Why Rust is not a mature programming language
codecs.multimedia.cx
codecs.multimedia.cx
Traits are sets. “Impl” means a type, or any type described by the constraints, is in that set.
trait A {}
struct S;
impl A for S {} // S is in A
impl<T> A for Vec<T> where T: A {} // Vecs of A are also in A
That’s it.
Here's a random example: if you want a trait object with more than one trait, that only works today if one of the two is an auto trait: https://play.rust-lang.org/?version=stable&mode=debug&editio...
Let me save others 6s https://vlang.io
The Go language spec is really awesome, but the core language is so much simpler, so it was probably a lot easier to come up with. E.g. if the spec included every detail about escape analysis, pragmas, assembly, and cgo, it would be significantly longer and more difficult to maintain.
Rust brought a novel idea to the landscape of programming languages. The borrow checker is a great invention, and a great contribution to making applications more secure.
The language itself is way too complicated. The learning curve is steep. Improvements are made (e.g. lifetime elision) but they don't flatten the curve.
Macros are very hard to debug and understand. Which is even more of an issue when they are used by dependencies and something goes wrong. Error messages don't help. Macros are powerful but can hide too much magic. Plus, they are a language of their own.
"Why doesn't have this structure any methods?" "Oh, they are hidden in traits, provided by other crates, themselves leveraging other traits from other crates". "Still, it doesn't compile. Why?" "Oh, because of that generic that requires traits from yet another crate that are only available when some Cargo features has been enabled". Once again, this is difficult to follow and debug.
Zero-cost abstractions are great, but in Rust, they are overused. This makes the language less accessible, but also constantly causes dependency issues due to incompatible versions.
Rust is powerful but it is not optimized for developer happiness. Fighting the compiler brings anger, not joy. Sure, the compiler may have good reasons to complain. But as a developer, I'm happier and more productive when code immediately runs, even if that means having to fix bugs later. This is from a happiness view, not from a security perspective. But eventually, happiness turns into productivity.
It's too late to fix this. But now that Rust demonstrated that memory safety and speed were not incompatible, future system languages will feel obligated to provide the same properties, maybe by using mechanisms similar to Rust.
So, my prediction is that Rust will slowly be replaced by new safe languages that will be more accessible, more productive, more stable, and bring novel ideas of their own. But Rust will have made history no matter what.
I disagree that the improvements that have been made to the language ergonomics haven't changed the slope of Rust's learning curve, and I'm not entirely convinced that the language features are the totality of the problem when it comes to the difficulty of picking Rust up by experienced developers[2]. I also have the further belief that the language itself is not the only thing that affects the learning curve of a language. The available libraries, documentation, tooling, platform support, and the surrounding community all affect how easy to learn a language is, and as importantly, how likely people are to stick around.
> Macros are very hard to debug and understand. Which is even more of an issue when they are used by dependencies and something goes wrong.
I agree. In their current incantation. Long term Rust will have other ways of declaring and using macros. proc_macros is one post-1.0 change, but multiple people in the team have their targets aimed beyond that feature to provide compile-time code expansion that are easier to learn, understand, declare, use, and debug than macro_rules and proc_macros.
> Error messages don't help.
Things got much better in the 2018-2019 period for macros error messages in particular, but they are still a long way off from what I'd personally like to see. It is a hard problem that I am intimately familiar with. There's one particular change I would like to implement in the next year (tracking spans of macro_rule macro arguments through operations on the macro body, regardless of operations on them, like if you call foo!(bar) and the problem is with bar, the underline should point at bar and not to both foo!(bar) and in the definition of foo where bar was used), but improvements in this area are non-trivial and take time.
> Macros are powerful but can hide too much magic.
True, this is a problem not only for crates that rely heavily on macros but also for ones that rely heavily on complex trait bounds. I would personally wish crate writers to exercise some restraint on how complex their APIs are considering also how difficult to understand it is what went wrong when using them wrong, but I have no authority to force anyone to do things the way I'd like to and cope by 1) leading by example and 2) slowly but steadily improving diagnostics for patterns used in the wild.
> Plus, they are a language of their own.
