entrait

A proc macro for designing loosely coupled Rust applications.

entrait is used to generate an implemented trait from the definition of a regular function. The emergent pattern that results from its use enable the following things:

• Zero-cost loose coupling and inversion of control
• Dependency graph as a compile time concept
• Mock library integrations
• Clean, readable, boilerplate-free code

The resulting pattern is referred to as the entrait pattern (see also: philosophy).

Introduction

The macro looks like this:

#[entrait(MyFunction)]
fn my_function<D>(deps: &D) {
}

which generates a new single-method trait named MyFunction, with the method signature derived from the original function. Entrait is a pure append-only macro: It will never alter the syntax of your function. The new language items it generates will appear below the function.

In the first example, my_function has a single parameter called deps which is generic over a type D, and represents dependencies injected into the function. The dependency parameter is always the first parameter, which is analogous to the &self parameter of the generated trait method.

To add a dependency, we just introduce a trait bound, now expressable as impl Trait. This is demonstrated by looking at one function calling another:

#[entrait(Foo)]
fn foo(deps: &impl Bar) {
    println!("{}", deps.bar(42));
}

#[entrait(Bar)]
fn bar<D>(deps: &D, n: i32) -> String {
    format!("You passed {n}")
}

Multiple dependencies

Other frameworks might represent multiple dependencies by having one value for each one, but entrait represents all dependencies within the same value. When the dependency parameter is generic, its trait bounds specifiy what methods we expect to be callable inside the function.

Multiple bounds can be expressed using the &(impl A + B) syntax.

The single-value dependency design means that it is always the same reference that is passed around everywhere. But a reference to what, exactly? This is what we have managed to abstract away, which is the whole point.

Runtime and implementation

When we want to compile a working application, we need an actual type to inject into the various entrait entrypoints. Two things will be important:

• All trait bounds used deeper in the graph will implicitly "bubble up" to the entrypoint level, so the type we eventually use will need to implement all those traits in order to type check.
• The implementations of these traits need to do the correct thing: Actually call the entraited function, so that the dependency graph is turned into an actual call graph.

Entrait generates implemented traits, and the type to use for linking it all together is Impl<T>:

#[entrait(Foo)]
fn foo(deps: &impl Bar) -> i32 {
    deps.bar()
}

#[entrait(Bar)]
fn bar(_deps: &impl std::any::Any) -> i32 {
    42
}

let app = Impl::new(());
assert_eq!(42, app.foo());

The linking happens in the generated impl block for Impl<T>, putting the entire impl under a where clause derived from the original dependency bounds:

impl<T: Sync> Foo for Impl<T> where Self: Bar {
    fn foo(&self) -> i32 {
        foo(self) // <---- calls your function
    }
}

Impl is generic, so we can put whatever type we want into it. Normally this would be some type that represents the global state/configuration of the running application. But if dependencies can only be traits, and we always abstract away this type, how can this state ever be accessed?

Concrete dependencies

So far we have only seen generic trait-based dependencies, but the dependency can also be a concrete type:

struct Config(i32);

#[entrait(UseTheConfig)]
fn use_the_config(config: &Config) -> i32 {
    config.0
}

#[entrait(DoubleIt)]
fn double_it(deps: &impl UseTheConfig) -> i32 {
    deps.use_the_config() * 2
}

assert_eq!(42, Impl::new(Config(21)).double_it());

The parameter of use_the_config is in the first position, so it represents the dependency.

We will notice two interesting things:

• Functions that depend on UseTheConfig, either directly or indirectly, now have only one valid dependency type: Impl<Config>1.
• Inside use_the_config, we have a &Config reference instead of &Impl<Config>. This means we cannot call other entraited functions, because they are not implemented for Config.

The last point means that a concrete dependency is the end of the line, a leaf in the dependency graph.

Typically, functions with a concrete dependency should be kept small and avoid extensive business logic. They ideally function as accessors, providing a loosely coupled abstraction layer over concrete application state.

Testing

Trait mocking with Unimock

The whole point of entrait is to provide inversion of control, so that alternative dependency implementations can be used when unit testing function bodies. While test code can contain manual trait implementations, the most ergonomic way to test is to use a mocking library, which provides more features with less code.

Entrait works best together with unimock, as these two crates have been designed from the start with each other in mind.

Unimock exports a single mock struct which can be passed as argument to every function that accept a generic deps parameter (given that entrait is used with unimock support everywhere). To enable mocking of entraited functions, they get reified and defined as a type called Fn inside a module with the same identifier as the function: entraited_function::Fn.

