A collection of software engineering techniques for effectively expressing intent with Rust.

• Cleanup
  • Combating Rightward Pressure
    • Basics
    • Tuple Matching
  • Iteration Issues
    • Pulling the First Error out of an Iterator over Results
    • Reverse Iterator Range
    • Empty and Singular Iterators
    • Enum Variants as Functions
• Blocks for Clarity
  • Closure Capture
• Ergonomics
  • Unification and Reading the Error Messages That Matter
  • Write-Compile-Fix Loop Latency
  • Caching with sccache
  • Editor support for jumping to compiler errors
• Lockdown
  • Never
  • Making traits unimplementable
  • Deactivating Mutability
• Avoiding Limitations
  • Box<FnOnce>

This section is about improving clarity.

After sparring with the compiler, it's not unusual to stand back and see several nested combinator chains or match statements. Much of the art of writing clean Rust has to do with the judicious application of de-nesting techniques.

• Use ? to flatten error handling, but be careful not to convert errors into top-level enums unless it makes sense to handle them at the same point in your code. Keep separate concerns in separate types.
• Split combinator chains apart when they grow beyond one line. Assign useful names to the intermediate steps. In many cases, a multi-line combinator chain can be more clearly rewritten as a for-loop.
• pattern match on the full complex type instead of using nested match statements
• If your match statement only has a single pattern that you care about, followed by a wildcard, replace the match statement with an if let My(Match(Pattern(thing))) = matched_thing { /*...*/ } possibly with an else branch if you cared about the wildcard earlier.
• Run cargo clippy! It can provide many legitimately helpful suggestions for cleaning up your code

If you find yourself writing code that looks like:

let a = Some(5);
let b = Some(false);

let c = match a {
    Some(a) => {
        match b {
            Some(b) => whatever,
            None => other_thing,
        }
    }
    None => {
        match b {
            Some(b) => another_thing,
            None => a_fourth_thing,
        }
    }
};

it can be de-nested by doing a tuple match:

let a = Some(5);
let b = Some(false);

let c = match (a, b) {
    (Some(a), Some(b)) => whatever,
    (Some(a), None) => other_thing,
    (None, Some(b)) => another_thing,
    (None, None) => a_fourth_thing,
};

As a special case, matching on tules of booleans can be used to encode decision tables. For example, here's roughly how cargo new handles --bin and --lib arguments:

let kind = match (args.is_present("bin"), args.is_present("lib")) {
    (true, true) => failure::bail!("can't specify both lib and binary outputs"),
    (false, true) => NewProjectKind::Lib,
    // default to bin
    (_, false) => NewProjectKind::Bin,
};

The collect method is extremely powerful, and if you have an iterator of Result types, you can use it to either return a collection of the Ok items, or the very first Err item.

From the std docs on collect:

let results = [Ok(1), Err("nope"), Ok(3), Err("bad")];

let result: Result<Vec<_>, &str> = results.iter().cloned().collect();

// gives us the first error
assert_eq!(Err("nope"), result);

let results = [Ok(1), Ok(3)];

let result: Result<Vec<_>, &str> = results.iter().cloned().collect();

// gives us the list of answers
assert_eq!(Ok(vec![1, 3]), result);

Seen in Sunjay's tweet

This functionality is unlocked by the Result type implementing FromIterator<Result<A, E>> for Result<V, E> where V implements FromIterator<A>. This may look a bit hairy, but it means that collect (which relies on the FromIterator trait for its functionality) can output a Result where the success type is some collection that can be built from A - the success type of the original results. The error type E is the same in both, meaning that the returned Result will just return the first encountered error.

In Rust, we can write for item in 0..50 to go from 0 to 49 but what if we wanted to iterate from 49 to 0? Many of us have written for item in 50..0 and been surprised that nothing happened. Instead, we can write:

// iterate from 49 to 0
for item in (0..50).rev() {}

// iterate from 50 to 0
for item in (0..=50).rev() {}

// iterate from 50 to 1
for item in (1..=50).rev() {}

Under the hood, when we write a range with this syntax, we are constructing a RangeInclusive instead of the normal Range. You can also construct ranges for everything with .., or have a range be half-open like ..50 or ..=50 or 0...

