My “grand vision” for Rust
blog.yoshuawuyts.com133 points by todsacerdoti 4 days ago
133 points by todsacerdoti 4 days ago
I write production Rust code that becomes critical infra for our customers. I got tired of nil checks in Go and became a squeaky wheel in incident retros, where I finally got the chance to rewrite parts of our system in Rust during a refactor.
I admit the skill issue on my part, but I genuinely struggled to follow the concepts in this article. Working alongside peers who push Rust's bleeding edge, I dread reviewing their code and especially inheriting "legacy" implementations. It's like having a conversation with someone who expresses simple thoughts with ornate vocabulary. Reasoning about code written this way makes me experience profound fatigue and possess an overwhelming desire to return to my domicile; Or simply put, I get tired and want to go home.
Rust's safety guardrails are valuable until the language becomes so complex that reading and reasoning about _business_ logic gets harder, not easier. It reminds me of the kid in "A Christmas Story" bundled so heavily in winter gear he cant put his arms down[0]. At some point, over-engineered safety becomes its own kind of risk even though it is technically safer in some regards. Sometimes you need to just implement a dang state machine and stop throwing complexity at poorly thought-through solutions. End old-man rant.
I've been on both sides of the fence here - I've bounced between two camps:
1. Go with a better type system. A compiled language, that has sum types, no-nil, and generics.
2. A widely used, production, systems language that implements PL-theory up until the year ~2000. (Effects, as described in this article, was a research topic in 1999).
I started with (1), but as I started to get more and more exposed to (2), you start looking back on times when you fought with the type system and how some of these PL-nerds have a point. I think my first foray into Higher-Kinded Types was trying to rewrite a dynamic python dispatch system into Rust while keeping types at compile time.
The problem is, many of these PL-nerd concepts are rare and kind of hard to grok at first, and I can easily see them scaring people off from the language. However I think once you understand how they work, and the PL-nerds dumb down the language, I think most people will come around to them. Concepts like "sum types" and "monads", are IMO easy to understand concepts with dumb names, and even more complex standard definitions.
> Reasoning about code written this way makes me experience profound fatigue and possess an overwhelming desire to return to my domicile;
I didn't understand that you were making fun of verbosity until the word 'domicile'. I must be one of those insufferable people who expresses simple thoughts with ornate vocabulary...
The article was comprehensible to me, and the additional function colorings sound like exciting constraints I can impose to prevent my future self from making mistakes rather than heavy winter gear. I guess I'm closer to the target audience?
I really think that golang makes it easy to read code, rust makes it easy to write code. If Golang had sum types it would be a much nicer language to write complex applications with
I find Go code mind numbing to read. There's just _so much of it_ that the parts of the code that should jump out at me for requiring greater attention get lost in the noise. Interfaces also make reading Go more difficult than it could be without LSP - there's no `impl Xyz for` to grep for.
Golang makes easy-to-skim code, with all the `if err != nil` after every function call.
Rust requires actual reading, like Typescript, only more detailed.
It's hard to see features through the programming language theory jargon, but solid theoretical foundations have worked well for Rust so far.
Jargon terms like "sum types" or "affine types" may seem complicated, but when you see it's actually "enums with data fields", it makes so much sense, and prevents plenty of state-related bugs.
Proposed "effects" mean that when you're writing an iterator or a stream, and need to handle error or await somewhere in the chain, you won't suddenly have a puzzle how to replace all of the functions in the entire chain and your call stack with their async or fallible equivalents.
"linear types" means that Rust will be able to have more control over destruction and lifetime of objects beyond sync call stack, so the tokio::spawn() (the "Rust async sucks" function) won't have to be complaining endlessly about lifetimes whenever you use a local variable.
I can't vouch for the specifics of the proposed features (they have tricky to design details), but it's not simply Rust getting more complex, but rather Rust trying to solve and simplify more problems, with robust and generalizable language features, rather than ad-hoc special cases. When it works it makes the language more uniform overall and gives a lot of bang for the buck in terms of complexity vs problems solved.
This may be too much advanced type theory for a useful language.
You can go all the way to formal verification. This is not enough for that. Or you can stop at the point all memory error holes have been plugged. That's more useful.
You can go way overboard with templates/macros/traits/generics. Remember C++ and Boost. I understand that Boost is now deprecated.
