I'm sorry not to add value to the discussion, but I wanted to say the back-and-forth has been a very educational look at some of the practical trade-offs that framework designers have to make.
Lots of interesting topics in Jonas' post worth discussing ... to pick a few:
> I tried to follow the Smalltalk-80 style of MVC very closely
> in Serenade.
Aside from ideological purity, did you notice any other benefits from sticking more closely to the Smalltalk-80 style? Personally, I can't imagine the benefit in "clearly defining the role of the controller as only reacting to user events" in a JavaScript app -- that's pretty much a single line of jQuery to listen for DOM events, and then call the appropriate method on the model. Having the same object responsible for rendering DOM elements also be responsible for listening to events on them seems much more useful.
> In Ember.js for example, all objects you want to use inherit
> from Ember.Object. In Serenade you can use any JavaScript
> object as a controller or a model.
This distinction seems entirely academic, as in the documentation, you show:
var model = {};
Serenade.extend(model, Serenade.Properties);
... effectively ignoring the prototype chain by copying properties instead, but achieving the same effect.
> I came to the conclusion that doing this is only really
> possible by writing a purpose built template language.
[...]
ul#comments
- collection @comments
li @title
Do you worry about the constraints that are entailed by tightly coupling a logic-less template language to your framework? Whenever I talk to folks that are trying to use logic-less templates to build big apps, they always seem to run into dead ends where being able to call a little bit of arbitrary code would make their templates so much easier -- whether it's zebra striping a table, checking the length of the array you're rendering, or what have you. To that extent, things like Handlebars provide escape hatches from Mustache's purity, with "registerHelper".
> Serenade has no dependencies. It achieves everything it
> does by using normal DOM manipulation.
Take care, as "normal DOM manipulation" is one of the slowest possible ways to build and render a UI. Do you have any benchmarks looking at how a Serenade template performs versus a more traditional innerHTML-based template of the same view, in, say IE7?
> Let’s first clear up that CoffeeScript classes are nothing
> else than fancy syntax for JavaScript’s prototypal inheritance,
> and the fact that they are called classes, is just calling
> them what prototypal inheritance is used for in 99% of cases.
[Applause]. Thanks for putting that in your post -- JavaScript's standard prototypal inheritance as practiced with constructor functions is just a longer way to say "class".
The advantage is that the controller knows nothing about the DOM, so it is usable and testable without the DOM. And while I agree that it's a single line to add a binding, that quickly becomes a single line for every single event you ever want to listen to. It ends up being a sizable amount of code. Aside from that, it seems clear that binding data is best done in the view, not in the model or controller, why are events different?
2) Using normal JavaScript objects
You don't have to extend the objects with Serenade.Properties. The downside is that data in the view isn't automatically updated. But as in the first examples in the README, it's perfectly acceptable to send in "JSON" directly to the view.
In the example right after the one you posted, I show how to add properties to the prototype. The intent is not to sidestep prototypal inheritance, but since multiple inheritance or mixins aren't available in JavaScript, copying over the functions is as close as we can get.
3) Break out
I just realized that it's not documented in the README, I'll have to fix that, but it's possible to add custom instructions. Any function added to `Serenade.Helpers` is available as an instruction in the view. This is very similar to handlebar's registerHelper. See the spec here: https://github.com/elabs/serenade.js/blob/master/spec/integr...
4) Performance
I have not done any benchmarks, so honestly I can't say how it compares performance wise. I'm not terribly worried about IE7. At the moment, we don't even support it, and it'll be dead soon. We could possible use document fragments to speed up the rendering, but I'm unsure if it actually makes a difference. I would argue that Rails is slow, Ruby is slow, PHP is slow, Java is slow, C is slow, etc..
Right -- you mentioned testing, but I'd like to hear more about benefits in your actual app. Putting something in a separate object just to make it more "testable", but without any actual benefit to, or effect on your app's behavior, strikes me as the essence of unnecessary abstraction. You can certainly test views that "know something" about the DOM without actually using the DOM as well...
4) View/Template Performance
Ok then, IE8. It has largely the same DOM performance characteristics (ie, slow) as IE7 -- and it'll be around forever, or as long as Windows XP lasts -- whichever comes first ;)
For the purposes of a web application, Rails/PHP/Python being slow matters far less than your template library being slow. They're in the backend, you can run many instances across many servers, you can cache their output, etc.
If your JS templating / View library is too slow, it means you can't use it to build the most interactive and interesting parts of your UI -- precisely the parts for which using client-side views would be most helpful.
1) I don't really see how it introduces any abstractions. The only thing we're doing is moving where event bindings are declared from the controller (or Backbone View) to the template. Other than that it's really pretty much the same.
4) I did a very dirty quick benmark. See http://pastie.org/3296683. This just inserts 1000 words into a <ul> in three different ways. With jQuery, by appending each item separately, with jQuery through string concatenation, and with Serenade. I tested this repeatedly with different numbers of words, in different orders and on different browsers.
jQuery with append is around 5x SLOWER than Serenade
jQuery with an entire template is around an order of magnitude FASTER than Serenade
In other words, the difference in how jQuery is used is way bigger than the difference to using Serenade.
I then did another dirty benchmark: http://pastie.org/3296747. I've only added a primitive event binding so that clicking the <li>s pops open an alert. A really interesting thing happens. In IE (7,8,9), the difference shrinks to around a factor 5. So jQuery is 5x faster. In all other browsers I tested however, Serenade outperforms even the full string concatenation jQuery example by a slight margin.
I don't think this proves that Serenade is fast, but it's at least not deficiently slow.
Jonas wrote:
> The advantage is that the controller knows nothing about the DOM, so it is usable and testable without the DOM.
Jeremy wrote:
> You can certainly test views that "know something" about the DOM without actually using the DOM as well...
I've been wondering for a while how to eliminate the slow and brittle Selenium layer from integration tests (perhaps I'm not the only one?). I'm thinking that if you can do away with the DOM, you might be able to somehow run the integration tests on the server, without launching a browser.
Do you have any concrete ideas how to do this, or are you only talking about unit tests in the text I quoted?
If you move most of the application client side, you could use something like jsdom to test it. That's obviously not going to get you the same level of confidence that a full stack test in a real browser would though. Maybe a combination of the two?
I dont see any reason to follow the mvc pattern slave-like. I mean as long as the framework you use gives you the abstraction you need to prevent DOM soup you are fine.
It all boils down to taste of coding style, fundamentalism etc.
Anyway, you made a framework - you should be proud of it!
