Without something like that, I think it just would have been impossible for Rust to gain enough momentum, and also attract the sort of people that made its culture what it is.

Otherwise, IMO Rust would have ended up just like D, a language that few people have ever used, but most people who have heard of it will say "apparently it's a better safer C++, but I'm not going to switch because I can technically do all that stuff in C++"

When my function gets an exclusive reference to an object, I know for sure that it won't be touched by the caller while I use it, but I can still mutate it freely. I never need to make deep copies of inputs defensively just in case the caller tries to keep a reference to somewhere in the object they've passed to my function.

And conversely, as a user of libraries, I can look at an API of any function and know whether it will only temporarily look at its arguments (and I can then modify or destroy them without consequences), or whether it keeps them, or whether they're shared between the caller and the callee.

All of this is especially important in multi-threaded code where a function holding on to a reference for too long, or mutating something unexpectedly, can cause painful-to-debug bugs. Once you know the limitations of the borrow checker, and how to work with or around them, it's not that hard. Dealing with a picky compiler is IMHO still preferable to dealing with mysterious bugs from unexpectedly-mutated state.

In a way, borrow checker also makes interfaces simpler. The rules may be restrictive, but the same rules apply to everything everywhere. I can learn them once, and then know what to expect from every API using references. There are no exceptions in libraries that try to be clever. There are no exceptions for single-threaded programs. There are no exceptions for DLLs. There are no exceptions for programs built with -fpointers-go-sideways. It may be tricky like a game of chess, but I only need to consider the rules of the game, and not odd stuff like whether my opponent glued pieces to the chessboard.

I've been writing C++ for almost 30 years, and a few years of Rust. I sometimes struggle with the Rust borrow checker, and it's almost always my fault. I keep trying to write C++ in Rust, because I'm thinking in C++ instead of Rust.

The lesson is always the same. If you want to use language X, you must learn to write X, instead of writing language Y in X.

Using indexes (or node ids or opaque handles) in graph/tree implementations is a good idea both in C++ and in Rust. It makes serialization easier and faster. It allows you to use data structures where you can't have a pointer to a node. And it can also save memory, as pointers and separate memory allocations take a lot of space when you have billions of them. Like when working with human genomes.

Probably I just haven't been writing very "advanced" rust programs in the sense of doing complicated things that require advanced usages of lifetimes and references. But having written rust professionally for 3 years now, I haven't encountered this once. Just putting this out there as another data point.

Of course, partial borrows would make things nicer. So would polonius (which I believe is supposed to resolve the "famous" issue the post mentions, and maybe allow self-referential structs a long way down the road). But it's very rare that I encounter a situation where I actually need these. (example: a much more common need for me is more powerful consteval.)

Before writing Rust professionally, I wrote OCaml professionally. To people who wish for "rust, but with a garbage collector", I suggest you use OCaml! The languages are extremely similar.

There are open ideas for how to handle “view types” that express that you’re only borrowing specific fields of a struct, including Self, but they’re an ergonomic improvement, not a semantic power improvement.

    struct Id(u32);

    fn main() {
        let id = Id(5);
        let mut v = vec![id];
        println!("{}", id.0);
    }

He does point out two significant problems in Rust. When you need to change a program, re-doing the ownership plumbing can be quite time-consuming. Losing a few days on that is a routine Rust experience. Rust forces you to pay for your technical debt up front in that area.

The other big problem is back references. Rust still lacks a good solution in that area. So often, you want A to own B, and B to be able to reference A. Rust will not allow that directly. There are three workarounds commonly used.

- Put all the items in an array and refer to them by index. Then write run-time code to manage all that. The Bevy game engine is an example of a large Rust system which does this. The trouble is that you've re-created dangling pointers, in the form of indices kept around after they are invalid. Now you have most of the problems of raw pointers. They will at least be an index to some structure of the right type, but that's all the guarantee you get. I've found bugs in that approach in Rust crates.

