I set a rule for myself that I'll spend up to one minute trying to save an allocation. Beyond that it's not worth getting sidetracked.
- Never combine two things that can be valid for different amounts of time into one struct. For example: A file struct should not contain both information about how it is formatted, which is true forever and can be used for many files, and a file handle, which could be invalidated by the OS at any time. Breaking this rule will fill your code with Box/Arc as you try to imitate classical OOP.
- Don't be afraid to frequently pass contextual information to functions; you don't need to put everything that will remain the same between two calls into `self`. For example, every function that works with the file can take the file handle and the format information as separate arguments. Trying to DRY function arguments by combining data with different lifetimes into a single struct and then hiding that argument in the `self` parameter might feel like simplification, but in Rust it triggers the above problem.
- Functions that call functions that take mutable references as output locations should do the same, unless they have to allocate for another reason. This rule of thumb will tend to push allocation as far outside of loops as possible.
With these in hand, I almost never need to box or reference count anything. If you fail to heed the fact that Rust is not really an OOP language, your entire program will start to look like the hairy parts of C libraries that interface with the Python interpreter.
Any tips for those of us who see PyO3 as one of Rust's biggest killer apps?
(Honest question. I hate how unmaintainable Python is but I don't know of any better equivalent to PyQt/PySide's memory-safe QWidget bindings or the RAD-friendly ORM migrations in Django ORM or SQLAlchemy+Alembic.)
My rule is that I'll do what is the easiest, unless it's an inner loop that runs all the time, where I'll try to make it fast.
In rust, usually the easiest is to not allocate or copy data. When it's easier to allocate or copy, I don't do a cost-benefit analysis at all, I just do it.
Honestly, this is a really good way to put it, and that one minute rule sounds like a pretty good rule (ignoring cases where performance requirements led to profiling, which pointed you at some specific piece that you need to optimize, obviously)
In Rust you control yourself the allocation on the stack or the heap. So before reaching out to a memory pool/arena, it's too tempting trying to put everything on the stack for "maximum performance".
Strangely i haven't had many of the issues that the previous poster was discussing, though. My issues are usually trying to work around the lack of GATs, lack of trait aliasing, etc. But i use `async_trait` so maybe i'm sidestepping many of the issues from the previous post. /shrug
https://raw.githubusercontent.com/luser/keep-calm-and-call-c...
True, but then those languages are heavily optimized for that scenario. This is why Rust written like Java often performs worse than Java (and optimized Rust).
This! Though I modified my rule to mostly being "I'm punting this to the weekend as a fun exercise", since I often find stupid enjoyment in seeing if it's possible.
Programming in Rust can often be a very good exercise in telling yourself to stop trying to be clever.
In rust the unspoken rule seems to be to allocate, allocate, allocate - unless you are writing a special purpose library etc. Which makes me wonder, is rust actually faster than a GC'd language when doing heavy async work?
> In rust the unspoken rule seems to be to allocate, allocate, allocate when you run into a lifetime issue
Lifetimes work fine with async, but some types of lifetimes can be problematic, for sure.
Yeah. And that is almost always an error on those languages too.
In rust you will have to declare it exactly like that. That's not really an example of passing mutable references between threads.
Rust gives you the tools to write very high performance code, but it doesn’t have to be about that.
I’ve had to point out to people quite a few times that garbage collectors can improve performance… especially compared to naive implementations of manual memory management. GCs are not just a tool for lazy programmers. GCs can make allocation incredibly fast, and you get to defer cleanup work to another thread(s), which means less work in the critical path. Removing work from the critical path is how you make software faster. Every tool has tradeoffs, and GCs are a tool. GCs often use more memory as a tradeoff.
I like Rust well enough, but I do wish we had a language that combined the ergonomics of Rust with the dead simple concurrency model of Go/Erlang. I haven’t tried it, but Luantic looks promising: https://github.com/lunatic-solutions/lunatic
As it is, we’re fortunate to have quite a few great languages and platforms these days.
It's almost like Rust and Elixir had a baby :D
As for the rest of the comment: totally. Almost none of the code I write in Rust needs C-like performance and I still choose Rust for it.
Aren't these correctness guarantees only for performance related factors though? (allocation/memory and concurrency). The rest of your program would be just as correct in any other language.
It’s natural that a lot of programs have some form of concurrency, so that’s an extremely common thing for Rust to help with, but it’s not the only thing.
For me, the number-one Rust feature I love is that, by baking monadic optionality and error return in from the beginning (Option<T> and Result<T, E>), I can trust that, unless an author abuses panic! (in which case I never trust their code again), I can see all a function's return paths in its type signature. (Without having to choke down a pure functional language like Haskell with that currying-based function call syntax that I can never get used to.)
The runners-up are how fast Rust starts compared to Python or Java or similar when writing a CLI tool and how nice PyO3 makes safely writing libraries or backends for tools that need to be in Python for some reason like "nothing but PyQt, PySide and possibly QtJambi offers memory-safe QWidget bindings... and I've never found a Java app that didn't feel laggy and sluggish on X11".
- Can compile to anything, you can use one language to write Cuda Kernels, backend services, distributed processing jobs, WASM apps, and native apps.
- Performance, comparable to better than C.
I like fast languages as it avoids headaches that come from slow languages. If rust wasn't fast, I'd probably just stick to a polyglot language portfolio and pass it by.
EDIT: I also just clicked through to lunatic, that does look very promising!
and that's when you usually get data races ;)
> In rust the unspoken rule seems to be to allocate, allocate, allocate - unless you are writing a special purpose library etc. Which makes me wonder, is rust actually faster than a GC'd language when doing heavy async work?
I think it's a bit more nuanced. `Arc` is still a "smart pointer", so while it's not a straight Rust reference, it acts as a pointer. Yes, it allocates and it needs to do a reference count, but the overhead is very small. So while in practice it is not "zero cost", it's usually negligible.