Productivity of Rust teams at Google

@[email protected] · edit-2 3 months ago

Productivity of Rust teams at Google

@sugar_in_your_tea · 3 months ago

There are of course macros

Yup, and Rust’s macros are pretty cool, but in D you can just do:

static if (condition) {
    ...
}

There’s a whole compile-time reflection library as well, so you can take a class and make a super-optimized serialization/deserialization library if you want. It’s super cool, and I built a compile-time JSON library just because I could…

You kind of get that with Rust for free

Yup, Rust is awesome.

But in D you can do explicit scope guards:

scope(exit) - basically Go’s defer()
scope(success) - only runs when no exceptions are run
scope(failure) - only runs when there’s an exception

I didn’t use them much, but they are really cool, so you can do explicit cleanup as you go through the logic flow, but defer them until they’re needed.

It’s a neat alternative to RAII, which D also supports.

Some really great work is being done on that pretty much all the time

I still need to try out Cranelift, which was posted here recently. Cranelift release mode could mostly solve this for me.

That said, I haven’t touched D in years since moving to Rust, so I obviously find more value in it. But I do miss some of the candy.

@[email protected] · 3 months ago

But in D you can do explicit scope guards

Hmm… that is interesting.

scope(exit) is basically just an inline std::ops::Drop trait, I actually think it’s a bad thing that you can mix that randomly into your code as you go instead of collecting all of the cleanup actions into a single function. Reasoning about what happens when something gets dropped seems much more straightforward in the Rust case. For instance it wasn’t immediately clear that those statements get evaluated in reverse order from how they’re encountered which is something I assumed, but had to check the documentation to verify.

scope(success) and scope(failure) are far more interesting as I’m not aware of a direct equivalent in Rust. There’s the nightly only feature of std::ops::Try that’s somewhat close to that, but not exactly the same. Once again though, I’m not convinced letting you sprinkle these statements throughout the code is actually a good idea.

Ultimately, while it is interesting, I’m actually happy Rust doesn’t have that feature in it. It seems like somewhat of a nightmare to debug and something ripe to end up as a footgun.

@sugar_in_your_tea · 3 months ago

For instance it wasn’t immediately clear that those statements get evaluated in reverse order

It’s a stack, just like Go’s defer().

scope(success) and scope(failure) are far more interesting as I’m not aware of a direct equivalent in Rust

Probably because Rust doesn’t have exceptions, and I’m pretty sure there are no guarantees with panic!().

Ultimately, while it is interesting, I’m actually happy Rust doesn’t have that feature in it

Same, but that’s because Rust’s semantics are different. It’s nice to have the option if RAII isn’t what you want for some reason (it usually is), but I absolutely won’t champion it since it just adds bloat to the language for something that can be solved another way.

@[email protected] · 3 months ago

Probably because Rust doesn’t have exceptions

Well, it has something semantically equivalent while being more explicit, which is Result (just like Option is the semantic equivalent of null).

and I’m pretty sure there are no guarantees with panic!().

I actually do quite a bit of bare metal Rust work so I’m pretty familiar with this. There are sort of guarantees with panic. You can customize the panic behavior with a panic_handler function, and you can also somewhat control stack unwinding during a panic using std::panic::catch_unwind. The later requires that anything returned from it implement the UnwindSafe trait which is sort of like a combination Send + Sync. That said, Rust very much does not want you to regularly rely on stack unwinding. Anything that’s possible to recover from should use Result rather than panic!() to signal a failure state.

@sugar_in_your_tea · 3 months ago

Yup. My point is just that scope(failure) could be problematic because of the way Rust works with error handling.

What could maybe be cool is D’s in/out contracts (example pulled from here):

int fun(ref int a, int b)
in
{
    assert(a > 0);
    assert(b >= 0, "b cannot be negative!");
}
out (r)
{
    assert(r > 0, "return must be positive");
    assert(a != 0);
}
do
{
    // function body
}

The scope(failure) could partially be solved with the out contract. I also don’t use this (I find it verbose and distracting), but maybe that line of thinking could be an interesting way to generically handle errors.

@[email protected] · 3 months ago

Hmm… I think the Rust-y answer to that problem is the same as the Haskell-y answer, “Use the Types!”. I.E. in the example above instead of returning an i32 you’d return a NonZero<u32>, and your args would be a: &NonZero<u32>, b: u32. Basically make invalid state unrepresentable and then you don’t need to worry about the API being used wrong.

@sugar_in_your_tea · 3 months ago

I’m more referring to a more general application, such as:

fn do_stuff() -> Result<...> {
    if condition {
        return Error(...)
    }

    return Ok(...)
} out (r) {
    if r.is_err() {
        // special cleanup (maybe has access to fn scope vars)
    }
}

That gives you some of the scope(failure) behavior, without as many footguns. Basically, it would desugar to:

fn do_stuff() -> Result<...> {
    let ret = if condition { Error(...) } else { Ok(eee) };

    if ret.is_err() {
        ...
    }

I’m not proposing this syntax, just suggesting that something along these lines may be interesting.

@[email protected] · edit-2 3 months ago

I think the issue with that is that it’s a little bit of a solution in search of a problem. Your example of:

fn do_stuff() -> Result<...> {
    if condition {
        return Error(...)
    }

    return Ok(...)
} out (r) {
    if r.is_err() {
        // special cleanup (maybe has access to fn scope vars)
    }
}

isn’t really superior in any meaningful way (and is arguably worse in some ways) to:

fn do_stuff() -> Result<...> {
    if condition {
        // special cleanup (maybe has access to fn scope vars)
        return Error(...)
    }

    return Ok(...)
}

For more complicated error handling the various functions on Result probably have all the bases covered.

For what it’s worth a lot of my day to day professional work is actually in Java and our code base has adopted various practices inspired by Rust and Haskell. We completely eliminated null from our code and use Optional everywhere and use a compile time static analysis tool to validate that. As for exception handling, we’re using the Reactor framework which provides a type very similar to Result, and we essentially never directly throw or catch exceptions any more, it’s all handled with the functions Reactor provides for error handling.

I just don’t think the potential footguns introduced by null and exceptions are worth it, the safer type level abstractions of Option and Result are essentially superior to them in every way.

@sugar_in_your_tea · edit-2 3 months ago

We completely eliminated null from our code

Nice. We use Python and use None everywhere. I ran pyright on our codebase, and while we use typing religiously, our largest microservice has ~6k typing errors, most of which are unchecked Nones. We also use exceptions quite a bit, which sucks (one thing really annoys me is a function like check_permissions() which returns nothing, and throws if there’s an issue, but it could totally just return a bool. We have nonsense like that everywhere.

I use Rust for all of my personal projects and love not having to deal with null everywhere. I’d push harder for it at work if others were interested, but I’m the only one who seems passionate about it (about 2-3 are “interested,” but haven’t even done the tutorial).