Seen in Andrea Pessino's tweet

The standard library also includes helpers for empty and singular iterators, using the functions std::iter::empty and std::iter::once, which can be a small cleanup of common code like vec![].into_iter() or vec![my_item].into_iter().

You may have received an error message at some point when you wrote an enum variant, but not the members inside it, and it complained about how you supplied a function instead of an enum:

enum E {
    A(u64),
}

// ERROR: expected enum `E`, found `fn(u64) -> E {E::A}`
let a: E = E::A;

Well, it turns out that enum tuple variants can be used as functions from their members to an instance of that enum. This can be used to encapsulate items in a collection inside that variant:

// create a vector of E::A's using the variant as a constructor function
let v_of_es: Vec<E> = (0..50).map(E::A).collect();

// v_of_es is now vec![A(0), A(1), A(2), A(3), A(4), ..]

// create a vector of Options using Some as a constructor function
let v_of_options: Vec<Option<u64>> = (0..50).map(Some).collect();

Blocks allow us to detangle complex expressions, and can be used anywhere that a one-liner expression would be valid.

Specifying variables for use in a closure can be frustrating, and it's common to see code that jumps through hoops to avoid shadowing variables. This is quite common when cloning an Arc before spawning a new thread that will own it. But a closure definition is an expression. Anywhere a closure is accepted, we could use a block that evaluates to a closure. In the example below, we use blocks to avoid shadowing the config that we want to pass to several threads, without creating gross names like config1, config2 etc... Seen in salsa and described in more detail in Rust pattern: Precise closure capture clauses.

Before, painfully avoiding shadowing config:

fn spawn_threads(config: Arc<Config>) {
    let config1 = Arc::clone(&config);
    thread::spawn(move || do_x(config1));
    
    let config2 = Arc::clone(&config);
    thread::spawn(move || do_y(config2));
}

After, no need to invent config_n names:

fn spawn_threads(config: Arc<Config>) {
    thread::spawn({
        let config = Arc::clone(&config);
        move || do_x(config)
    });
    
    thread::spawn({
        let config = Arc::clone(&config);
        move || do_y(config)
    });
}

One of the most important aspects of feeling at peace with the Rust programming language is to find harmony with the compiler. We've all introduced a single error and been whipped in the face by dozens of error messages. Even after years of professional Rust usage, it can feel like a cause for celebration when there are no errors after introducing more than a few new lines of code. Remember that the strictness of the compiler is what gives us so much freedom. Rust is useful for building back-ends, front-ends, embedded systems, databases, and so much more because the compiler knows how long our variables are valid for without using a garbage collector at runtime. Any lifetime-related bug that fails to compile in Rust might have been an exploitable memory corruption issue in C or C++. The compiler pain frees us from exploitation and gives us the ability to work on a wider range of projects.

Rust requires that arguments and return types are made explicit in function definitions. The compiler will use these explicit types at the boundaries of a function to drive type inference. It will take the input argument types and work from the top of the function toward the bottom. It will take the return type and work its way up. Hopefully they can meet in the middle. The process under the hood is actually a little more complicated than this but this simplified model is adequate to reason about this particular subject. The point is, there has to be an unbroken chain of type evidence that connects the input arguments to the return type through the body. When there is a gap in the chain, all ambiguous types will turn into errors. This is partially why rust will emit many pages of errors sometimes when there's actually only a single thing that needs to be fixed.

A big part of avoiding compiler fatigue is to just filter out the errors that don't matter. Start with the first one, and work your way down. See the next section for a command that will do this automatically when your code changes.

Programming Rust is a long game. It's common to see beginners spending lots of energy switching back and forth between their editor and a terminal to run rustc, and then scrolling around to find the next error that they want to fix. This is high-friction, and will tire you out faster than if this was automated.

There is a cargo plugin called cargo watch that will look for changes in source files descendent from the current working directory, and then run cargo check which skips the LLVM codegen and only looks for compilation errors in your Rust code. It can be installed by typing cargo install cargo-watch.

