The Run Loop

Very often, batching similar work has benefits. Web browsers do something quite similar by batching changes to the DOM.

Consider the following HTML snippet:

<div id="foo"></div>
<div id="bar"></div>
<div id="baz"></div>

and executing the following code:

foo.style.height = '500px' // write
foo.offsetHeight // read (recalculate style, layout, expensive!)

bar.style.height = '400px' // write
bar.offsetHeight // read (recalculate style, layout, expensive!)

baz.style.height = '200px' // write
baz.offsetHeight // read (recalculate style, layout, expensive!)

In this example, the sequence of code forced the browser to recalculate style, and relayout after each step. However, if we were able to batch similar jobs together, the browser would have only needed to recalculate the style and layout once.

foo.style.height = '500px' // write
bar.style.height = '400px' // write
baz.style.height = '200px' // write

foo.offsetHeight // read (recalculate style, layout, expensive!)
bar.offsetHeight // read (fast since style and layout is already known)
baz.offsetHeight // read (fast since style and layout is already known)

Interestingly, this pattern holds true for many other types of work. Essentially, batching similar work allows for better pipelining, and further optimization.

Let's look at a similar example that is optimized in Ember, starting with a User object:

var User = Ember.Object.extend({
  firstName: null,
  lastName: null,
  fullName: Ember.computed('firstName', 'lastName', function() {
    return `${this.get('firstName')} ${this.get('lastName')}`;
  })
});

and a template to display its attributes:

{{firstName}}
{{fullName}}

If we execute the following code without the run loop:

var user = User.create({ firstName: 'Tom', lastName: 'Huda' });
user.set('firstName', 'Yehuda');
// {{firstName}} and {{fullName}} are updated

user.set('lastName', 'Katz');
// {{lastName}} and {{fullName}} are updated

We see that the browser will rerender the template twice.

However, if we have the run loop in the above code, the browser will only rerender the template once the attributes have all been set.

var user = User.create({ firstName: 'Tom', lastName: 'Huda' });
user.set('firstName', 'Yehuda');
user.set('lastName', 'Katz');
user.set('firstName', 'Tom');
user.set('lastName', 'Huda');

In the above example with the run loop, since the user's attributes end up at the same values as before execution, the template will not even rerender!

It is of course possible to optimize these scenarios on a case-by-case basis, but getting them for free is much nicer. Using the run loop, we can apply these classes of optimizations not only for each scenario, but holistically app-wide.

As mentioned earlier, we schedule work (in the form of function invocations) on queues, and these queues are processed to completion in priority order.

What are the queues, and what is their priority order?

Ember.run.queues
// => ["sync", "actions", "routerTransitions", "render", "afterRender", "destroy"]

Because the priority is first to last, the "sync" queue has higher priority than the "render" or "destroy" queue.

• The sync queue contains binding synchronization jobs
• The actions queue is the general work queue and will typically contain scheduled tasks e.g. promises
• The routerTransitions queue contains transition jobs in the router
• The render queue contains jobs meant for rendering, these will typically update the DOM
• The afterRender queue contains jobs meant to be run after all previously scheduled render tasks are complete. This is often good for 3rd-party DOM manipulation libraries, that should only be run after an entire tree of DOM has been updated
• The destroy queue contains jobs to finish the teardown of objects other jobs have scheduled to destroy

The algorithm works this way:

1. Let the highest priority queue with pending jobs be: CURRENT_QUEUE, if there are no queues with pending jobs the run loop is complete
2. Let a new temporary queue be defined as WORK_QUEUE
3. Move jobs from CURRENT_QUEUE into WORK_QUEUE
4. Process all the jobs sequentially in WORK_QUEUE
5. Return to Step 1

Rather than writing the higher level app code that internally invokes the various run loop scheduling functions, we have stripped away the covers, and shown the raw run-loop interactions.

Working with this API directly is not common in most Ember apps, but understanding this example will help you to understand the run-loops algorithm, which will make you a better Ember developer.

You should begin a run loop when the callback fires.

The Ember.run method can be used to create a runloop. In this example, jQuery and Ember.run are used to handle a click event and run some Ember code.

This example uses the => function syntax, which is a new ES2015 syntax for callback functions that provides a lexical this. If this syntax is new, think of it as a function that has the same this as the context it is defined in.

$('a').click(() => {
  Ember.run(() => {  // begin loop
    // Code that results in jobs being scheduled goes here
  }); // end loop, jobs are flushed and executed
});

As mentioned above, you should wrap any non-Ember async callbacks in Ember.run. If you don't, Ember will try to approximate a beginning and end for you. Consider the following callback:

$('a').click(() => {
  console.log('Doing things...');

  Ember.run.schedule('actions', () => {
    // Do more things
  });
});

The runloop API calls that schedule work i.e. run.schedule, run.scheduleOnce, run.once have the property that they will approximate a runloop for you if one does not already exist. These automatically created runloops we call autoruns.

Here is some pseudocode to describe what happens using the example above:

$('a').click(() => {
  // 1. autoruns do not change the execution of arbitrary code in a callback.
  //    This code is still run when this callback is executed and will not be
  //    scheduled on an autorun.
  console.log('Doing things...');

  Ember.run.schedule('actions', () => {
    // 2. schedule notices that there is no currently available runloop so it
    //    creates one. It schedules it to close and flush queues on the next
    //    turn of the JS event loop.
    if (! Ember.run.hasOpenRunloop()) {
      Ember.run.start();
      nextTick(() => {
        Ember.run.end()
      }, 0);
    }

    // 3. There is now a runloop available so schedule adds its item to the
    //    given queue
    Ember.run.schedule('actions', () => {
      // Do more things
    });

  });

  // 4. This schedule sees the autorun created by schedule above as an available
  //    runloop and adds its item to the given queue.
  Ember.run.schedule('afterRender', () => {
    // Do yet more things
  });
});

Although autoruns are convenient, they are suboptimal. The current JS frame is allowed to end before the run loop is flushed, which sometimes means the browser will take the opportunity to do other things, like garbage collection. GC running in between data changing and DOM rerendering can cause visual lag and should be minimized.

Relying on autoruns is not a rigorous or efficient way to use the runloop. Wrapping event handlers manually is preferred.

When your application is in testing mode then Ember will throw an error if you try to schedule work without an available runloop.

Autoruns are disabled in testing for several reasons:

1. Autoruns are Embers way of not punishing you in production if you forget to open a runloop before you schedule callbacks on it. While this is useful in production, these are still situations that should be revealed in testing to help you find and fix them.
2. Some of Ember's test helpers are promises that wait for the run loop to empty before resolving. If your application has code that runs outside a runloop, these will resolve too early and give erroneous test failures which are difficult to find. Disabling autoruns help you identify these scenarios and helps both your testing and your application!

Check out the Ember.run API documentation, as well as the Backburner library that powers the run loop.