Unimock support is enabled by passing the unimock option to entrait (#[entrait(Foo, unimock)]), or turning on the unimock feature, which makes all entraited functions mockable, even in upstream crates.

#[entrait(Foo)]
fn foo<D>(_: &D) -> i32 {
    unimplemented!()
}
#[entrait(Bar)]
fn bar<D>(_: &D) -> i32 {
    unimplemented!()
}

fn my_func(deps: &(impl Foo + Bar)) -> i32 {
    deps.foo() + deps.bar()
}

let mocked_deps = unimock::mock([
    foo::Fn.each_call(matching!()).returns(40).in_any_order(),
    bar::Fn.each_call(matching!()).returns(2).in_any_order(),
]);

assert_eq!(42, my_func(&mocked_deps));

Deep integration testing with unimock

Entrait with unimock supports un-mocking. This means that the test environment can be partially mocked!

#[entrait(SayHello)]
fn say_hello(deps: &impl FetchPlanetName, planet_id: u32) -> Result<String, ()> {
    Ok(format!("Hello {}!", deps.fetch_planet_name(planet_id)?))
}

#[entrait(FetchPlanetName)]
fn fetch_planet_name(deps: &impl FetchPlanet, planet_id: u32) -> Result<String, ()> {
    let planet = deps.fetch_planet(planet_id)?;
    Ok(planet.name)
}

pub struct Planet {
    name: String
}

#[entrait(FetchPlanet)]
fn fetch_planet(deps: &(), planet_id: u32) -> Result<Planet, ()> {
    unimplemented!("This doc test has no access to a database :(")
}

let hello_string = say_hello(
    &unimock::spy([
        fetch_planet::Fn
            .each_call(matching!(_))
            .answers(|_| Ok(Planet {
                name: "World".to_string(),
            }))
            .in_any_order(),
    ]),
    123456,
).unwrap();

assert_eq!("Hello World!", hello_string);

This example used unimock::spy to create a mocker that works mostly like Impl, except that the call graph can be short-circuited at arbitrary, run-time configurable points. The example code goes through three layers (say_hello => fetch_planet_name => fetch_planet), and only the deepest one gets mocked out.

Alternative mocking: Mockall

If you instead wish to use a more established mocking crate, there is also support for mockall. Note that mockall has some limitations. Multiple trait bounds are not supported, and deep tests will not work. Also, mockall tends to generate a lot of code, often an order of magnitude more than unimock.

Enabling mockall is done using the mockall entrait option. There is no cargo feature to turn this on implicitly, because mockall doesn't work well when it's re-exported through another crate.

#[entrait(Foo, mockall)]
fn foo<D>(_: &D) -> u32 {
    unimplemented!()
}

fn my_func(deps: &impl Foo) -> u32 {
    deps.foo()
}

fn main() {
    let mut deps = MockFoo::new();
    deps.expect_foo().returning(|| 42);
    assert_eq!(42, my_func(&deps));
}

Modular applications consisting of several crates

A common technique for Rust application development is to divide them into multiple crates. Entrait does its best to provide great support for this kind of architecture. This would be very trivial to do and wouldn't even be worth mentioning here if it wasn't for concrete deps.

Further up, concrete dependency was mentioned as leaves of a depdendency tree. Let's imagine we have an app built from two crates: A main which depends on a lib:

mod lib {
    //! lib.rs - pretend this is a separate crate
    pub struct LibConfig {
        pub foo: String,
    }

    #[entrait_export(pub GetFoo)]
    fn get_foo(config: &LibConfig) -> &str {
        &config.foo
    }

    #[entrait_export(pub LibFunction)]
    fn lib_function(deps: &impl GetFoo) {
        let foo = deps.get_foo();
    }
}

// main.rs
struct App {
    lib_config: lib::LibConfig,
}

fn main() {
    use entrait::*;

    let app = Impl::new(App {
        lib_config: lib::LibConfig {
            foo: "value".to_string(),
        }
    });

    use lib::LibFunction;
    app.lib_function();
}

How can this be made to work at all? Let's deconstruct what is happening:

1. The library defines it's own configuration: LibConfig.
2. It defines a leaf dependency to get access to some property: GetFoo.
3. All things which implement GetFoo may call lib_function.
4. The main crate defines an App, which contains LibConfig.
5. The app has the type Impl<App>, which means it can call entraited functions.
6. Calling LibFunction requires the caller to implement GetFoo.
7. GetFoo is somehow only implemented for Impl<LibConfig>, not Impl<App>.