I should work some more on my solution to the back-reference problem in Rust. The general idea is that Rc/Weak/upgrade/downgrade provide enough expressive power for back references, but the ergonomics are awful. That could be fixed, and some of the checking moved to compile time for the single owner/multiple users case.
> You can go way overboard with templates/macros/traits/generics.
You can go overboard on any language concept imaginable, but conflating all these mechanisms makes it sound like you haven't interacted much with non-C++ languages—particularly since rust doesn't have templates or anything like templates, traits are an entirely unrelated composition mechanism, and macros are entirely unrelated to the type discussion in the article.
This isn't really "advanced type theory" so much as picking up programming language developments from the 90s. I suppose it's "advanced" in the sense that it's a proper type system and not a glorified macro ala templating, but how is that a bad thing?
Agreed. Generics are in most modern typed languages, and traits are essentially interfaces. Maybe templates means C++ templates, which are essentially generics?
My guess is they meant metaprogramming in general (templates/generics, macros), but traits are not quite like the others.
In C++, concepts are essentially generics where templates are more like weird macros.
Thanks for posting this! As a long-time Rust user (and contributor, in the good old days), the thing that has always fascinated me about Rust is the healthy balance it strikes between academic brilliance and industry pragmatism. Radical changes like the ones suggested by the OP risk damaging that balance IMO. I'd rather put up with some language quirks and see Rust achieve "boring technology" status...
But who knows, maybe the "academic brilliance" from the article is more pragmatic than I give it credit for. I sure hope for it if these changes ever go through.
The lack of use cases in that document is a concern. They're all "nice to have" features, but is the payoff there for real work? The "effects" section mentions properties useful for a proof system. But it's not part of a proof system. If it were, most of those could be found automatically by static analysis, without bothering the programmer. (I was doing that decades ago in a very early proof of correctness system.) Getting programmers to annotate items is tough. Just getting C++ programmers to use "const" enough is hard.
"View types" are interesting. But how much of that generality is really needed? We already have it for arrays, with "split_at_mut" and its friends. That's a library function which uses "unsafe" but exports a safe interface. The compiler will already let you pass two different fields of a struct as "&mut". That covers the two most pressing cases.
"View types" addresses one of the top pain points of the borrow checker, which is disjoint borrows on a single value. The compiler can understand values where only some fields are borrowed, so e.g. you can mutably borrow the currently-unborrowed fields, but there's no way to write these partial borrows and so you cannot write functions that either take or return these borrows. This often means having to restructure code to split up one struct into several child structs just to be able to borrow those substructs independently.
I would also have liked to see some motivational examples, but I think the most interesting upside of an effect system is composability.
Rust is actually really unique among imperative languages in its general composability - things just compose really well across most language features.
The big missing pieces for composability are higher-kinded types (where you could be generic over Option, Result, etc.), and effects (where you could be generic over async-fn, const-fn, hypothetical nopanic-fn, etc.)
The former becomes obvious with the amount of interface duplication between types like Option and Result. The latter becomes obvious with the number of variants of certain functions that essentially do the same thing but with a different color.
Some uses of higher-kinded types (though not all of them) can be addressed by leveraging Generic Associated Types (GAT).
Part of the problems is that the "things just compose really well" point becomes gradually less and less applicable as you involve the lower-level features Rust must be concerned with. Abstractions start to become very leaky and it's not clear how best to patch things up without a large increase in complexity. A lot of foundational PL research is needed to address this sensibly.
How are these suggestions not pragmatic? You don't have to use them, but if you need them they are there. From a security point of view I can see many of these being incredibly useful.
If it slows down Rust development it's not pragmatic. And if it creates a cultural schism between full commitment and pragmatic approaches, it's also trouble. Remember Scala?
Rust is nowhere near the complexity of Scala wrt. seemingly arbitrary high-level features. There's a low-level, systems programming featureset that involves quite a bit of complexity but that's also less arbitrary when comparing across similarly low-level languages.
> This may be too much advanced type theory for a useful language.
I think a lot of things taken for granted these days were considered "too complicated" some time ago: think of how widespread pattern matching, closures, generics, or functional idioms in imperative languages are, and compare to e.g. Java 1.0.
My feeling is that the "acceptable level of complexity" for programming languages goes up over time, so probably stuff like effect types will be almost everywhere in another 10 years.
Counterpoint: if any language could thrive in that valley of despair between pragmatic and theoretical excellence you're referring to, it would be Rust. Because so much of the cost is already paid for once you have satisfied the borrow checker. At least that's what I'd imagine, I could certainly be wrong.
