5 ms·

Interesting article, but I don't buy the conclusion. I think this is partly because the term MVC has been distorted considerably from its original meaning when

Interesting article, but I don't buy the conclusion. I think this is partly because the term MVC has been distorted considerably from its original meaning when used in the context of web sites. The key thing about real MVC is to separate not just the I/O details from the model, but the input details from the output. The view's only purpose is to render data from the model to some output device so the user can see it (or otherwise detect it). No matter how many monads you use to dress this up, it is fundamentally effectful: it updates the screen, plays sounds through speakers, etc. This is where the confusion with web-based "MVC" comes in. If all your view does is generate some HTML as pure data, it isn't much of a view at all. No-one can see the results. You need a web server sitting on top, to take that HTML and send it back to the browser client using HTTP, and in the traditional sense of MVC that would be part of the view code as well. This may be a useful distinction, in that it provides separation between the (necessarily impure) outside-facing code and the (potentially pure) rendering operation. A view module can be built up in layers from submodules just like any other part of a software system, and likewise the top layers can be effectful while depending only on pure code below. But the view module as a whole is always effectful, or it serves no purpose.

Periodic 16y ago
In terms of pure code, the view returns data, which the controller then can route to the appropriate place to show. In this case the view generates an HTML response, but it could also generate an audio response or a video response. I understand that from an OO stand point it makes a lot of sense for the view to directly control that, but in a functional world it makes more sense for the view to pass back something which the controller then routes to the user, making it more of a formatting layer than a display layer. So yes, a functional MVC and OO MVC may have very different meanings.
- stan_rogers 16y ago
  Huh? In the world I grew up in, the controller is the middleman sitting between the model (the raw data representation or, at the highest level, the ORM) and the view (the "client", if you will). The controller is where the application logic lives. In a web app, the view is the HTML chain (it requires the web server, the browser and the HTML/JS/CSS to be complete) and the controller is just about everything that sits between the database and the HTML page. Complete separation of the view and controller can only really happen with a templating system (or with a page that essentially self-assembles in the browser using AJAX/AJAJ).
  - Chris_Newton 16y ago
    This is a fascinating discussion, because I don't think either of the architectures described in the parent and grandparent posts is even close to MVC. Views returning data that controllers then show? Controllers containing application logic?! To me, it seems these things are almost the opposite of what MVC is supposed to achieve. Then again, distortion of the term is commonplace in server-side web programming frameworks, as I mentioned before. I did once look at the Wikipedia article on this subject, but it's one of those where people with no understanding of the original concept have taken over so comprehensively that it is beyond hope and needs a complete rewrite... preferably by someone who knows that Smalltalk is a programming language, not to be confused with idle chat at web developer conferences. :-)
    - stan_rogers 16y ago
      Hmmm... model as persistent state, controller as state-change manager and view as, well, what you see -- that's pretty much the way Burbeck described it. The model may "know" how to reflect state changes (and that would include things like object-relational mapping where appropriate, maintaining relational integrity and so on), but what to change (the application logic) is and was the responsibility of the controller -- yes, even in Smalltalk 80. Some of us do go back rather a long way, you know.
      - Chris_Newton 16y ago
        I suppose it depends on what you mean by "state-change manager". My understanding has always been that a controller's job was to translate user actions into domain-level operations to be carried out. You might start the foobazifier by clicking a toolbar icon, pressing a keyboard shortcut, or selecting a menu item, but they all have the same effect. The controller's job is to interpret those user actions to determine the required effect. The model's job is to understand what that effect means and update things accordingly. In practice, of course, it has never really been quite that simple. Views tend to have metadata (such as the current cursor position in a document or zoom level) that are also maintained by commands from the controller, so not all data is stored in the model, only the important, domain-related state. Observer patterns get implemented so that the model can remain independent of any particular controller/view pair, but it still has to provide both an API for updates from controllers and a defined set of events that interested modules can observe. In any case, the model's role is not just to be a thin wrapper around the database/file system/whatever. It is also the guardian of "business logic". I think one danger with the term MVC is that it suggests a system composed of only three modules. In reality, MVC is a high-level architectural pattern, and in non-trivial applications all three of those components are going to need a lot of internal structure as well. Representing domain concepts and persistent state may well be at the core of the model, but just about any code that works entirely in terms of domain concepts probably belongs within the model component somewhere.
ezyang 16y ago
Good point, but I think you’re thinking too much about the single client case. Consider if you have an application that serves multiple clients at the same time: the side effect of the view model is no longer “update a screen” or “play a sound through a speaker”, it’s “update screen X” or “play a sound through user X’s speakrs.” You can make a case for making this logic live in the logic, but it should start making you itchy, and I think this sort of routing functionality should live in the controller. In which case, your life is a lot easier if the View is pure.
- Chris_Newton 16y ago
  I'm sorry, but I don't see how what I described is limited to a single client. There is no reason you could not have your model running on one machine, and a view/controller pair running on another. If you have multiple clients, just start a view/controller process on each of them. Indeed, I have worked on systems with essentially this architecture, and it works fine. Fundamentally, MVC is a three-way relationship, designed to support an interative user interface. I don't see what is gained by trying to shoehorn a web-based system with a linear structure into those boxes. It's not wrong, it's just different, but it is unhelpful to call something different by the same name. Funnily enough, the MVC idea does work rather well for the client side of a web-based interface, as the event model in JavaScript lends itself quite neatly to handling by a controller, and the whole HTML/DOM set-up lends itself to rendering by a view.