The clue to get around this problem lies in trait delegation. Trait delegation is an implementation of a trait that contains no logic, but just forwards each method to another implementation of the same trait, through some kind of indirection.

A leaf dependency is a trait which gets delegated from Impl<T> to T, conditional on where T: Trait.

In our example, GetFoo is automatically implemented for Impl<T> where T: GetFoo and for LibConfig. This means it is definitely implemented for Impl<LibConfig>. To make the trait implemented for Impl<App>, we only have to implement it for App. That impl will be another delegation, to LibConfig:

impl GetFoo for App {
    fn get_foo(&self) -> &str {
        self.lib_config.get_foo()
    }
}

Using entrait with a trait

An alternative way to achieve something similar to the above is to use the entrait macro directly on a trait.

A typical use case for this is to put core abstractions in some "core" crate, letting other libraries use those core abstractions as dependencies.

// core_crate
#[entrait]
trait System {
    fn current_time(&self) -> u128;
}

// lib_crate
#[entrait(ComputeSomething)]
fn compute_something(deps: &impl System) {
    let system_time = deps.current_time();
    // do something with the time...
}

// main.rs
struct App;
impl System for App {
    fn current_time(&self) -> u128 {
        std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .unwrap()
            .as_millis()
    }
}

Impl::new(App).compute_something();

This is similar to defining a leaf dependency for a concrete type, only in this case, core_crate really has no type available to use. We know that System eventually has to be implemented for the application type, and that can happen in the main crate.

The reason that the #[entrait] attribute has to be present in core_crate, is that it needs to define a blanket implementation for Impl<T> (as well as mocks), and those need to live in the same crate that defined the trait. If not, this would have violated the orphan rule.

(NB: This example's purpose is to demonstrate entrait, not to be a guide on how to deal with system time. It should contain some ideas for how to mock time, though!)

dyn leaf dependencies

Most application configuration is data-based, but some applications also require dynamically changing implementation. Rust can express dynamically changing implementation using dyn Trait.

Entrait supports this for leaf dependencies. As an example, imagine a trait for abstracting away how to read some config:

#[entrait]
trait ReadConfig {
    fn read_config(&self) -> String;
}

How to read this config could be different depending on which platform the application runs on, and the application would like to support several implementations of this trait.

The problem we face here, is that the generated delegation code for Impl<T> will require that T: ReadConfig. In other words, there can only be one implementation of this trait per T. This means that we would have to create a separate application type for each way of reading config. This doesn't scale well, and leads to combinatorial type explosion when we have several such traits to implement. Additionally, this approach would lead to several copies of our entire generic application written with the entrait pattern, because it would need to get monomorphized for each type.

The solution is to borrow a dyn reference to this trait. Since Impl<T> delegations does not know anything about the concrete T, we need to describe in an abstracted way how to borrow it. For this, Rust has the Borrow trait.

Instead of generating a T: ReadConfig bound, we can generate a T: Borrow<dyn ReadConfig> bound, by using delegate_by = Borrow.

#[entrait(delegate_by = Borrow)]
trait ReadConfig: 'static {
    fn read_config(&self) -> String;
}

#[entrait(DoSomething)]
fn do_something(deps: &impl ReadConfig) {
    let config = deps.read_config();
}

struct App {
    read_config: Box<dyn ReadConfig + Sync>,
};

impl ReadConfig for String {
    fn read_config(&self) -> String {
        self.clone()
    }
}

impl std::borrow::Borrow<dyn ReadConfig> for App {
    fn borrow(&self) -> &dyn ReadConfig {
        self.read_config.as_ref()
    }
}

let app = Impl::new(App {
    read_config: Box::new("hard-coded!".to_string()),
});
app.do_something();

Options and features

Trait visibility

by default, entrait generates a trait that is module-private (no visibility keyword). To change this, just put a visibility specifier before the trait name:

use entrait::*;
#[entrait(pub Foo)]   // <-- public trait
fn foo<D>(deps: &D) { // <-- private function
}

async support

Since Rust at the time of writing does not natively support async methods in traits, you may opt in to having #[async_trait] generated for your trait. Enable the async-trait cargo feature and pass the async_trait option like this:

#[entrait(Foo, async_trait)]
async fn foo<D>(deps: &D) {
}

This is designed to be forwards compatible with static async fn in traits. When that day comes, you should be able to just remove that option and get a proper zero-cost future.

