.org/ghc/doc/users_guide/exts/scop... are generally worth a read, they start out with an example of what lexically scoped type variables allow, and follow with the design principles for the extension.

They’re useful on occasion, but don’t seem anywhere close to essential.

GHC seems to be more of a platform for crafting your own language than rust. Yes, rust nightly has experimental features that can enabled, but they seem to have gone through more vetting and design, whereas in GHC the features seem to be much more exploratory, or more explicitly, research oriented. It seems as though every contemporary haskell program is written in a different language, though rust at least, always has stable, which seems a much more sensible default than GHC's.

> GHC seems to be more of a platform for crafting your own language than rust.

It’s always been that tho, one of the motivations behind Haskell was to have a unified basis for FP language research: at the time (late 80s) there were a dozen half-assed half-implemented languages in the space, and the main and most stable platform (Miranda) was proprietary software.

Alas I had fun with rust and made some small apps but as I got deeper I just got scared off by the learning curve. I'm kind of over the hump with c++, javascript and python and I just couldn't do it again. Maybe I will revisit in a few years if it actually slows down.

The language evolution speed has nothing whatsoever to do with the learning curve - there's like two major features a year, one of which is only for advanced users. The rolling release cycle just means the tiny incremental improvements arrive sooner. The learning curve has been in place since 1.0 and it only gets harder the more times you give up.

> language evolution speed has nothing whatsoever to do with the learning curve

I don't think that's true at all. Little by little, ergonomic changes make it in, that are actually quality of life improvements for the end user. Better error messages, smarter faster compiler, better idioms etc. These are all small, but they add up. For someone coming into the language cold, it can be the difference between a terrible experience and a decent one. No language was born perfect, so let's not pretend Rust is there, or even close.

Rust keeps making small usability improvements all the time, but the major improvements have already landed in the 2018 edition (smarter borrow checker, modules syntax, forgiving match patterns).

I think everyone in this thread can be in agreement:

* there are language and tooling improvements all the time, that make learning the language easier

* waiting to learn the language can make it easier due to the above

* if you learn Rust today the amount of things you need to learn going forward are few to none, the evolution of the language doesn't make your old knowledge useless

To be honest, and maybe I'm just an outlier... As someone who uses Haskell for most of my personal stuff (sadly, not for work), I really don't mind the complexity. Haskell98 is... nice, at best, but it lacks a lot for me to be truly pleasant. GHC Haskell with extensions however I find a pleasure to use, and I've always felt like the complexity was well placed, right where I wanted it or needed it. Rust is also nice, but for me personally the extra features are just making it better, not worse.

No worries, Rust has a looong way to go before they reach C++ levels. And the editions system even makes it feasible to pursue some ex-post simplification, though that requires a number of years to complete and cross-linking across editions must be preserved regardless.

I get the same feeling, Rust seems to be going down the path of "#[feature]" collection.

Rust feature flags aren't the same as Haskell feature flags. In Haskell you design your own programming language by selecting which extensions to enable. In Rust, feature flags only exist on the unstable nightly compiler and either get folded into the language proper or deprecated and dropped entirely. It's not a combinatorial explosion of interacting extensions, it's a single base for everyone that grows over time.

Really?

https://doc.rust-lang.org/cargo/reference/features-examples....

Oh, those features, you are talking about stuff that is included/excluded in libraries. Those have nothing to do with compiler extensions, they don't change the language.

Indeed, perhaps it was bad to name this (er) feature of Cargo with the word "features", since it invites conflation with the unstable #[feature] attribute of nightly Rust, but it has nothing to do with the language.

It shares that as well as horrendous compile times with Swift.

The irony being that while C++ is known for its compile times, the ecosystem relience on binary libraries makes it much more tolerable.

Microsoft is also shipping pre-compiled projections for Rust/WinRT.

Then we have examples like image, that compile rayon twice, with different versions, because we just love to watch third party crates being compiled.

However I should also add that compile times have improved greatly, my travel netbook can finally handle small Rust projects without killing the battery.

C++ binary libraries are a disaster. At least Rust is honest and tells you to just fall back to the C ABI for interop, and rebuild high-level abstractions around it in code that's built with the project - much like a 'header-only library' in C++.

I beg to differ, and until Rust community acknowledges they are relevant, there are industry domains where C++ will keep its dominance.

Take a look at the C++ community discussion around the feasibility of "ABI breaks". The C++ community is acknowledging the issues, and how they're getting in the way of continued "dominance" in many sectors.

I am fully aware of them, a vocal minority that only cares about FOSS libs.

Why do you think Apple has spent such a big effort making Swift work for shipping binary libraries?

Most devs would just stick with Objective-C, C and C++ otherwise.

> Microsoft is also shipping pre-compiled projections for Rust/WinRT.

Eh, sort of. They're basically shipping the Rust equivalent of C header files and import libraries. Whereas before they would parse metadata in a build script which would then generate the Rust code on the fly (which was hell for a number of reasons). And the installed Windows SDK was used for the import libraries.

Nope, I was speaking about this,

https://github.com/microsoft/windows-rs/tree/23ff38bbbf46fb5...

Apparently since last November they decided to replace it with another approach, so my information is outdated.

Reply on news.ycombinator.com