maybe-async

Why bother writing similar code twice for blocking and async code?

When implementing both sync and async versions of API in a crate, most API of the two version are almost the same except for some async/await keyword.

maybe-async help unifying async and sync implementation by procedural macro.

• Write async code with normal async, await, and let maybe_async handles those async and await when you need a blocking code.
• Switch between sync and async by toggling is_sync feature gate in Cargo.toml.
• use must_be_async and must_be_sync to keep code in specified version
• use async_impl and sync_impl to only compile code block on specified version
• A handy macro to unify unit test code is also provided.

These procedural macros can be applied to the following codes:

• trait item declaration
• trait implementation
• function definition
• struct definition

RECOMMENDATION: Enable resolver ver2 in your crate, which is introduced in Rust 1.51. If not, two crates in dependency with conflict version (one async and another blocking) can fail compilation.

Motivation

The async/await language feature alters the async world of rust. Comparing with the map/and_then style, now the async code really resembles sync version code.

In many crates, the async and sync version of crates shares the same API, but the minor difference that all async code must be awaited prevent the unification of async and sync code. In other words, we are forced to write an async and a sync implementation respectively.

Macros in Detail

maybe-async offers 4 set of attribute macros: maybe_async, sync_impl/async_impl, must_be_sync/must_be_async, and test.

To use maybe-async, we must know which block of codes is only used on blocking implementation, and which on async. These two implementation should share the same function signatures except for async/await keywords, and use sync_impl and async_impl to mark these implementation.

Use maybe_async macro on codes that share the same API on both async and blocking code except for async/await keywords. And use feature gate is_sync in Cargo.toml to toggle between async and blocking code.

• maybe_async

  Offers a unified feature gate to provide sync and async conversion on demand by feature gate is_sync, with async first policy.

  Want to keep async code? add maybe_async in dependencies with default features, which means maybe_async is the same as must_be_async:

  [dependencies]
  maybe_async = "0.2"
  

  Want to convert async code to sync? Add maybe_async to dependencies with an is_sync feature gate. In this way, maybe_async is the same as must_be_sync:

  [dependencies]
  maybe_async = { version = "0.2", features = ["is_sync"] }
  

  There are three usage variants for maybe_async attribute usage:

  • #[maybe_async] or #[maybe_async(Send)]

    In this mode, #[async_trait::async_trait] is added to trait declarations and trait implementations to support async fn in traits.

  • #[maybe_async(?Send)]

    Not all async traits need futures that are dyn Future + Send. In this mode, #[async_trait::async_trait(?Send)] is added to trait declarations and trait implementations, to avoid having "Send" and "Sync" bounds placed on the async trait methods.

  • #[maybe_async(AFIT)]

    AFIT is acronym for async function in trait, stabilized from rust 1.74

  For compatibility reasons, the async fn in traits is supported via a verbose AFIT flag. This will become the default mode for the next major release.

• must_be_async

  Keep async.

  There are three usage variants for must_be_async attribute usage:

  • #[must_be_async] or #[must_be_async(Send)]
  • #[must_be_async(?Send)]
  • #[must_be_async(AFIT)]

• must_be_sync

  Convert to sync code. Convert the async code into sync code by removing all async move, async and await keyword

• sync_impl

  A sync implementation should compile on blocking implementation and must simply disappear when we want async version.

  Although most of the API are almost the same, there definitely come to a point when the async and sync version should differ greatly. For example, a MongoDB client may use the same API for async and sync version, but the code to actually send reqeust are quite different.

  Here, we can use sync_impl to mark a synchronous implementation, and a sync implementation should disappear when we want async version.

• async_impl

  An async implementation should on compile on async implementation and must simply disappear when we want sync version.

  There are three usage variants for async_impl attribute usage:

  • #[async_impl] or #[async_impl(Send)]
  • #[async_impl(?Send)]
  • #[async_impl(AFIT)]

• test

  Handy macro to unify async and sync unit and e2e test code.

  You can specify the condition to compile to sync test code and also the conditions to compile to async test code with given test macro, e.x. tokio::test, async_std::test, etc. When only sync condition is specified,the test code only compiles when sync condition is met.

  # #[maybe_async::maybe_async]
  # async fn async_fn() -> bool {
  #    true
  # }
  
  ##[maybe_async::test(
      feature="is_sync",
      async(
          all(not(feature="is_sync"), feature="async_std"),
          async_std::test
      ),
      async(
          all(not(feature="is_sync"), feature="tokio"),
          tokio::test
      )
  )]
  async fn test_async_fn() {
      let res = async_fn().await;
      assert_eq!(res, true);
  }
  

What's Under the Hook

maybe-async compiles your code in different way with the is_sync feature gate. It removes all await and async keywords in your code under maybe_async macro and conditionally compiles codes under async_impl and sync_impl.

Here is a detailed example on what's going on whe the is_sync feature gate set or not.

#[maybe_async::maybe_async(AFIT)]
trait A {
    async fn async_fn_name() -> Result<(), ()> {
        Ok(())
    }
    fn sync_fn_name() -> Result<(), ()> {
        Ok(())
    }
}

struct Foo;

#[maybe_async::maybe_async(AFIT)]
impl A for Foo {
    async fn async_fn_name() -> Result<(), ()> {
        Ok(())
    }
    fn sync_fn_name() -> Result<(), ()> {
        Ok(())
    }
}

#[maybe_async::maybe_async]
async fn maybe_async_fn() -> Result<(), ()> {
    let a = Foo::async_fn_name().await?;

    let b = Foo::sync_fn_name()?;
    Ok(())
}

When maybe-async feature gate is_sync is NOT set, the generated code is async code:

// Compiled code when `is_sync` is toggled off.
trait A {
    async fn maybe_async_fn_name() -> Result<(), ()> {
        Ok(())
    }
    fn sync_fn_name() -> Result<(), ()> {
        Ok(())
    }
}

struct Foo;

impl A for Foo {
    async fn maybe_async_fn_name() -> Result<(), ()> {
        Ok(())
    }
    fn sync_fn_name() -> Result<(), ()> {
        Ok(())
    }
}

async fn maybe_async_fn() -> Result<(), ()> {
    let a = Foo::maybe_async_fn_name().await?;
    let b = Foo::sync_fn_name()?;
    Ok(())
}

When maybe-async feature gate is_sync is set, all async keyword is ignored and yields a sync version code:

// Compiled code when `is_sync` is toggled on.
trait A {
    fn maybe_async_fn_name() -> Result<(), ()> {
        Ok(())
    }
    fn sync_fn_name() -> Result<(), ()> {
        Ok(())
    }
}

struct Foo;

impl A for Foo {
    fn maybe_async_fn_name() -> Result<(), ()> {
        Ok(())
    }
    fn sync_fn_name() -> Result<(), ()> {
        Ok(())
    }
}

fn maybe_async_fn() -> Result<(), ()> {
    let a = Foo::maybe_async_fn_name()?;
    let b = Foo::sync_fn_name()?;
    Ok(())
}

Examples

rust client for services

When implementing rust client for any services, like awz3. The higher level API of async and sync version is almost the same, such as creating or deleting a bucket, retrieving an object, etc.

The example service_client is a proof of concept that maybe_async can actually free us from writing almost the same code for sync and async. We can toggle between a sync AWZ3 client and async one by is_sync feature gate when we add maybe-async to dependency.

License

MIT