246
Productivity of Rust teams at Google
(programming.dev)
Slide with text: “Rust teams at Google are as productive as ones using Go, and more than twice as productive as teams using C++.”
In small print it says the data is collected over 2022 and 2023.
I don't know. After writing rust for a while, and slowly putting programs together, I tried Go and I feel so relived I can just write what I want in 10 seconds instead of messing with lifetimes, borrow checker and other stuff I actually don't care about at all.
A more experienced colleague said that yes that is true, but Go can't guarantee your code is correct, so you will spend time fixing your code also in Go. Probably true.
Right, it's essentially the same argument as strong vs. weak typing. The weak typing proponents say JavaScript is best, because you can just write anything and you don't need to worry about all those pesky types getting in your way. The strong typing proponents (which if it's not obvious I am one of) point out that you can write incorrect code quickly in just about any language, but writing correct code is much harder, and the cost of correcting code increases the later the mistake is found. Errors that can't even be written are better than errors that are found at compile time which are better than errors that are reliably caught at runtime, which are all infinitely better than errors that only randomly appear under very specific circumstances.
That is why many people switched to using TypeScript for their websites instead of JavaScript, because even though you have to spend more time putting type annotations on everything, and at the end of the day at runtime TypeScript is literally just JavaScript, the errors it lets you find at compile time instead of runtime make the effort necessary to include those types worth it. Same thing applies with Rust vs. Go. Yes it requires more thinking up front when you're writing Rust code, and yes it might take you longer to write that code, but it's also going to be correct code you can be confident in and not have a bunch of ticking timebombs waiting in it that you don't even know about.
An extra 30 minutes spent having to think about a dozen lines of code, is infinitely preferable to spending 3 hours pouring over stack traces and single stepping debuggers to find that one subtle mistake you made.
I totally agree, though I think it's worth adding:
The advantages of static types is not just finding bugs (though it does do that quite well). It also massively helps with productivity because a) types are now documented, b) you can use code intelligence tools like renaming variables, go-to-definition, find-references, etc. (assuming you use a good editor/IDE).
In general stronger types are better but I do think there is a point at which the effort of getting the types right is too high to be worth the benefit. I would say Rust hasn't reached that point, but if you look at formal verification languages like Dafny, it's pretty clear that you wouldn't want to use that except in extreme circumstances. Similarly I think the ability to use an
anyordynamicescape hatch is quite useful, even if it should be used very sparingly.You are right. But I think similar secondary benefits also come from using the borrow checker. Rust developers, by necessity, try to avoid using circular references and prefer immutability where they can. Both of these are advantages because they tend to make for systems that are easier to understand and are easier to maintain.
Yeah I agree. The borrow checker definitely pushes you to write less buggy code.
Absolutely! Types are as much about providing the programmer with information as they are the compiler. A well typed and designed API conveys so much useful information. It's why it's mildly infuriating when I see functions that look like something from C where you'll see like:
rather than a better strongly typed version like:
I disagree with this, I don't think those are ever necessary assuming a powerful enough type system. Function arguments should always have a defined type, even if it's using dynamic dispatch. If you just want to not have to specify the type on a local,
letbindings where you don't explicitly define the type are fine, but even in that case it still has a type, you're just letting the compiler derive it for you (and if it can't it will error).You can go to definition / find references / rename for dynamically typed languages too.
E.g. https://github.com/palantir/python-language-server
Without static type annotations you can only make best effort guesses that are sometimes right. Better than nothing but not remotely the same as actual static types. The LSP you linked works best when you use static type annotations.
Also I would really recommend Pylance over that if you can - it's much better but is also closed source unfortunately.
Why would it just be best effort? To find references for a specific thing, it still would parse an AST, find the current scope, see it's imported from some module, find other imports of the module, etc.
Where is the definition of x? What is the type of x? If you can't identify it, neither can the LSP.
This kind of thing actually happens when implementing interfaces, inheritance, etc. Thus, LSPs in dynamic languages are best effort both theoretically and in practice.
Tbf this example can be deducted as
string | intjust fine.The real problem is when you start using runtime reflection, like
getattr(obj, "x")Types are not necessary at all.
Saying "x is defined somewhere in the entire program" isn't satisfactory to many users. Also, you didn't tell me what type x has. Can I do
x + 5?It absolutely does. Without static types an IDE/LSP can't reliably find all the references / definition and therefore can't refactor reliably either.
Consider something like this:
Now imagine you want to rename
Foo.baror find all references to it. Impossible without the type annotations.Ah, I see. You're talking about object properties. I don't see any issue with finding references to variables, but for properties, yeah.
It breaks down when you do runtime reflection, like
getattr(obj, "x").Ok imagine you are a LSP. What type is
x? Isprizea typo? What auto-complete options would you return forx.?I didn't say types. I said find references / go to definition / rename.
How are you going to find references to
prize, go to its definition or rename it without knowing what typexis? It's impossible without static types.Preach 🙏
There's nothing wrong with putting Rc<_> or Rc<RefCell<_>> around data if you don't want to fight the borrow checker or think about lifetimes even if you know it can be written without.
Or even just clone. Depending on use case the performance cost would be negligible.
It's mainly the visual pollution that bothers me. Wrapping everything in the reference counting smart pointers just because you can't be bothered dealing with the borrow checker seems like an antipattern
I don't know why so many recommend Rc or Arc as a catchall. 90% of the time if you want to avoid the borrow checker then a clone or copy is good enough.
Really? I might have agreed for some other languages, but Go is so bare bones it feels like it takes way longer to write simple stuff than with Rust - you have to tediously write out loops all the time for example.
Tbf I haven't used it since it got generics. Maybe it is better now.
#Rust is not high level at all, change my mind.
Rust gives you high level abstractions but also allows lower level control over the hardware. These are not mutually exclusive.
You can easily argue almost any language is high level though, it is such a nebulous term that it is almost meaningless.
But Go has garbage collection so that code is as correct as that of Rust. Go is just a little less performant
Rusts ownership model is not just an alternative to garbage collection, it provides much more than that. It's as much about preventing race conditions as it is in making sure that memory (and other resources) get freed up in a timely fashion. Just because Go has GC doesn't mean it provides the same safety guarantees as Rust does. Go's type system is also weaker than Rusts even setting aside the matter of memory management.
Rust still can have race conditions
True, but ownership does eliminate a lot of the possible sources of them.
Also go is quite a lot slower than rust. It seems fast compared to python or course but it's probably half the speed of Rust.