It's not like that you can't write relational models in the safe Rust. The only forbidden thing is a reference pointing arbitrary memory, which is typically worked around via indices and often more performant in that way. It is much rarer to find applications that need an absolutely random pointer that can't be hidden in abstractions in my opinion.
While beginner resources don't dwell too much upon cyclic references, they don't consider unsafe blocks as unusual. All the material I've seen say that there are certain domains where Rust's compile-time safety model simply won't work. What Rust allows you to do instead, is to limit the scope of unsafe blocks. However, the beginner material often won't give you too much details on how to analyze and decide on these compromises.
Anyway, compile-time safety checks (using borrow checker) and manual safety checks (using unsafe) aren't the only way to deal with safety. Cyclic references can be dealt with runtime safety checks too - like Rc and Weak.
Indeed. Starting out with code sprinkled with Rc, Weak, RefCell, etc is perfectly fine and performance will probably not be worse than in any other safe languages. And if you do this, Rust is pretty close to those languages in ease of use for what are otherwise complex topics in Rust.
A good reference for different approaches is Learn Rust With Entirely Too Many Linked Lists https://rust-unofficial.github.io/too-many-lists/
It is no accident that while Val/Hylo, Chapel and Swift have taken inspiration from Rust, they have decided to not inflict the affine types directly into the language users, rather let the compiler do part of the work itself.
a doubly linked list
Parent field is something where you have a clear hierarchy (it's not really “cyclic”, so it's the perfect use-case for weak references).
When coming from a managed-memory language, this obviously requires some conceptual effort to understand why this is a problem at all and how to deal with it, but when compared to C or C++, the situation is much better in Rust.
It's very noticeable how often people who "need" a linked list actually only wanted a queue (thus Rust's VecDeque) or even a growable array (ie Vec).
Aria has a long list of excuses people offer for their linked lists, as well as her discussion of the time before Rust 1.0 when she sent lots of Rust's standard library collections to that farm up-state but wasn't able to send LinkedList.
I've still been meaning to write something in / learn Rust's ways of thinking; is there not an intended replacement for these data structures? Or do they expect it all to go under Unsafe?
For linked-lists, there's one in std and the majority of people should never have to write their own as it's error prone and requires unsafe.
For graph use-case then you can either use ECS, arena, ref counting or unsafe, but you're probably better off using/developing a dedicated crate that optimizes it and abstract it away behind an easy to use (and safe) interface.
Sometimes such a primitive doesn't exist and you should use unsafe yourself, but then you're the one supposed to make sure that your code is in fact sound. If you keep unsafe for small portions of the code you can reason about and extensively test so Miri gives you a good level of confidence, then it's fine to use unsafe. But it's the more expensive option, not the default.
I just put up a blog post about this actually :) https://jacko.io/object_soup.html
> promoting the shared state to a container object for the set
Yeah I think that's a good way to describe these "ECS-ish" patterns.
They are.
I have not had to use cyclic references ever, except once doing experiments with fully connected graphs, that was very unusual
If you're doing a lot cyclic references Rust is not the right choice. Horses for courses
But are you sure you're using the best algorithm?
Doesn't it just come down to the question of who owns the node?
If it's a tree, and parents are never removed before children, just make the child owned by the parent and keep a weak reference to the parent.
If it's a general graph, and vertices can exist or not exist regardless of edges, keep a list of them independent of the edges, and keep weak references in the edges.
If it's a graph where a one or a few roots exist, and nodes exist as long as there's a path from a root node to them, that sounds like a classic use case for Rc<>.
Is there a common use case I'm missing?
https://jacko.io/object_soup.html
Agreed that I wish more beginner Rust books had a section about this. The pattern is quite simple, but it's hard for beginners who get stuck to realize that they need it.
While the Rust devs believe such cyclic references are rare - I think this speaks mostly to the problem domain they are focused on. Relational models are everywhere in Apps, they are often common in complex systems software like databases, and they are fairly rare in firmware/drivers/system code code.
There are a few patterns for dealing with cyclic references, but they all end up requiring either unsafe or a main "owner" object which you clean up occassionally (effectively arena allocation). Having now worked in C/CPP - the idea of having unsafe blocks sprinkled around the code doesn't bother me, and many C/CPP components have some form of arena allocation built-in. I just wish Rust learning resources would be more upfront about this.