They are. As alluded to but not explicitly said earlier, macro_rules where a placeholder for the 1.0 release. They are in every way an MVP, scheduled to be replaced at some point[1]. The thing is that they work "well enough™️" so there hasn't been a quick push to replace them, and we wouldn't want to rush a new way of doing something you can already do and get stuck with two subpar features. And proc_macros have their own host of issues to deal with, from wrapping your head around writing Rust code to generate Rust code, dealing with the AST, finding the right crates to use, learning how to use the available hooks to provide reasonable error messages, etc.
> "Why doesn't have this structure any methods?" "Oh, they are hidden in traits, provided by other crates, themselves leveraging other traits from other crates". "Still, it doesn't compile. Why?"
Agree, and that can be even exacerbated by the use of Derive and proc_macros, which can generate the impl for the "hidden" trait in a way that rust-analyzer and rg won't find them.
> "Oh, because of that generic that requires traits from yet another crate that are only available when some Cargo features has been enabled". Once again, this is difficult to follow and debug.
This is one of the mentioned "hard problems" that diagnostics could help with but nailing the experience down to make it "enjoyable" (or even bearable) require changes and coordination between rustc, cargo, and in some cases but likely not this one in particular, crate writers.
> Zero-cost abstractions are great, but in Rust, they are overused.
In parts of the ecosystem they absolutely are. I don't think there's anything about the language itself that encourage their overuse, I think it is more of the "new toy" effect: "Oh! Neat feature! And this lets me make this problem unrepresentable! NEAT! Let's use it! (Wait, what does it look like when someone misuses this? OH, NO!)".
> This makes the language less accessible, but also constantly causes dependency issues due to incompatible versions.
> Rust is powerful but it is not optimized for developer happiness.
Again, I disagree, but I can see your point of view. IMO Rust optimizes for debugability: the behavior of the code is always laid bare, which leads to verbosity (although it might be partially hidden behind Derives or macros, there is no "implicit" behavior). I find that this brings me joy when working on production-ready projects.
> Fighting the compiler brings anger, not joy.
It saddens me to hear that and would love to take any extra feedback in order to mitigate the anger it brings you.
> Sure, the compiler may have good reasons to complain. But as a developer, I'm happier and more productive when code immediately runs, even if that means having to fix bugs later.
How would you even find out about the bugs later? There's always the possibility of a buggy branching code path is never run, until it is, months after the fact.
> This is from a happiness view, not from a security perspective. But eventually, happiness turns into productivity.
I migh guess part of the issue is the old "make as many pots as possible in the allotted time" vs "take the allotted time and make a single great pot" experiment: more, faster iterations yields better quality in the same amount of time. This is a perfectly valid complaint about how fussy rustc can be, but the target is for rustc to be more akin to a pair programmer: "Hey! You forgot this trait bound here so that it matches what you're calling. Add it over there and that should do it."
> It's too late to fix this. But now that Rust demonstrated that memory safety and speed were not incompatible, future system languages will feel obligated to provide the same properties, maybe by using mechanisms similar to Rust.
And I've gotten to the part I actually wanted to answer to :)
What do you feel would get in the way from making Rust easier to use? I agree that there are ergonomic related things that could be done that won't because they would affect some of the use cases that Rust targets (like embedded, kernels or databases), like auto-cloning or auto-boxing which would make the language feel way higher level than it currently does, but I think there are lots of other things that could be done (thinking of the match pattern ergonomics work done some time back where you don't have to write & in patterns nearly as much now) that would make the experience of writing code nicer without sacrificing any of the project's stated goals.
> So, my prediction is that Rust will slowly be replaced by new safe languages that will be more accessible, more productive, more stable, and bring novel ideas of their own. But Rust will have made history no matter what.
If Rust's impact is nothing else but improving the larger ecosystem through the indirect impact of raising the bar for whatever comes after, then it will have been worth it.
[1]: https://github.com/rust-lang/rfcs/blob/master/text/1584-macr...
[2]: https://youtu.be/Z6X7Ada0ugE?t=705 Transcript of the relevant excerpt:
> I have the unsubstantiated theory that experienced developers have a harder time than less experienced developers when learning Rust. You need to forget a lot of constructs that work well enough in the languages you already know because they introduce things that go against the single owner enforcement that Rust has, whereas somebody with less experience will simultaneously accept restrictions as "just the way it is" and not seek out more performant constructs that can be much harder to understand or implement. Rust has a curse (it has many, but this one is critical): inefficient code is generally visible. Experienced developers hate to notice that their code is inefficient. They will recoil at seeing `Arc<RefCell<T>>`, but won't bat an eye at using Python. I know because I have the same instinct! This makes it much harder to learn Rust for experienced developers because they start with the "simple Rust code that will work but is slightly inefficient" and in an effort to improve it they land squarely in parts of the language they haven't yet developed a mental model for.