I'm not 100% convinced that "plugging memory error holes" was right at the compiler level.
Currently building out clr, which uses a heuristic (not formal verification) method for checking soundness of zig code, using ~"refinement types". In principle one could build a more formal version of what I'm doing.
> This may be too much advanced type theory for a useful language.
Maybe but:
- Move fixes Pin
- Linear types, prevent memory leaks
- potentially effects simplify so many things
Each of these functionalities unlock capabilities people have complained about Rust. Namely async, gen blocks, memory leaks.
Boost is as actual as ever.
Also the way nowadays is with constexpr, templates, and around the corner, static reflection.
> I understand that Boost is now deprecated.
Huh?? Boost is used basically everywhere.
Definitely not. Boost is specifically prohibited in many companies. I haven’t run into Boost in a source tree in over a decade.
There are many reasons for this. Boost has uneven quality. Many of the best bits end up in the C++ standard. New versions sometimes introduce breaking changes. Recent versions of C++ added core language features that make many Boost library features trivial and clean to implement yourself. Boost can make builds much less pleasant. Boost comes with a lot of baggage.
Boost was a solution for when template metaprogramming in C++ was an arcane unmaintainable mess. Since then, C++ has intentionally made massive strides toward supporting template metaprogramming as a core feature that is qualitatively cleaner and more maintainable. You don’t really need a library like Boost to abstract those capabilities for the sake of your sanity.
If you are using C++20 or later, there isn’t much of a justification for using Boost these days.
Not at all. Most of the good parts of boost are now part of the standard library, and the rest of boost that is of high quality have stand-alone implementations, like ASIO, pybind11 (which was heavily influenced by Boost.python), etc...
A lot of the new stuff that gets added into boost these days is basically junk that people contribute because they want some kind of resume padding but that very few people actually use. Often times people just dump their library into boost and then never bother to maintain it thereafter.
Reposting my comment from Reddit,
I had some Scala 3 feelings when reading the vision, I hope Rust doesn't gets too pushy with type systems ideas.
That is how we end with other ecosystems doubling down in automatic memory management with a good enough ownership model for low level coding, e.g. Swift 6, OxCaml, Chapel, D, Linear Haskel, OCaml effects,...
Where the goal is that those features are to be used by experts, and everyone else stays on the confort zone.
> I had some Scala 3 feelings when reading the vision, I hope Rust doesn't gets too pushy with type systems ideas.
I don't know if it is true or not, but my feeling is that Scala brought a lot of new ideas. But as I read somewhere, "Scala was written by compiler people, to write compilers", and I can understand that feeling.
Kotlin came after Scala (I think?) and seems to have gotten a lot of inspiration from Scala. But somehow Kotlin managed to stay "not too complex", unlike Scala.
All that to say, Rust has been innovating in the zero-cost abstraction memory safe field. If it went the way of Scala, I wonder if another language could be "the Kotlin of Rust"? Or is that Zig already? (I have no idea about Zig)
> But somehow Kotlin managed to stay "not too complex", unlike Scala.
It's not really true anymore, Kotlin has slowly absorbed most of the same features and ideas even though they're sometimes pretty half-baked, and it's even less principled than the current Scala ecosystem. JetBrains also wants to make Kotlin target every platform under the sun.
At this point, the only notable difference are HKTs and Scala's metaprogramming abilities. Kotlin stuck to a compiler plugin exposing a standard interface (kotlinx.serialization) for compile-time codegen. Scala can do things like deriving an HTTP client from an OpenAPI specification on the fly, by the LSP backend.
> JetBrains also wants to make Kotlin target every platform under the sun
So did Scala long before. It's just that Kotlin got a lot more traction for different reasons.
Not to the same extent. Scala.JS and Kotlin.JS are somewhat comparable, other targets not so much. There was no serious attempt at making Scala target mobile devices, even during the window of opportunity with Scala on Android.
> I wonder if another language could be "the Kotlin of Rust"?
Some people would say that Swift is that language since it's potentially memory safe like Rust and is described as friendlier to novices. There's some room for disagreement wrt. the latter point.
Well many languages are memory safe. Java has been memory safe forever. Rust is the one that is memory-safe with zero-cost abstractions, right?