- Unsafe code with raw pointers. That seldom ends well. Crates which do that are almost the only time I've had to use a debugger on Rust code.

- Rc/RefCell/run-time ".borrow()". This moves all the checking to run time. It's safe, but you panic at run time if two things borrow the same item.

This is a fundamental problem in Rust. I've mentioned this before. What's needed to fix this is an analyzer that checks the scope of explicit .borrow() and .borrow_mut() calls, and determines that all scopes for the same object are disjoint. This is not too hard conceptually if all the .borrow() calls produce locally scoped results. It does mean a full call chain analysis. It's a lot like static detection of deadlock, which is a known area of research [1] but something not seen in production yet.

I've discussed this with some of the Rust developers. The problem is generics. When you call a generic, the calling code has no idea what code the generic is going to generate. You don't know what it's going to borrow. You'd have to do this static analysis after generic expansion. Rust avoids that; generics either compile for all cases, or not at all. Such restricted generic expansion avoids the huge compile error messages from hell associated with C++ template instantiation fails. Post template expansion static analysis is thus considered undesirable.

Fixing that could be done with annotation, along the lines of "this function might borrow 'foo'". That rapidly gets clunky. People hate doing transitive closure by hand. Remember Java checked exceptions.

This is a good PhD topic for somebody in programming language theory. It's a well-known hard problem for which a solution would be useful. There's no easy general fix.

[1] https://dl.acm.org/doi/10.1145/3540250.3549110

I don't think that was ever the intent behind the borrow checker but it is definitely an outcome.

So yes, the borrow checker makes some code more awkward than it would be in GC languages, but the benefits are easily worth it and they stretch far beyond memory safety.

A huge part of the spirit of rust is fearless concurrency. The simple seeming false positive examples become non-trivial in concurrent code.

The author admits they don't write large concurrent - which clearly explains why they don't find much use in the borrow checker. So the problem isn't that the rust doesn't work for them - it's that a central language feature of rust hampers them instead of helping them.

The conclusion for this article should have been: if you're like me and don't write concurrent programs, enums and matches are great. The language would be work better for me if the arc/box syntax spam went away.

As a side note, if your code is a house of cards, it's probably because you prematurely optimized. A good way to get around this problem is to arc/box spam upfront with as little abstraction as possible, then profile, then optimize.

The other end of the spectrum is something like gamedev: you write code that pretty explicitly has an end-date, and the actual shape of the program can change drastically during development (because it's a creative thing) so you very much don't want to slowly build up rigidity over time.

  struct Point {
      x: f64,
      y: f64,
  }

  impl Point {
      fn x_mut(&mut self) -> &mut f64 {
          &mut self.x
      }

      fn y_mut(&mut self) -> &mut f64 {
          &mut self.y
      }
  }

I also disagree with the author that his rejected code:

  fn main() {
      let mut point = Point { x: 1.0, y: 2.0 };
      let x_ref = point.x_mut();
      let y_ref = point.y_mut();
      *x_ref *= 2.0;
      *y_ref *= 2.0;
  }

I think the spirit of Rust's ownership rules is quite clear that when calling a function whose signature is

  fn f<'a>(param: &'a mut T1) -> &'a mut T2;

This is often necessary for correctness (because there are many scenarios where you need to be guaranteed exclusive access to an object beyond just wanting to satisfy the LLVM "noalias" rules) and is not just an implementation detail: the language would be fundamentally different if instead the borrow checker tried to loosen this requirement as much as it could while still respecting the aliasing rules at a per-field level.

It doesn't need to be Rust: Rust's borrow checker has (mostly reasonable) limitations that eg. make some interprocedural things impossible while being possible within a single function (eg. &mut Vec<u32> and &mut u32 derived from it, both being used at the same time as shared references, and then one or the other being used as exclusive later). Maybe some other language will come in with a more powerful and omniscient borrow checker[^1], and leave Rust in the dust. It definitely can happen, and if it does then I suppose we'll enjoy that language then.