Periodic 16y ago

In terms of pure code, the view returns data, which the controller then can route to the appropriate place to show. In this case the view generates an HTML response, but it could also generate an audio response or a video response. I understand that from an OO stand point it makes a lot of sense for the view to directly control that, but in a functional world it makes more sense for the view to pass back something which the controller then routes to the user, making it more of a formatting layer than a display layer. So yes, a functional MVC and OO MVC may have very different meanings.

stan_rogers 16y ago
Huh? In the world I grew up in, the controller is the middleman sitting between the model (the raw data representation or, at the highest level, the ORM) and the view (the "client", if you will). The controller is where the application logic lives. In a web app, the view is the HTML chain (it requires the web server, the browser and the HTML/JS/CSS to be complete) and the controller is just about everything that sits between the database and the HTML page. Complete separation of the view and controller can only really happen with a templating system (or with a page that essentially self-assembles in the browser using AJAX/AJAJ).
- Chris_Newton 16y ago
  This is a fascinating discussion, because I don't think either of the architectures described in the parent and grandparent posts is even close to MVC. Views returning data that controllers then show? Controllers containing application logic?! To me, it seems these things are almost the opposite of what MVC is supposed to achieve. Then again, distortion of the term is commonplace in server-side web programming frameworks, as I mentioned before. I did once look at the Wikipedia article on this subject, but it's one of those where people with no understanding of the original concept have taken over so comprehensively that it is beyond hope and needs a complete rewrite... preferably by someone who knows that Smalltalk is a programming language, not to be confused with idle chat at web developer conferences. :-)
  - stan_rogers 16y ago
    Hmmm... model as persistent state, controller as state-change manager and view as, well, what you see -- that's pretty much the way Burbeck described it. The model may "know" how to reflect state changes (and that would include things like object-relational mapping where appropriate, maintaining relational integrity and so on), but what to change (the application logic) is and was the responsibility of the controller -- yes, even in Smalltalk 80. Some of us do go back rather a long way, you know.
    - Chris_Newton 16y ago
      I suppose it depends on what you mean by "state-change manager". My understanding has always been that a controller's job was to translate user actions into domain-level operations to be carried out. You might start the foobazifier by clicking a toolbar icon, pressing a keyboard shortcut, or selecting a menu item, but they all have the same effect. The controller's job is to interpret those user actions to determine the required effect. The model's job is to understand what that effect means and update things accordingly. In practice, of course, it has never really been quite that simple. Views tend to have metadata (such as the current cursor position in a document or zoom level) that are also maintained by commands from the controller, so not all data is stored in the model, only the important, domain-related state. Observer patterns get implemented so that the model can remain independent of any particular controller/view pair, but it still has to provide both an API for updates from controllers and a defined set of events that interested modules can observe. In any case, the model's role is not just to be a thin wrapper around the database/file system/whatever. It is also the guardian of "business logic". I think one danger with the term MVC is that it suggests a system composed of only three modules. In reality, MVC is a high-level architectural pattern, and in non-trivial applications all three of those components are going to need a lot of internal structure as well. Representing domain concepts and persistent state may well be at the core of the model, but just about any code that works entirely in terms of domain concepts probably belongs within the model component somewhere.

ezyang 16y ago

Good point, but I think you’re thinking too much about the single client case. Consider if you have an application that serves multiple clients at the same time: the side effect of the view model is no longer “update a screen” or “play a sound through a speaker”, it’s “update screen X” or “play a sound through user X’s speakrs.” You can make a case for making this logic live in the logic, but it should start making you itchy, and I think this sort of routing functionality should live in the controller. In which case, your life is a lot easier if the View is pure.

Chris_Newton 16y ago
I'm sorry, but I don't see how what I described is limited to a single client. There is no reason you could not have your model running on one machine, and a view/controller pair running on another. If you have multiple clients, just start a view/controller process on each of them. Indeed, I have worked on systems with essentially this architecture, and it works fine. Fundamentally, MVC is a three-way relationship, designed to support an interative user interface. I don't see what is gained by trying to shoehorn a web-based system with a linear structure into those boxes. It's not wrong, it's just different, but it is unhelpful to call something different by the same name. Funnily enough, the MVC idea does work rather well for the client side of a web-based interface, as the event model in JavaScript lends itself quite neatly to handling by a controller, and the whole HTML/DOM set-up lends itself to rendering by a view.