I doubt those languages would have the same level of traction as Rust, especially now that Rust has already gotten said traction over the past decade with even the Linux kernel using them. It's more likely that Rust will be written as today and then these extra features are added for more type safety in certain functions as like I said in another comment I doubt people are going to write type contracts for every single function (maybe LLMs will, but that's an orthogonal discussion).
Apparently you missed Swift.
Linux kernel adoption of Rust hasn't been a smooth ride, exactly because of its type system among C folks.
It is only happening because the likes of Google and Microsoft want to see it through.
Swift is not really for systems level programming like Rust and interestingly some projects like Ladybird have moved away from Swift towards Rust.
My understanding is that Scala 3 came with many large breaking changes that made adoption difficult. I at least hadn't heard users complain that new features weren't desired.
Doubt Rust will ever get to implicit hell of Scala 2.
If for anything, Rust isn't married to C as Scala is to Java.
Really excited for the possibilities here.
People undoubtedly thought going for Affine types was too much, and even simple things like null safety or enums-with-values and the prevalence of Result saw debate with minimalists voicing concerns.
A world where you could write a Rust program that is memory leak free with Affine types is one I want to live in. Haskell can do it now, but its just not easy and Rust has beat out Haskell with its mix of ML-strength types and practicality.
IMO these changes maintain Rusts winning mix of academia and practicality. Heres a proof point — dependent types weren't mentioned :)
From the historical sources I could find online, it appears that Rust's borrow system was independently invented, or at least they don't mention linear logic or anything substructural. This is kind of interesting to me, especially given the reactions in this thread, and ties into the general difficulty of PL research to find acceptance among practitioners, especially when presented by researchers (which I think is regretful, I like the ideas in the article!). Perhaps we really should stick to terminology like "function colors" to make effect systems more popular (or not, because the color framing makes it sound bad to have different colors in a program, IIRC).
It's the jargon, I think. PL research is in an awkward position, where the jargon is not shared with the much wider community of people using programming languages daily. From the other side, it looks like there is a small body of theoreticians using impenetrable language for discussing topics I'm supposed to be familiar with, because they are a core part of my day job. It's much easier to accept jargon, when it's used in a clearly separate field.
Some of the terminology is just unfortunate. For example, I have an intuitive understanding of what a type means. The meaning used in PL theory is somehow wider, but I don't really understand how.
And then there is my pet peeve: side effect. Those should be effects instead, because they largely define the observable behavior of the program. Computation, on the other hand, is a side effect, to the extent it doesn't affect the observable behavior.
But then PL theory is using "effect" for something completely different. I don't know what exactly, but clearly not something I would consider an effect.
Man who uses arithmetic upset at research mathematicians for using words like R-module when they clearly do not mean a module in C++
More at 11
I don't remember where I read it, but I think Rust cited Cyclone as an influence, a variation of C with "region-based" memory management - more or less the literature name for "lifetimes". I think Rust may be the first to use it directly for stack variables, however.
No-one ever has the "Grand Vision" to cut something down to it's essential 25% and delete the rest.
By all means, write some Go instead—you might want to order a new `if err != nil` button to stave off the RSI. The rest of us are trying to make things better.
Finally seeing more movement on effects or what started as keyword generics, there was a big blog post a few years ago but not much public facing news although of course they've been working on it as Yoshua says in the post.
I truly do wish we get closer to Ada and even Lean in terms of safety, would be great to see all these theoretical type system features become reality. I use the `anodized` crate right now for refinement type features, and who knows, maybe we get full fledged dependent types too as there aren't many production languages with them and certainly not popular languages.
When I wrote my very first Rust code, I was trying to write to a socket. I got stuck on this task with misleading error messages for the longest time. I finally realized I had not made the socket object mutable. I’m used to Posix where you have an integer file descriptor and I don’t tend to think of socket write as a mutable operation. At least it doesn’t mutate state that my app manages. Perhaps something in the kernel gets mutated. I believe the socket interface may have been intended to support queuing which is perhaps why it needed to be mutable. I might have needed a lower level api. I just mention this because I think it’s interesting as to how it should be typed when mutation is external to the app. I didn’t follow through on using Rust and this was long ago so I’m sure some details are wrong.
I also found it really unintuitive what needs to be made mutable. And still do to some degree.
When I first had learned that rust had this concept of “opt-in” mutability, I thought that it must then be an accepted pattern that we make as little as possible be mutable in an attempt to help us better reason about the state of our program. I had come to rust after learning some clojure so I was like “ahh! Immutable by default! So everything’s a value!”