My take on the state of these issues:
> Rust does not have a formal language specification... I understand that adding new features is more important than documenting them but this is lame.
Most languages do not. It also really depends on what you mean by "formal."
It's not about being more important, it's that we value stability very strongly, and don't have the ability to document things with the guarantees we'd prefer. You might call it... not mature enough yet :)
There's been a bunch of movement here, I'm excited to see it continue to develop!
> Function/method calling convention. ... I’m told that newer versions of the compiler handle it just fine but the question still stands
The objection here doesn't have to do with calling conventions, actually, this is about "two-phase borrowing," described in a series of blog posts ending here http://smallcultfollowing.com/babysteps/blog/2017/03/01/nest...
I believe this will get even better with polonius https://nikomatsakis.github.io/rust-belt-rust-2019/
Regarding argument evaluation order, technically it is not yet documented https://github.com/rust-lang/reference/issues/248 but has been left-to-right for basically forever https://internals.rust-lang.org/t/rust-expression-order-of-e... and I actually thought that it was documented as such. I would expect this to shake out the exact same way as struct field destruction order, that is, something that's been one way for a long time and so we wouldn't change it even if maybe it's a good idea to.
> Traits... the problem is not me being stupid but rather the lack of formal description on how it’s done and why what I want is so hard. Then I’d probably at least be able to realize how I should change my code to work around the limitations.
Upcasting/downcasting is rarely used, and so has less love, generally, it's true.
> First of all, bootstrapping process is laughably bad.
So, Kostya acknowledges
> Of course it’s a huge waste of resources for rather theoretical problem but it may prove beneficial for compiler design itself.
Which is I think the way most feel about it. However, there is some desire to improve this, for other, related reasons. https://matklad.github.io/2020/09/12/rust-in-2021.html talks about some of them.
> Then there’s LLVM dependency.
While this stuff is all true, we wouldn't be where we are without it. Everything has upsides and downsides.
> And finally there’s a thing related to the previous problem. Rust has poor support for assembly.
He mentions asm; we're almost there! It took some time because it is not a simple problem. At the end of this, we'll have better support than C or C++ according to his metrics; these are not part of the language standard, so give his previous comments about maturity, I find this one a little weird, but it is what it is. :)
> There’s also one problem with Rust std library that I should mention too. It’s useless for interfacing OS.
Yes, the intention of the standard library is to be portable, so that's not really a goal.
> But the proper solution would be to support at least OS-specific syscall() in std crate
This may in fact be a good idea! I'm not sure how much use it would actually get.
That is: CRust (a hypothetical C-based Rust compiler) can be made to target XX99 hardware, but in order to run CRust on that hardware you also have to make the C compiler support it. Achieving self-hosting, especially for a language that's targeting low-level capabilities like Rust is, is rather important.
Writing a compiler on its own language has a bunch of benefits.
Early on, it let's both evolve in tandem, even before you know what the language itself might be. Having real world experience in a complex enough codebase for the language will inform some design decisions. Things that are hard to do might get ergonomic work poured into, sharp edges tapered. Things that are too hard to implement or that might cause exponential evaluation might be redesigned to allow for a linear algorithm.
Later, having the compiler written in its language is beneficial for contributors: people that use the language can jump in and help with the development of the compiler. This has the caveat if any large codebase, but it certainly was my case. I would go as far as saying that I really learned Rust through my rustc contributions. (BTW, doing that has the nice benefit of fixing your mental models if the language to actually match reality, instead of some approximation based on the documentation and observed behavior.)
Finally, setting the debugging scaffolding in place will be made a priority in order to debug the compiler, so even early users of the language will benefit of some tooling in that area, however crude it might be at the start.
> Most languages do not (have a formal specification)
Most languages also aren't trying to replace languages that do. C and C++ are both languages that Rust, afaik rather officially, aims to replace in some areas. It can be a great language like so many others, but if it wants to replace these old giants, it needs to have a proper spec, maybe in the form of an ISO standard. Of course that will come when its time, but thats a good indicator of when a language can compete as an answer to the question of "what tech will we use for our next big, important and highly specialized project?".