But: it is my opinion that a borrow checker is an absolutely massive thing in a (non-GC) programming language, and one that cannot be ignored in the future. (Though, Zig is proving me wrong here and it's doing a lot of really cool things. What memory safety vulnerabilities in the Ziglang world end up looking like remains to be seen.) Memory is always owned by some_one_, its validity is always determined by some_one_, and having that validity enforced by the language is absolutely priceless.

Wanting GC for some things is of course totally valid; just reach for a GC library for those cases, or if you think it's the right tool for the job then use a GC language.

[^1]: Or something even better that can replace the borrow checker; maybe Graydon Hoare's original idea of path based aliasing analysis would've been that? Who knows.

If this is true for Rust, it's 10x more true for C++!

Lifetime issues are puzzles, yes, but boring and irritating ones.

But in C++? Select an appetizer, entree, and desert (w/ bottomless breadsticks) from the Menu of Meta Programming. An endless festival of computer science sideshows living _in the language itself_ that juices the dopamine reward of figuring out a clever way of doing something.

Language support: You can implement extension traits on an integer so you can do things like current_node.next(v) (like if you have an integer named 'current_node' which is an index into a vector v of nodes) and customize how your next() works.

Also, I disagree there is 'zero safety', since the indexes are into a Rust vector, they are bounds checked by default when "dereferencing" the index into the vector (v[i]), and the checking is not that slow for vast majority of use cases. If you go out of bounds, Rust will panic and tell you exactly where it panicked. If panicking is a problem you could theoretically have custom deference code that does something more graceful than panic.

But with using indexes there is no corruption of memory outside of the vector where you are keeping your data, in other words there isn't a buffer overflow attack that allows for machine instructions to be overwritten with data, which is where a huge amount of vulnerabilities and hacks have come from over the past few decades. That's what is meant by 'safety' in general.

I know people stick in 'unsafe' to gain a few percent speed sometimes, but then it's unsafe rust by definition. I agree that unsafe rust is unsafe.

Also you can do silly optimization tricks like if you need to perform a single operation on the entire collection of nodes, you can parallelize it easily by iterating thru the vector without having to iterate through the data structure using next/prev leaf/branch whatever.

In principle, the language already has raw pointers with the same expressive power as in C, and unlike references they don't have aliasing restrictions. That is, so long as you only use pointers to access data, this should be fine (in the sense of, it's as safe as doing the same thing in C or Zig).

Note that this last point is not the same as "so long as you don't use references" though! The problem is that aliasing rules apply to variables themselves - e.g. in safe rust taking a mutable reference to, say, local variable and then writing directly to that variable is forbidden, so doing the same with raw pointers is UB. So if you want to be on the safe side, you must never work with variables directly - you must always take a pointer first and then do all reads and writes through it, which guarantees that it can be aliased.

However, this seems something that could be done in an easy mechanical transform. Basically a macro that would treat all & as &raw, and any `let mut x = ...` as something like `let mut x_storage = ...; let x = &raw mut x_storage` and then patch up all references to `x` in scope to `*x`.

The other problem is that stdlib assumes references, but in principle it should be possible to mechanically translate the whole thing as well...

And if you make it into a macro instead of patching the compiler directly, you can still use all the tooling, Cargo, LSP(?) etc.

If you try to write Java in Rust, you will fail. Rust is no different in this regard from Haskell, but method syntax feels so friendly that it doesn't register that this is a language you genuinely have to learn instead of picking up the basics in a couple hours and immediately start implementing `increment_counter`-style interfaces.

And this is an inexperienced take, no matter how eloquently it's written. You can see it immediately from the complaint about CS101 pointer-chasing graph structures, and apoplexy at the thought of index-based structures, when any serious graph should be written with an index-based adjacency list and writing your own nonintrusive collection types is pretty rare in normal code. Just Use Petgraph.

