This is a really nice cleanup/update on Knockout. The source code at ~400 lines is very understandable. I've looked at a lot of reactive frameworks, this is great work!
I'm glad you're thinking about how to manage garbage collection. It's tricky with these push-based frameworks.
Have you given consideration to asynchronous vs synchronous reactivity? The advantage of asynchronous is you don't propagate the changes as they are made but instead once all the changes are done. You can avoid redundant computations this way, for example if a bind is dependent on multiple observables that change in the same "turn". And in some ways (worth debating) asynchronous semantics are more understandable than synchronous since unknown code is not being run underneath you while you are in the middle of changing observables.
Yes! There's already a lag-delay primitive, but it's a one-off hack we needed to get things working - more general asynchronous propagation is definitely a requirement.
"z = bind -> x.get() + y.get()" seems nice, but when are the attached listeners removed again?
Typically this is solved with weak listeners, but Javascript doesn't support weak references. I'm worried that this library never removes the listeners, effectively leaking like crazy?
Astute observation - this is actually something we do call out in http://yang.github.io/reactive-coffee/infelicities.html#garb.... As you note, JS doesn't have it as easy as in other platforms, but it's not a fundamental problem and is still reasonably straightforward to address (as described). The fix is slated for this coming release, and we'll need to more thoroughly document/explain what this all means.
Thanks for the explanation in the link. The problem that is described in the link is clearly solvable, however what I'm referring to is more fundamental.
I'm not talking about nested binds, but just a simple bind such as "z = bind -> x.get() + y.get()". This will add listeners to x and y. When will those listeners be removed again (without adding new ones during the recalculation of z's new value in case of a change to x or y)?
It may be that x and y are the basic models in my app, and during my clicking around in the application I create and close many panels, and each panel has bind-operations that add listeners to x and y. Now the panels may be long gone, and will never be used again, but when will the listeners that was added to x and y as part of creating the panels be removed?
Yeah, this is what I'm talking about - a fuller explanation would require more text. :)
The question is what you're doing with z. As long as you're also adding and removing your hypothetical panels in turn with bind, you're good:
div {class: 'container'}, bind -> [
if show.get()
div {class: 'z'}, bind -> x.get() + y.get()
else
div {class: 'nothing'}
]
The question is what happens at the top level or when you want to break out and do your own thing. We don't have the luxury of weak refs in JS, and as a result it's less forgiving if you do "silly" things like create binds that go nowhere and that you don't want to keep. But even with weak references, one can't tell if you added a bind only to subscribe a console.log caller to its changes, save to localStorage, or some other side effect that you do want to keep. In any case, it's incumbent on us to fully document what "silly" means, include simple-to-use API calls to capture and dispose entire subgraphs (for when you want to do your own thing manually), and provide good debugging introspection tools for finding these `bind`s to nowhere, in case you really don't want them lingering around.
We've also been bouncing around ideas for reversing the reference DAG and having cells named by the reversed paths through the DAG, which we may experiment with. You then get to deal with the converse problem of manually needing to hold references to the sinks in the DAG. In any case, we're very open to learning from how things play out in practice, and shape the direction of the library accordingly.
Ahh, you're putting the entire UI in binds. That's a really clever way to handle the lifecycle of the listeners!
bind is "stupid" in the way that it completely recalculates a cells value on any change of a cell it's previous calculation depended on. It doesn't "understand" the calculation in a way to only recalculate the part that was strictly needed to update its value to a change.
I mention this because if the entire UI is in a bind, then on every change the entire UI would be recreated. I worry that this could get prohibitively slow on a sufficiently complex UI? (It can also lead to problems with widgets losing focus, but that can be solved in a similar way as in Immediate Mode GUI's).
A bind only re-evaluates if changes happen within its own immediate scope, and not if the changes are within any descendant binds' scopes. Whenever you want to carve off finer-granularity updates, scope things to a separate bind. This has scaled to, for instance, a complex WYSIWYG web page editor, where the whole page/document being edited is structured using cells and rendered with recursive binds - the app needs to be very responsive to each change a user makes, whether it's typing in text or dragging a style slider, even if the focus is on the top-level DOM element. That's not to say we've seen all the use cases - we definitely want to see where things fall down as well.
Say I have something like a button with an action, and the action has a cell whose value is used when the action is executed by clicking the button.
If I create the button in a bind, and the cell in the action is also made with a bind, but during the calculation to create the button the cell for the action is never actually read, then no listeners will be attached to the cell for the action, and the cell will therefore not update itself to underlying changes.