This is the pertect indicator of it not being as mature as what it aims to replace.
And also, about the invocation of "formal" there...
Rust doesn't have that problem; there aren't yet any alternative Rust compilers that have any other purpose than to run as a batch-scheduled crate-at-a-time compile step at the command-line.
In such a case, where there's only one real stake-holder, "standardization" becomes less about declaring what should happen; and more about specifying what does happen, in exacting detail, such that someone could build an alternative conforming implementation from the spec without looking at the source of your reference implementation.
I don't feel like the existence of such a descriptive specification would have "strangled Rust" at any point. At most, this would have roughly doubled the work of any fix: writing the code, and then writing the change in the spec. But it wouldn't have actually been double the overall labor overhead, since the increased clarity-of-purpose of modifying the spec to declare a change in intention, would likely have mooted a lot of requesting-clarification and debating at code-review time.
But besides, software-engineering as a discipline now has tools like Behavior-Driven Development to minimize the costs of maintaining a parallel descriptive spec for a project. BDD tests are just regular tests that embed a lot of descriptive strings in them—those strings being words you are already mostly thinking at the time of writing the test. So they're only a little more costly than writing ordinary tests (which the Rust compiler already has), yet can also be compiled out into a descriptive spec. (And then you can diff the generated spec, between versions, and turn that diff into the spec errata for the "minor specification addendum" of that minor release.)
Sort of, ISO has some rules that are antithetical to Rust's ethos, like requiring that conversations not be recorded. Rust's development chooses when to be public and when to be private where it makes sense.
I don't actually know if the "meet in person" aspect is a formal ISO rule or a peculiarity of the C and C++ committees, but that would be another vast difference that matters a lot. Especially at this historical moment.
> Rust doesn't have that problem;
We do have this problem, it's just not driven by compiler authors, but by the relevant stakeholders directly. The language team and the compiler team, while sharing some people, are separate.
> less about declaring what should happen; and more about specifying what does happen
This is not how the process plays out in Rust, though you're right that it could, if the compiler team wanted to act in bad faith.
"Does not have a formalized spec"
See [1] page 35.
[1] https://www.slideshare.net/AdaCore/securing-the-future-of-sa...
For an example of how this can play out, the parent is referring to things like https://marcan.st/2017/12/debugging-an-evil-go-runtime-bug/
That will work for most OS's out there, and windows
Ditto libsystem on macos.
Freebsd also provides a stable syscall interface, like linux.
I can't find anything that guarantees FreeBSD's system call ABI. Do you have a source? It would be very interesting if, for example, FreeBSD 10 applications that use syscalls can run on modern FreeBSD without a compatibility layer. However, if the FreeBSD project does not provide guarantees it would be folly to rely on this behaviour in the future. If it does provide such a guarantee then I stand corrected.
EDIT: After some more research it seems the FreeBSD kernel needs the `COMPAT_FREEBSD10` option enabled for my hypothetical example to work. The default options for amd64 include compatibility options back to FreeBSD 4. Defaults for other platforms seems to differ (perhaps depending on when the platform was first supported).
I can't find good documentation on if these provide full compatibility or if any `COMPAT_` options could be dropped in future versions.
I disagree.
System calls on linux comprise the interface provided to applications to talk to the kernel.
On windows, ntdll serves the same function--it is itself a 'very thin wrapper around calling the kernel directly by number'. (Especially important since a libc may be hard to come by on windows.)
(Libsystem it seems I was mistaken about; it looks like that's just a bundle of libc, libm, libpthread, etc. Though possibly libsystem_kernel is nearer the mark? Difficult to find information on the subject, and I don't have a mac.)
>> Regarding argument evaluation order, technically it is not yet documented https://github.com/rust-lang/reference/issues/248 but has been left-to-right for basically forever https://internals.rust-lang.org/t/rust-expression-order-of-e... and I actually thought that it was documented as such.
I am aware the word "undefined" references the order of evaluation. However, I just want to clear possible confusion on the matter. Code that depends on evaluation order doesn't produce undefined behaviour. It produces unspecified behaviour. (This is not a directly reply to Steve, who am I sure knows more about this than I ever will).