A beginner is told to 'Just' use borrow-splitting functions, and this feels like a hoop to jump through. This is because it's not the real answer. The real answer is that once you have properly learned Rust, once your reflexes are procedural instead of object-oriented, you stop running into the problem altogether; you automatically architect code so it doesn't come up (as often). The article mentions this point and says 'nuh uh', but everyone saying it is personally at this level; 'intermittent' Rust usage is not really a good Learning Environment.

Many projects are written in Rust that would absolutely be fine in Go, Swift or a JVM language. And I don't understand: it is nicer to write in those other languages, why choose Rust?

On the other hand, Rust is a lot nicer than C/C++, so I see it as a valid alternative there: I'm a lot happier having to make the borrow-checker happy than tracking tricky memory errors in C.

Every time I write Go, I find it so annoying all the defensive deep copying I see. In JS I always find myself getting confused on whether a mutation is safe (as in my program won't break some assumption) to do or not. Marking arguments as shared or exclusive is really great for me to know what kind of access I can have. It needs to be enforced so that the owner also doesn't accidentally mutate while a shared borrow is still active (again, not just for memory safety but for my own invariants). The classic example could be inserting into a collection while iterating over it

I think the borrow checker is necessary if you have ADTs like Rust and you want memory safety. You could pattern match a union into one of its variants and get a pointer to one of the fields, but without a borrow checker there's nothing to stop you from changing the variant stored in the union. This would obviously cause issues and the only way you'd solve this with GC alone is by allocating each variant individually where the union is just a tagged pointer.

This is true both in theory and in practice, as you can write any program with a borrow checker as you can without it.

TFA also dismisses all the advantages of the borrow checker and focuses on a narrow set of pain points of which every Rust developer is already aware. We still prefer those borrowing pain points over what we believe to be the much greater pain inflicted by other languages.

The compiler changed the type of my variable based on its usage. Usage in code I didn't write. There was no warning about this (even with clippy). The program crashed at runtime.

I found this amusing because it doesn't happen in dynamic languages, and it doesn't happen in languages where you have to specify the types. But Rust, with its emphasis on safety, somehow lured me into this trap within the first 15 minutes of programming.

I found it more amusing because in my other attempts at Rust, the compiler rejected my code constantly (which was valid and worked fine), but then also silently modified my program without warning to crash at runtime.

I saw an article by the developers of the Flow language, which suffered from a similar issue until it was fixed. They called it Spooky Action at a Distance.

This being said, I like Rust and its goals overall. I just wish it was a little more explicit with the types, and a little more configurable on the compiler strictness side. Many of its errors are actually just warnings, depending on your program. It feels disrespectful for a compiler to insist it knows better than the programmer, and to refuse to even compile the program.

- Marking and sweeping cause latency spikes which may be unacceptable if your program must have millisecond responsiveness.

- GC happens intermittently, which means garbage accumulates until each collection, and so your program is overall less memory efficient.

With modern concurrent collectors like Java's ZGC, that's not the case any longer. They show sub-millisecond pause times and run concurrently. The trade-off is a higher CPU utilization and thus reduced overall throughput, which if and when it is a problem can oftentimes be mitigated by scaling out to more compute nodes.

Indices aren't simply "references but worse". There are some advantages:

- they are human readable

- they are just data, so can be trivially serialized/deserialized and retain their meaning

- you can make them smaller than 64 bits, saving memory and letting you keep more in cache

Also I don't see how they're unsafe. The array accesses are still bounds-checked and type-checked. Logical errors, sure I can see that. But where's the unsafety?

It might be surprising to some folks, but there is a lot of unsafe code in Rust, and a lot of that is in the standard’s data structure implementations.

Also —

Common in network programming, the pain of lifetimes, is in async.

The model sort of keels over and becomes obtuse when every task requires ownership of its data with static lifetimes.

Both have rust-like flavor and neither has a borrow checker.

Data races prevented not just in concurrent programs. I personally was hit once by a panic from RefCell. I have coded a data race, but I wouldn't notice it without RefCell. I borrowed the value and then called some function, and a several stack frames deeper I tried to borrow it mutably.

