Improving spotify-tui: going async

In spotify-tui versions 0.15.0 and prior there was a major issue that caused the UI to freeze and not respond to user key events.

This post attempts to outline the problem and the solution.

Architecture

The architecture of spotify-tui is one big loop that draws to the terminal, receives new key events and performs network calls.

The freezing behaviour lies somewhere in here.

Drawing the UI

We draw the terminal UI on each loop. Could something in here involve a heavy CPU bound operation?

There are plenty of data transformations happening on each loop. But from what I can tell, nothing so serious to clog up the CPU. So unlikely that the rendering loop is causing this freezing problem.

User key events

These events are emitted from their own thread and wonโ€™t block the main UI rendering loop. So, again, the freezing problem was unlikely to be here.

Blocking network requests

So the cause of the freezing behaviour is now obvious - the network calls to the Spotify API were blocking the main UI loop ๐Ÿ˜„.

When the internet connection is fast itโ€™s not noticeable, but this is a different story on bad connections.

Solutions

I had thought of two main solutions to this:

  1. Refactor the underlying Rust Spotify API wrapper rspotify to use the new async/await syntax and update spotify-tui to use the tokio async runtime (with prayers that this would magically solve the issue).
  2. Within spotify-tui, use the main process to handle the UI rendering loop and implement a network thread to handle the network calls.

I new that the author of rspotify had plans to migrate to using async/await anyway so I started by contributing to that effort.

Fix #1 - No dice

After upgrading spotify-tui to the new async/await version of rspotify I simply kept the spotify-tui architecture as it was and followed the compiler errors telling me I needed to await the future.

Rather expectedly, using async/await alone changed nothing with regards to the main freezing issue.

Even with an async runtime, we were still awaiting within the main rendering loop and so pausing execution until the future was resolved.

Thus, it was time to work on solution 2 from above.

Fix #2 - Major architectural change

Now we need to spawn a new thread that can receive messages to perform network calls, thereby not blocking the rendering loop. We also need a way of sharing state between these threads.

The Rust book has an excellent section on shared-state concurrency. To safely share read data in the UI thread and mutate it from the network thread, we must wrap the shared state within a Mutex.

And to share the Mutex across threads we must wrap it in a thread-safe reference-counting pointer Arc. You can also read more on atomic reference counting here.

Most of the state within spotify-tui is contained in the App struct.

So, in effect, I needed to do this

let app = Arc::new(Mutex::new(App::new()));

Now we can safely access App from the main process that implements the UI rendering loop, and from the network thread

let app = Arc::new(Mutex::new(App::new()));

// Clone the Arc so we can acquire a lock within the UI loop below
let cloned_app = Arc::clone(&app);
std::thread::spawn(move || {
  // Start a tokio async runtime in another thread and give it a
  // reference to the App (which is wrapped in Arc and Mutex).
  // We will explore what this is doing below
  start_tokio(&app);
});


loop {
  // Acquire a lock on the Mutex so we can access data contained in the app.
  let mut app = cloned_app.lock();

  terminal.draw(|| {
    // Draw the terminal UI and use data contained in app.
  });

  match events.next()? {
    // Match a key press event
  }

}

Message passing

Since we receive user input from the main rendering loop (see events.next() above) we need a way to send a message to the network thread so it can perform the network action.

To do this, we create a channel, pass the sending end into App and the receiving end into the network thread.

let (tx, rx) = std::sync::mpsc::channel();

let app = Arc::new(Mutex::new(App::new(tx)));

let cloned_app = Arc::clone(&app);
std::thread::spawn(move || {
// Send the receiving end of the channel into the network thread
  start_tokio(&app, rx);
});

loop {
  ...
}

Now we can receive messages in our network thread.

But what do we send from the UI loop? And how do we handle the message in the network thread?

Enums

This is a perfect usecase for Enums, which allows us to leverage the might of pattern matching.

pub enum IoEvent {
  GetCurrentPlayback,
  GetAlbum(String),
}

// Send a message to the network thread (this is synchronous)
tx.send(IoEvent::GetCurrentPlayback).unwrap();

And to receive the message we use the recv method on the Receiver. Now we can finally implement start_tokio method that is invoked in the thread above.

This function will attempt to receive a message, handle it via pattern matching, make async network requests, and update the app state.

// Receive a message and handle it
#[tokio::main]
async fn start_tokio(io_rx: std::sync::mpsc::Receiver<IoEvent>) {
  while let Ok(io_event) = io_rx.recv() {
    match io_event {
      IoEvent::GetCurrentPlayback => {
        // Make a network request and await it
        let current_playback = get_current_playback().await;

        // Acquire a lock on the App Mutex and mutate the state
        let mut app = app.lock();
        app.current_playback = current_playback;
      }

      IoEvent::GetAlbum(album_id) => {
        let album = get_album(album_id).await;

        let mut app = app.lock();
        app.current_album = album;
      }
    }
  }
}

Conclusion

Refactoring both rspotify and spotify-tui to use async/await produced a huge number of changes.

On top of this, moving all Spotify API calls into the network thread was took a good deal of effort too.

Without vim macros this surely would have been a carapl tunnel inducing effort ๐Ÿ˜„.

Learnings

  • Rust Futures
  • async/await
  • tokio - the asynchronous runtime
  • How to share state between threads using Arc and Mutex
  • How to send data between threads using channel

Gotchas

In the network thread, we must be careful to only acquire the Mutex lock after the network call is done (after the await), otherwise the lock we acquire in the UI loop will block that loop and so weโ€™d be back to square one.

This could be hard to maintain and so perhaps I should investigate a different solution.

Boilerplate

There is lots more boilerplate in this new architecture - now, all async tasks need to be added to the IoEvent enum so we can synchronously dispatch events from the UI thread and handle events in the network thread.

UX

The user experience is much much better, however, as now the UI wonโ€™t freeze when network calls are slow!

Power of Rust ๐Ÿ’ช

As ever, Rust has made it much easier to handle rather tricky problems - async network calls and shared-state concurrency.

Given the scale of these changes, without the power of the compiler to guide me it would have been so much more difficult!

It was an amazing feeling following the compiler errors, fixing them, and then after about 4 hours of coding finally running spotify-tui for it to work as expected!

This is not to say that this change wonโ€™t introduce new issues - we now need to think of loading states, navigating early before the network has responded etc.

These are familiar problems encountered in all (T)UI applications, and I look forward to solving them in spotify-tui!

Suggestions

You might know a better architecture to the one outlined here, if so Iโ€™d leave to hear from you! Please tweet @AlexKeliris or open an issue on the spotify-tui.