There is a cargo feature to automatically apply #[async_trait] to every generated async trait: use-async-trait.

Zero-cost async inversion of control - preview mode

Entrait has experimental support for zero-cost futures. A nightly Rust compiler is needed for this feature.

The entrait option is called associated_future, and depends on generic_associated_types and type_alias_impl_trait. This feature generates an associated future inside the trait, and the implementations use impl Trait syntax to infer the resulting type of the future:

#![feature(generic_associated_types)]
#![feature(type_alias_impl_trait)]

use entrait::*;

#[entrait(Foo, associated_future)]
async fn foo<D>(deps: &D) {
}

There is a feature for turning this on everywhere: use-associated-future.

Integrating with other fn-targeting macros, and no_deps

Some macros are used to transform the body of a function, or generate a body from scratch. For example, we can use feignhttp to generate an HTTP client. Entrait will try as best as it can to co-exist with macros like these. Since entrait is a higher-level macro that does not touch fn bodies (it does not even try to parse them), entrait should be processed after, which means it should be placed before lower level macros. Example:

#[entrait(FetchThing, no_deps)]
#[feignhttp::get("https://my.api.org/api/{param}")]
async fn fetch_thing(#[path] param: String) -> feignhttp::Result<String> {}

Here we had to use the no_deps entrait option. This is used to tell entrait that the function does not have a deps parameter as its first input. Instead, all the function's inputs get promoted to the generated trait method.

Conditional compilation of mocks

Most often, you will only need to generate mock implementations for test code, and skip this for production code. A notable exception to this is when building libraries. When an application consists of several crates, downstream crates would likely want to mock out functionality from libraries.

Entrait calls this exporting, and it unconditionally turns on autogeneration of mock implementations:

#[entrait_export(pub Bar)]
fn bar(deps: &()) {}

or

#[entrait(pub Foo, export)]
fn foo(deps: &()) {}

It is also possible to reduce noise by doing use entrait::entrait_export as entrait.

Feature overview

Feature	Implies	Description
unimock		Adds the [unimock] dependency, and turns on Unimock implementations for all traits.
use-async-trait	async_trait	Automatically applies the [async_trait] macro to async trait methods.
use-associated-future		Automatically transforms the return type of async trait methods into an associated future by using type-alias-impl-trait syntax. Requires a nightly compiler.
async-trait		Pulls in the [async_trait] optional dependency, enabling the async_trait entrait option (macro parameter).

"Philosophy"

The entrait crate is central to the entrait pattern, an opinionated yet flexible way to build testable applications/business logic.

To understand the entrait model and how to achieve Dependency Injection (DI) with it, we can compare it with a more widely used and classical alternative pattern: Object-Oriented DI.

In object-oriented DI, each named dependency is a separate object instance. Each dependency exports a set of public methods, and internally points to a set of private dependencies. A working application is built by fully instantiating such an object graph of interconnected dependencies.

Entrait was built to address two drawbacks inherent to this design:

• Representing a graph of objects (even if acyclic) in Rust usually requires reference counting/heap allocation.
• Each "dependency" abstraction often contains a lot of different functionality. As an example, consider DDD-based applications consisting of DomainServices. There will typically be one such class per domain object, with a lot of methods in each. This results in dependency graphs with fewer nodes overall, but the number of possible call graphs is much larger. A common problem with this is that the actual dependencies—the functions actually getting called—are encapsulated and hidden away from public interfaces. To construct valid dependency mocks in unit tests, a developer will have to read through full function bodies instead of looking at signatures.

entrait solves this by:

• Representing dependencies as traits instead of types, automatically profiting from Rust's builtin zero-cost abstraction tool.
• Having each dependency do only one thing, by abstracting over functions instead of modules. This is possible because we do not pay anything extra for having more detailed dependency graphs.

Limitations

This section lists known limitations of entrait:

Cyclic dependency graphs

Cyclic dependency graphs are impossible with entrait. In fact, this is not a limit of entrait itself, but with Rust's trait solver. It is not able to prove that a type implements a trait if it needs to prove that it does in order to prove it.

While this is a limitation, it is not necessarily a bad one. One might say that a layered application architecture should never contain cycles. If you do need recursive algorithms, you could model this as utility functions outside of the entraited APIs of the application.