functional reactive programming on the web

software engineering course at university of texas.

i read the flapjax paper for my software engineering class at UT. it presents a language and framework for functional reactive programming in javascript that’s suitable for the web. i think it has some really powerful ideas that could make it a better design pattern than mvc (or mv*)

aside from sounding like academic bullhankey, functional reactive programming is a complicated concept. so let’s begin with an over-simplified example to motivate the problem.


pretend you’re writing an e-mail client web app (like gmail). in the main view on the left you’ve got a link to the inbox with the number of unread messages, and on the right you’ve got the list of messages with the unread ones marked in bold. if the user marks a message as read, you need to do two things: un-embolden the text, and change the count of unread messages. so you write a function called messageWasRead() to do so.

this seems pretty simple, but as your application grows it gets tricky to remember all the little things that need to happen for any given action. suddenly that messageWasRead() function is 700 brittle lines long and an absolute monolith that you can’t reuse in any other context.

we’re talking about application state here. the message list and unread count are just two different views of that same state, which, in large apps, becomes large, unruly (unsavory even), and impossible to reason about comprehensively. since we’re simple creatures, developers have come up with patterns for managing state, the most popular of which currently seems to be mv*. reactive programming is an alternative that’s been around for a while but is gaining popularity.

main concepts

flapjax elegantly distills the world of application programming down into two concepts (which is one less concept than mvc, ipso facto, functional reactive is more elegant </joking>):


behaviors are values that may change over time, like your age. if you set a regular variable to the value of your age, then it’s no longer correct after your birthday. behaviors on the other hand update over time, so a behavior representing your age will return your current age regardless of when you check it’s value.

behaviors are kind of like cells in a spreadsheet. if you’ve got a column of numbers and a cell that calculates the sum, updating one of the rows will also update the sum. behaviors similarly update themselves if any of the values upon which they depend change.

event streams

user interactions like clicks can’t be modeled with behaviors since they’re discrete in time. it doesn’t make sense to “check the value of a click,” so flapjax introduces event streams. event streams are instantiated with a function that takes some event data and returns a value. each time the event occurs, the function runs. the event stream can then be treated like a behavior whose value is the item returned the last time the event fired.

building applications

if you’re used to more common approaches of app development, it’s probably hard to see how we could use the above to actually build anything. there are numerous (higher quality) examples, but consider this short example from the paper: a timer that displays the number of seconds passed, and a button to reset it. the strait-forward implementation might look something like this:

      var timerID = null;
      var elapsedTime = 0;

      function doEverySecond() {
        elapsedTime += 1;
        document.getElementById("curTime").innerHTML = elapsedTime;

      function startTimer() {
        timerId = setInterval("doEverySecond()", 1000);

      function resetElapsed() {
        elapsedTime = 0;
  <body onload="startTimer()">
    <input id="reset" type="button" value="Reset" onclick="resetElapsed()"/>
    <div id="curTime"></div>

this is about as basic of a javascript app as it gets, but essentially on page load we’re setting an interval every second that does two things:

the problems may not be apparent in such a trivial example, but remember what we said about state – as the application grows it gets hard to manage. we’ve left ourselves no way of tracking updates to the elapsedTime variable (state changes) without re-writing doEverySecond and resetElapsed.

the functional reactive approach looks like this:

    <script src=''></script>
      function loader() {
        var nowB = timerB(1000);
        var resetEl = document.getElementById('reset');

        var clickTimeB = $E(resetEl, 'click')

        var elapsedB = liftB(function(now, start) {
          return Math.floor((now - start) / 1000);
        }, nowB, clickTimeB);

        insertDomB(elapsedB, 'curTime');
  <body onload='loader()'>
    <input id='reset' type='button' value='Reset'/>
    <div id='curTime'></div>

though this approach looks like just as much code, we have isolated state in the behavior variables. as long as we have a reference to elapsedB, we not only know the amount of time elapsed, but we can also observe updates whenever it changes. if want to react to changes, we don’t have to rewrite any of our existing code, we just need elapsedB.

a more involved example

none of this caught my eye until i saw a functional reactive implementation of a draggable UI component. drag and drop is easy to get wrong; i know because my first attempt was a train wreck. the authors do a much better job of explaining their code in section 2.4 of their paper than i could, but the approach is both succinct and elegant. it’s interesting to compare the level of complexity to jquery-ui’s draggable implementation, or even jeremy kahn’s relatively simple library, even if it is apples-to-oranges since those address a much wider array of use cases.

is it worth it?

if you’re thinking that all of this is solved by frontend mv* frameworks, you’re largely correct. the majority of webapps map pretty well to crud operations on restful api’s. mv* crushes in those applications because it makes sense to have a model class describing each resource. for example if you’ve got an api for fetching users, you have a User class that inherits from the base Model class, and as long as you’re rendering your views from a shared instance of that model, everything works well.

reactive programming upholds the same solid principles of isolating state to a single point of truth, but it applies them at a much finer granularity. the obvious benefit is it addresses situations where mv* patterns couldn’t help you (such as drag and drop or data models more complex than rest).

the two major downsides i see to reactive programming are:

there’s a lot of folks making libraries that implement various levels of functional reactive programming on the web. here are a few:

i’ve got a fair amount of experience with knockout, and i’ve really enjoyed working with it. i think it scales well with the messier parts of big apps, such as asynchronous input validation.

blog comments powered by Disqus
© aaron stacy 2013, all rights reserved