You can use the cargo watch plugin to call a specific command when your code changes as well. I like to filter out the lines after the beginning of the error messages, after clearing the terminal:

cargo watch -s 'clear; cargo check --tests --color=always 2>&1 | head -40'

This way I just save my code and it shows the next error.

sccache is a tool written by Mozilla that supports ccache-style build caching for Rust. This is particularly useful if you frequently clean and build projects with lots of dependencies, as normally they would all need to be recompiled, but with sccache they will be stored and used to back a cache that is accessible while building any project on your system. It takes care to do the right thing in the presence of different feature flags, versions etc... For projects with lots of dependencies it can make a huge difference over time.

The installation is simple:

cargo install sccache
export RUSTC_WRAPPER=sccache

It can also be configured to use a remote cache like s3, memcached, redis, etc... which is quite useful for building speedy CI clusters.

To go even farther than the last section, most editors have support for jumping to the next Rust error. In vim, you can use the vim.rust plugin in combination with Syntastic to automatically run rustc when you save a file, and to jump to errors using a keybind. Emacs users can use flycheck-rust for similar functionality.

This section is about preventing undesirable usage.

To make a type that can never be created, simply create an empty enum. Use this where you want to represent something that should never actually exist, but a placeholder is required. This is being brought into the standard library piece by piece, and it's already possible to have a function return ! if it exits the process or ends in an infinite loop (important for embedded work where main should never return). This can also be used when working with a Result that will never actually be an Err but you need to adhere to that interface.

enum Never {}

let never = Never:: // oh yeah, can't actually create one...

If you want to prevent others from implementing your Trait, use the following pattern to "seal" it so that only your implementations will ever exist. Seen in tokio-tls.

mod sealed {
    pub trait Sealed {}
}

// must have access to sealed::Sealed to
// implement this trait, which is not possible
// for other crates etc...
pub trait MyPublicTrait: sealed::Sealed {}

pub struct MyStruct;

impl sealed::Sealed for MyStruct {}

impl MyPublicTrait for MyStruct {}

Here's a pattern for disabling mutability for "finalized" objects, even in mutable owned copies of a thing, preventing misuse. Done by wrapping it in a newtype with a private inner value that implements Deref but not DerefMut:

mod config {
    #[derive(Clone, Debug, PartialOrd, Ord, Eq, PartialEq)]
    pub struct Immutable<T>(T);

    impl<T> Copy for Immutable<T> where T: Copy {}

    impl<T> std::ops::Deref for Immutable<T> {
        type Target = T;

        fn deref(&self) -> &T {
            &self.0
        }
    }
    
    #[derive(Default)]
    pub struct Config {
        pub a: usize,
        pub b: String,
    }
    
    impl Config {
        pub fn build(self) -> Immutable<Config> {
            Immutable(self)
        }
    }
}

use config::Config;

fn main() {
    let mut under_construction = Config {
        a: 5,
        b: "yo".into(),
    };
    
    under_construction.a = 6;
    
    let finalized = under_construction.build();
    
    // at this point, you can make tons of copies,
    // and even if somebody has an owned local version,
    // they won't be able to accidentally change some
    // configuration that
    println!("finalized.a: {}", finalized.a);
    
    let mut finalized = finalized;
    
    // the below WON'T work bwahahaha
    // finalized.a = 666;
    // finalized.0.a = 666;
}

Currently, it's not possible to call Box<FnOnce(T) -> R> on stable Rust. The common workaround is to use Box<FnMut(T) -> R>, store internal state inside of an Option and take the state out (with a potential run-time panic) in the call. However, a solution that statically guarantees that fn can be called at most once is possible. Seen in Cargo.

trait FnBox<A, R> {
    fn call_box(self: Box<Self>, a: A) -> R;
}

impl<A, R, F: FnOnce(A) -> R> FnBox<A, R> for F {
    fn call_box(self: Box<F>, a: A) -> R {
        (*self)(a)
    }
}

fn demo(f: Box<dyn FnBox<(), String>>) -> String {
    f.call_box(())
}

#[test]
fn test_demo() {
    let hello = "hello".to_string();
    let f: Box<dyn FnBox<(), String>> = Box::new(move |()| hello);
    assert_eq!(&demo(f), "hello");
}

Note that self: Box<Self> is stable and object-safe.