Now sometime later I click the button, but the cell for the action has an expired value, what happens?
I guess my question is, what happens when the reading of a cell value only happens outside the evaluation of binds?
One solution could be to always add a listener to new reactive cells created during a bind, another solution is simply to evaluate a cell everytime it's value is requested when it has no listeners added.
Think of cells (including binds) as just containers: when you call .get() (inside or outside a bind), you just get the currently stored value. Besides the container abstraction for its own sake, values also enable the scoping effect described earlier - if you have x.get() + y.get(), and x updates, you don't need to re-evaluate y (which in turn may be a bind - its body is scoped off so that you're not re-evaluating an entire sub-graph of the application). They're also necessary groundwork for smarter propagation strategies, e.g. stopping propagation if the value hasn't changed.
These are great questions! The topics deserve attention from the docs, esp. regarding suggested ways of doing things (e.g. the `shown` example from earlier vs. imperatively adding/removing elements).
Really cool stuff. I love the simplicity of reactive cells and I've been writing views in coffeescript for a while now. (https://github.com/jkarmel/space-pen)
As I was playing around with the JS fiddle todo example I noticed that there is no focus restoration when I have the todos update on 'keyup' instead of on the form submission:
Recently, I've seen the idea of reactive user interface pop up all over HackerNews and GitHub. What a lot of folks are missing is that whole applications themselves can be reactive, from UI to servers to data sources. I feel like many of these reactive UI libraries are missing this.
We elaborate briefly on various related work at http://yang.github.io/reactive-coffee/related.html, but absolutely - there's a large body of existing work and trailblazers we liberally draw inspiration from. In the case of RxJS in particular, the styles actually feel quite different - RxJS is more about explicitly merging streams and constructing the data flow graph with some set of combinators, whereas dependencies are inferred automatically in reactive.coffee (along with a UI-building layer).
We're also strong believers that the entire stack can be architected in this way. Stay tuned! For now, check out projects like Fun, Ur/Web, and Meteor/Derby:
A nice JS library for reactive programming is Bacon.js - https://github.com/raimohanska/bacon.js - it's smaller, the source code is easier to read and I found it a little easier to use.
This looks like just the sort of thing I've been looking for. Reactivity is very convenient, but it's so often tangled up in a massive library that's trying to do much more at once.
I've learned the hard way to embrace the "best in breed small libraries" philosophy over the "monolithic do-everything framework" approach, and I'm particularly excited to see a small, simple reactive library that goes as far as to use efficient DOM diffs.
You may be interested in what we've built with React (http://facebook.github.io/react/). We're focused exclusively on rendering and wiring up event handlers and offer a super straightforward, highly performant reactive programming model. There is some particularly interesting technical aspects around how we keep state consistent between React and the browser as well as how we've architected the core (it's a bit like a game!). And it's been proven at scale (both userbase and eng org size) time and time again.
Hi Peter - React is great! To those of you who haven't already, go check it out. It shares a lot of similarities with reactive.coffee (besides our highly original names), including the "code-first" approach to building up views.
I really like this. As an avid user of Coffeescript, Knockout.js and Erector-style code-as-templates, this is a great fit for me. Additionally, while I now understand how Knockout works internally, I was initially frustrated by how much magic it involved; Reactive.coffee looks like it does a better job of exposing the relevant parts of how it works.
Do you feel that you've lost anything in terms of composeability as compared to a more stream/combinator-oriented approach, like, say, Flapjax? If so, have you found that painful in the real world, or do you think it's more of an academic concern with little meaningful impact?
Good question - the reason why we built the library this way was exactly because we find declaring expressions to be more natural than wiring together streams by combinators—which, BTW, Flapjax does support. Far from academic-without-impact, Leo's work was actually directly influential on me—esp. since he worked in PL and I didn't, so much of my initial exposure to reactive programming came from a few projects including his.
That would require a new language/syntax layer, which we wanted to avoid. (There's also something to be mentioned for being explicit and reducing "magic," which was part of what drove us to create this library in the first place.)
That said, you could imagine a simple implementation by reflecting on each sub-expression (or auto-lifting operators/functions), where something like `z <- x + y` is compiled to:
z = bind ->
tmp0 = if x instanceof ObsCell then x.get() else x
tmp1 = if y instanceof ObsCell then y.get() else y
tmp0 + tmp1
Rather than forking CoffeeScript for the `<-`, you could start with a simple expression parser:
Also, our startup, Infer, is hiring great engineers, and we love open-source - come learn about us: https://www.infer.com/careers.html
(Thanks jashkenas!)