This is actually a learning lesson for the user to understand that the bugs one has seen in languages like c++ are inherent to using simple types.

The author goes about mentioning python. If you do change all your types to python equivalents, ref counted etc. Rust becomes as easy. But you don’t want to do that and so it becomes pain, but pain with a gain. You must decide if that gain is worth it.

From my point of view the issue is that rust defaults to be a system programming language. Meaning, simple types are written simple (i32, b32, mut ..), complex types are written complex (ref, arc, etc.). And because of that one wants to use the simple types, which makes the solutions complex.

Let’s imagine a rust dialect, where every type without annotation is ref counted and if you want the simple type you would have to annotate your types, the situation would change.

What one must realize is that verifiable correctness is hard , the simplicity of the given problematic examples is a clear indication of how close those screw ups are even with very simple code. And exactly why we are still seeing issues in core c libs after decades of fixing them.

I think there's another part of the story in a few of these examples that the author might be missing. Sometimes Rust doesn't want to infer/allow things because of the Halting Problem or whatever, but other times it's because that inference would amount to a compatibility hazard that the user might not've realized they were creating. "Types implement `Copy` automatically" is a good example of one of these. It would be an easy rule to implement (and it is in fact how `Send` and `Sync` and some other things work), but it would also mean that adding a non-Copy private field to a struct that didn't previously have one could break callers. I wouldn't call it a slam-dunk case for the rules being what they are -- and you could argue that it prioritizes the needs of library code over the needs of application code -- but anyway it's some context.

No, you could use destructuring. This doesn't work for all cases but it does for your examples without needing to derive copy or clone. Here's a more complex but also compelling example of the problem:

    struct Graph {
        nodes: BTreeMap<u32, Node>,
    }

    struct Node {
         edges: Vec<u32>,
    }

    impl Graph {
        fn visit_mut(&mut self, visit: impl Fn(&mut Node, &mut Node)) {
            let mut visited = BTreeSet::new();
            let mut stack = vec![0];
            while let Some(id) = stack.pop() {
                if !visited.insert(id) { continue; }
                let curr = self.nodes.get_mut(&id);
                for id in source.edges.clone() {
                    let next = self.nodes.get_mut(&id);
                    visit(curr, next);
                    stack.push(id);
                }
            }
        }
    }

This is a problem that shows up all the time when using collections, which Rust encourages within the ecosystem.

This is a real problem across the entire industry, and Rust is a particularly egregious example because you get to justify playing with the fun stimulating puzzle machine because safety—you don't want unsafe code, do you? Meanwhile there's very little consideration to whether the level of rigidity is justified in the problem domain. And Rust isn't alone here, devs snort lines of TypeScript rather than do real work for weeks on end.

Most of the other criticisms are pretty disappointing, though.

> The following example is a famous illustration of how it can't properly reason across branches, either:

Except that function body would have been better rewritten as map.entry(key).or_default() -- which passes the borrow checker just fine and is more performant as it avoids multiple lookups. I suspect many other examples would benefit from being re-written into higher-level primitives that can be easier to borrow-check in this manner.

> But what's the point of the rules in this case, though? Here, the ownership rules does not prevent use after free, or double free, or data races, or any other bug. It's perfectly clear to a human that this code is fine and doesn't have any actual ownership issues.

What if Id represents a file descriptor that ought to be closed when the value ceases to exist? Or some other type of handle? This is an ownership problem: these are not limited to memory safety issues. For very good API stability reasons, without an explicit #[derive(Clone,Copy)] the compiler is not free to assume the type represents pure information that can be copied at will, but can only treat it as one potentially containing owned resources, that just happens not to include any at this time.

> References to temporary values, e.g. values created in a closure, are forbidden even though it's obvious to a human that the solution is simply to extend the lifetime of the value to its use outside the closure.