> Aw, dang it. This post is going to get out and cause this server to crash again, isn’t it?
> Rust does not have a formal language specification […] A proper mature language (with 1.0 in its version) should have a formal specification
Python doesn’t have a formal specification. It doesn’t even have a semi-formal specification. I would consider Python a proper mature language. (… even if it doesn’t have 1.0 in its version!)
Some recently-published work in the general direction of formally specifying Rust behaviour: https://people.mpi-sws.org/~jung/thesis.html
> In Rust it’s undefined because there’s no format specification to tell you even that much.
That seems to me an very unreasonable definition of “undefined”. By that token, nothing in Rust is defined, which renders the term “undefined” devoid of meaning and completely useless. Rather, the implementation defines the specification, which, yes, is weaker than a formal specification, but not that much weaker. (Rust’s stability guarantees do allow it to retcon, but only if safety is on the line, or sometimes if you can convince people that affected code probably doesn’t exist, or was wrong anyway. Changing argument order evaluation in this sort of case would break plenty of real-life code, and so will never happen.)
> […] traits […] the problem is not me being stupid but rather the lack of formal description on how it’s done and why what I want is so hard.
I disagree that the lack of a formal description of how it’s done is the problem, though such a description would certainly help. I feel that the real problem here is that you’re trying to treat traits as something that they’re not, trying to do things that you’re used to doing in dynamic/GC languages that just don’t map to Rust because of it being a thin abstraction over what can be fastest. Upcasting and downcasting trait objects is really rare, because other constraints of the language make them just not very useful operations. The main reason you want those operations in other languages is inheritance, not interfaces, and Rust doesn’t do inheritance, and traits aren’t inheritance.
> rustc problems
The bootstrapping process is not laughably bad, the developers have merely prioritised their own ergonomics rather than pandering to making proving a theoretical problem false easier. Still more, they’ve prioritised actually using new functionality, which helps make sure that it both works properly, and is the right new functionality. So I find two or three good reasons for doing it the way they have, and one weak reason not to do it that way.
The LLVM dependency: I feel that this complaint is more about there being only a single Rust backend, rather than about the LLVM dependency. This point is then debatable. Yeah, Rust would be more mature if it had more backends (or better, more full implementations), but does having only a single implementation make it immature? Eh. Plenty of mature languages functionally have but a single implementation.
———
I definitely consider Rust not yet mature, I just don’t think that most of the reasons in this article are legitimate. I would focus in no small part on documentation, just not requiring formality of specification or documentation before I’d consider those parts mature. But note that I said that I’d consider Rust not yet mature, rather than the Rust programming language. Because in most regards I do actually consider the programming language itself mature, and a lot of its tooling (more than popular languages thrice its age), though there are definitely still gaps in both. But the ecosystem around the programming language, that’s the area where it’s not yet mature. Rust is much more complex than many languages (e.g. Go), so it takes a lot longer to fill in the spaces in library availability, and there are many very significant gaps that haven’t been filled yet. GUI, for example, is very immature space. Audio is… eh, it’s getting somewhere, but I wouldn’t call it mature. Web is approaching maturity for some sorts of tasks, but is not generally mature.
Ironically, I would disagree here; the language developers have not prioritized their own ergonomics, which is why building rustc is a big pain.
I do agree that making it easier to bootstrap in that way would not necessarily help them... there is a sense in which you're right, I just found the framing kind of funny, since I'd argue the same point but for completely opposite reasons.
This doesn't have to be all-or-nothing, though, if you check out matklad's post I linked, he points to a version where the standard library still gets to use unstable features, but the compiler does not, which is an interesting hybrid that may retain the best of both.
Therefore the article as a whole cannot be taken seriously, and reflecting that I stopped reading it immediately after.
In this day and age, you need to decide quickly, or you'll be wasting a lot of your time.
Now I don't know about you, but my time is valuable.
[1]: https://godbolt.org/z/er3zT6
[2]: https://godbolt.org/z/zKPao4
[3]: https://blog.rust-lang.org/2019/07/04/Rust-1.36.0.html#nll-f...
[4]: https://godbolt.org/z/WKczbv
[5]: https://blog.rust-lang.org/2018/12/06/Rust-1.31-and-rust-201...