This page represents a complete vision for where we want async to go. This vision is what we believe to be the best way to achieve the experiences that we want async to provide.
Note that while a lot of the steps needed are fairly clear, several of them also have significant unknowns or points of controversy. We have attempted to highlight those and expect to be working through those points as we go.
- 🌈 -- Implemented and stable
- 🌞 -- Everything is looking good
- 🌤️ -- Still some stuff to figure out, but unlikely to see major changes in the design
- 🌥️ -- Got one or two solid leads, but still have to figure out if it will work
- 🌧️ -- No clear path yet, this may not even be a good idea
- 🌤️ If you know sync Rust, getting started in Async Rust is straightforward ([more][async_fn_everywhere])
- 🌤️ Mostly, you change
async fn, add some calls to await, and change over to other parts of the stdlib, though supporting
dyn Traitrequires making some choices, particularly in a no-std environment
- 🌤️ It still has that "if it compiles, it generally works, and it runs pretty darn fast" feeling
- 🌤️ Destructors and cleanup also work the same way as in sync Rust, thanks to
- 🌤️ No need to write poll functions or to interact with pin except in quite specialized scenarios
- 🌤️ Mostly, you change
- 🌤️ High-quality documentation and tutorials helps you to get started and learn the ropes
- 🌤️ The docs also identify common patterns for structuring your async programs and their advantages and disadvantages
- 🌥️ Tooling and debugger integration gives insight into the behavior of your program
- 🌥️ Easy to get a snapshot of overall acitivity (e.g. to find out what tasks or exist or why a task is blocked)
- 🌥️ Easy to see aggregate performance trends over time (e.g., number of active connections, waiting connections, etc)
- 🌥️ Easy to profile things in terms of your async tasks (e.g., to get a flamegraph of a specific connection)
- 🌥️ Variety of high-quality runtimes available in cargo, and it's easy to change between them:
- 🌧️ When you use things from the standard library, they work across runtimes automatically
- 🌥️ There are standardized, foundational traits for common operations like I/O, spawning tasks, timers
- 🌥️ Hierarchical scopes allow you to easily spawn parallel and concurrent tasks
- 🌥️ These can reference borrowed data, enabling easy parallel processing of async iterators (think "async rayon")
- 🌥️ Cancellation works well and without surprises
- 🌥️ When cancellation is requested, it propagates to subtasks within a scope
- 🌧️ I/O operations and the like begin to fail, so that cancellation is automatic and flows through familiar error paths
- 🌥️ If desired, you can "opt-in" to synchronous cancellation, in which case any await becomes a cancellation point. This allows your
async fnto be used with
selectwithout spawning a task.
The full set of "submitted" shiny future stories have been moved here.