Which is to say, to what? The type signature does not say whether a closure will be called immediately or in another thread after two hours. And in which stack frame should the value be stored, the soon-to-be-exiting closure's?

Of course, designing for safety is quite complex and easy to get wrong. For example, Swift's "structured concurrency" is an attempt to provide additional abstractions to try to hide some complexity around life times and synchronization... but (personally) I think the results are even more confusing and volatile.

I’ll bite a little bit. If I write:

    struct Thing(u32);

Sure, I suppose someone could invent a language where an LLM reads the type definitions and fills in the blanks as to what uses are valid, but this seems like a terrible idea. Or one could use a language without affine types (most languages, and even C++’s move feature entirely fails to enforce any sort of good behavior).

This isn't really solvable in C/C++, and its worked around with a bunch of hacks, which might be overly convervative at times, and at others, generates buggy code.

Rust managed to fix this issue elegantly, but I'm wondering if the solution might be worse than the problem.

It's essentially impossible to write a Rust program without relying on many of its escape hatches like RefCell and unsafe, that make the borrow checker go away.

Essentially any program, that needs to communicate or store data in some sort of persistent structure, which is then accessed in some other part of the program, has multiple mutable references to said abstract central point, which is not allowed in Rust, without the aforementioned workarounds.

When I was going through its docs I was impressed with all those good ideas one after the other. Docs itself are really good (high information density that reads itself).

[0] https://www.moonbitlang.com

It seems to me that this is exactly the sort of thing that Rust is intended to prevent, and it makes complete sense to reject the code.

The borrow checker is certainly Rust’s claim to fame. And a critical reason why the language got popular and grew. But it’s probably not in my Top 10 favorite things about using Rust. And if Rust as it exists today existed without the borrow checker it’d be a great programming experience. Arguably even better than with the borrow checker.

Rust’s ergonomics, standardized cargo build system, crates.io ecosystem, and community community to good API design are probably my favorite things about Rust.

The borrow checker is usually fine. But does require a staunch commitment to RAII which is not fine. Rust is absolute garbage at arenas. No bumpalo doesn’t count. So Rust w/ borrow checker is not strictly better than C. A Rust without a borrow checker would probably be strictly better than C and almost C++. Rust generics are mostly good, and C++ templates are mostly bad, but I do badly wish at times that Rust just had some damn template notation.

> This is generally good advice. Usually, extra allocations are fine, and the resulting performance degradation is not an issue. But it is a little strange that it allocations are encouraged in an otherwise performance-forcused language, not because the program logic demands it, but because the borrowchecker does.

I often end up writing code that (seems) to do a million tiny clones. I've always been a little worried about fragmentation and such but it's never been that much of an issue -- I'm sure one day it will be. I've often wanted a dynamically scoped allocator for that reason.

    struct Id(u32);

    fn main() {
        let id = Id(5);
        let mut v = vec![id];
        println!("{}", id.0);
    }

1. The culture of licensing stuff under MIT by default rather than (A)GPL.

2. The syntax. I'd much rather have S-expressions.

As someone who writes Rust professionally this sentence is sus. Typically, the borrow checker is somewhere between 10th and 100th in the list with regards to things I think about when programming. At the end of the day, you could in theory just wrap something in a reference counter if needed, but even that hasn't happened to me yet.

This article didn't even make it 2 paragraphs before being incorrect. Safety isn't bug resistance - Safety is ensuring memory integrity. Bugs are still easy to write, they're just going to be logic bugs instead of use after frees, double frees, etc.

The first example of a function and access to fields misses point as well. If you just wanted to edit a field locally, you can do that without references. If you're sharing the struct around, then you don't want more than one mutable reference at a time anyway.

> But what's the point of the rules in this case, though? Here, the ownership rules does not prevent use after free, or double free, or data races, or any other bug. It's perfectly clear to a human that this code is fine and doesn't have any actual ownership issues

I mean, of course there is an obvious ownership issue with the code above, how are the destructors supposed to be ran without freeing the Id object twice?