But in reality it feels like rust code is not “designed” around immutability but instead around appeasing the borrow checker - which is actually pretty easy to reason about once you get the hang of the language. But it’s a ton to learn up front
> got stuck on this task with misleading error messages for the longest time.
Could you elaborate on that? We consider misleading error messages to be bugs and would like to know more on case we could fix them.
I couldn’t disagree more. Most of my company’s backend code is written in Scala, and most of our engineers dislike it because the language is difficult to understand, has way too many features, and has many ways to solve the same problem. I don’t want Rust to continue down this path, and I already worry with some of the syntactic sugar and type system additions being discussed that it already has.
A language’s type system doesn’t need to model every possible type of guarantee. It just needs to provide a type safe way to do 95% of things and force its users to conform to use the constructs it provides. Otherwise it becomes a buggy hodge podge of features that interact in poor and unpredictable ways. This is already the case in Scala; we’ve discovered almost 20 bugs in the compiler in the past year.
There is a middle ground. People seem to use Haskell and OCaml just fine and both are as expressive, so maybe it is just Scala having shoved in too many things. Based on what the article shows, it doesn't seem like they're making ten different ways to do the same thing but rather one way to (optionally) get more type safety out. I doubt everyone will be writing dependent type contracts for every single function, it's more for certain pieces of the codebase.
Ah yes Haskell, the reasonable centrist position in language design.
Compared to the other languages listed in this thread, it definitely is, and it has production systems unlike them.
I'm terrified by the notion of try fns. Are we getting exceptions (and therefore losing one of rust's greatest features)?
Isn't his point exactly that we don't want to have too many function colors and instead want a generic way of declaring side effects so people can do what they want (be it try fns, IO, async, etc..., no panicking)?
This sounds insane at this point. The language already has too many features. Would be cool if all these people with amazing visions could move it elsewhere.
Rust is fast tracking being as bad as c++ in terms of just garbage in it.
IMO the worst thing about c++ isn't that it is unsafe but it is extemely difficult to learn to a satisfying degree.
This is already kind of feels true for Rust and it will be surely true if people just keep shoving their amazing ideas into it.
IMO even async/tokio/error-handling aren't that well though out in rust. So much for keeping things out of the language.
Maybe Rust just wasn't what I wanted and I am salty about it but it feels a bit annoying when I see posts like this and considering where Rust is now after many years of shoving stuff into it
> The language already has too many features.
That's actually the point. Many of these additions can be phrased as unifying existing features and allowing them to be used in previously unusable ways and contexts. There's basically no real increase in user-perceived complexity. The Rust editions system is a key enabler of this, and C++ has nothing comparable.
It has clang tidy, -std=lang-version, and preprocessor that is version aware.
Rust editions don't cover all use cases that one can think of regarding language evolution, and requires full access to source code.
> requires full access to source code
What do you mean? Editions don't require full access to source code. Rust in general relies heavily on having access to source code, but that has nothing to do with how editions work
You can write a binary library that exposes a C ABI using Rust (which is indistinguishable from an ordinary C/C++ library) and then provide source for a Rust wrapper crate that provides a "safe" interface to it, much like a C header file.
Yes they do, when mixing crates from various editions and how changes interact together.
> when mixing crates from various editions and how changes interact together.
Could you elaborate more on this? It's not obvious to me right now why (for example) Crate A using the 2024 edition and Crate B using the 2015 edition would require both full access to both crates' source beyond the standard lack of a stable ABI.
Because in order to have standard library breaking changes across editions, if those types are exposed in the crate public types, or change their semantics across editions, the compiler has to be able to translate between them when generating code.
See the Rust documentation on what editions are allowed to change, and the advanced migration guide on examples regarding manual code migration.
Not so much what has happened thus far, rather the limitations imposed in what is possible to actually break across editions.
Or put another way, a hypothetical feature that you made up in your head is the thing that requires source access. Editions do not let you change the semantics of types.
To be fair, Rust tooling does tend toward build-from-source. But this is for completely different reasons than the edition system: if you had a way to build a crate and then feed the binary into builds by future compilers, it would require zero additional work to link it into a crate using a different edition.
Exactly, hence why people should stop talking about editions as if they sort out all Rust evolution problems, in your own words it doesn't allow changing type semantics