This is a discussion of the Model-View-Controller pattern in web applications. I wrote this for perlmonks.org, so some of the examples are Perl-specific.
For centuries people have sought ways to separate style from substance, good manners from good science, and user interface from application logic. The Model View Controller (MVC) design pattern is a way of separating the user-interface from the substance of the application.
In recent years, MVC has become a popular strategy for building websites. There are now web-MVC frameworks available for many programming languages, for instance Struts for Java, Maypole for Perl and Rails for Ruby.
This article introduces the MVC pattern and its use in web development. I will mention some Perl modules, but most of it will be applicable to any language supporting object-oriented programming.
MVC is usually implemented in an object-oriented style, so I'll mention objects, classes etc here. This text might not make very much sense without some experience in OO programming and web development. Information about all the patterns mentioned here can be found at the C2 Wiki: http://c2.com/cgi/wiki
Since MVC was originally invented for traditional GUI applications, certain details in original MVC pattern don't map well to web applications. (Note 1) Since I'm describing MVC for the web, I will simplify, change, misrepresent and ignore those details here.
The idea of the MVC pattern is to divide an application into 3 parts: the Model, View and Controller. Let's take a look at each of them.
The Model handles the state of the application. The state is what your application is about. If your application is a forum, your Model might contain Class::DBI objects representing threads, users and postings. The Model does not know anything about HTML, or web servers or anything like that. It just supplies ways to query the state, and ways to change that state.
The View is the representation of the user interface. Usually there are many (possibly nested) Views in a single application. A view can query the model, but it is not supposed to change the state. In web based MVC systems, a view can be implemented using a template that renders an HTML page. In our hypothetical forum application, the Views would be the templates for rendering a full thread, the login page, the posting page etc.
User actions on the View are send to the Controller. In a web environment, this is usually done by having the Controller handle the incoming HTTP requests.
The Controller receives user requests, and translates them into actions that the Model should take. Then it selects the appropriate View to handle the response.
It is possible to have more than one Controller, but most web application frameworks I've used assume you only have one.
Figure 1. A sequence diagram of a single request/response pair.
Browser View Controller Model . . . . . HTTP Request . . . +---------------------------------->+ . | . | update model . | . +-------------->+ | . | | | . | return status | | . +<--------------+ | . select view | . | +<----------------+ . | | . . | | query state . . | +-------------------------------->+ | | . | | | . return state | | +<--------------------------------+ | HTTP Response | . . +<----------------+ . . . . . .
Notice that the Controller does not handle the communication between the View and the Model: the Views make direct requests to the Model.
Since the Controller is in charge of handling the requests and selecting an appropriate page (View), there is no immediate coupling between the request made by the user and the resulting page.
This turns out to be very useful if the page-flow in the application is complex, but even for simple applications the Controller is a good place to handle common actions - authentication and session management can be handled in the Controller, for instance.
Since all code that does anything except building a nice page for the user is outside the View objects, changing the layout does not involve touching the logic of the application. Since the part of the application that changes the most during and after development is the layout, this means much less chance of adding bugs.
Since all actions on the application state are handled by the Model, it is possible to change the Model's implementation without touching the user interface, as long as the Model's public API doesn't change. (but see Coupling between View and Model).
Sometimes is just hard to figure out where a specific piece of the application is supposed to go. Especially dividing the Model from the Controller can be hard. As a rule of thumb, the Controller should be as minimal as possible - it is only responsible for translating HTTP requests into Model actions and selecting the right View - the Model should provide all the behavior it can without handling the HTTP requests or output formatting details.
One problem with having the View and Controller querying the Model is that changing the Model's public API means you also have to adapt the Controller and Views that act on it (Note 2). Adapting the Controller is usually not too much work, but changing a large number of Views will be annoying.
The Model Model View Controller pattern tries to minimize the impact of these changes by using two Models: a Domain Model and an Application Model - the View only queries the Application Model, and the Application model can query the Domain Model. The Application Model usually partly generated by the GUI design tools. I haven't used this pattern at all, so I don't know how useful it is for web applications
Creating an MVC application can result in more classes and objects than a ``page-based'' system. That means more design up front. On the other hand, a well designed MVC system will be easier to adapt and expand, because the code will be separated better.
Possibly :-) I think that any web-based application that uses more than a handful of templates, or has a complex interaction between pages would be a good candidate for the MVC pattern. There are alternatives, of course. See Resources - Alternatives.
Note: There are more frameworks. These are two examples I picked because they were the only two I had tried at the time I wrote this article.
CGI::Application, by Jesse Erlbaum is a relatively simple module that provides a good base-class for a Controller. It's easy to understand and it has some convenient HTML::Template support (though it doesn't force you to use HTML::Template if you like another templating system better).
You derive your Controller from the CGI::Application class, handling user requests via the run_modes system and you can use a templating system to implement the Views.
How you implement your Model is entirely up to you. IMO this is a Good Thing; the model is the most specific thing in an application, and a framework should not make unnecessary restrictions on it.
Maypole, a fairly new project by Simon Cozens is a complete MVC framework that uses lots of other CPAN modules: CGI::Untaint, Class::DBI and extensions, Template::Toolkit etc. It ``aims to be the most generic and extensible [ ... ] MVC-based web application framework'', and comes bundled with templates for building CRUD (Create Update Delete) applications. Maypole runs best on mod_perl (mod_perl 2 now supported).
While CGI::Application's approach is to set up a minimal base for your Controller, giving the programmer as much choice as possible, Maypole's focus is on rapid development: the example BeerDB application requires about 20 lines of Perl and a database schema. If you're building a CRUD application on mod_perl, Maypole might be just what you're looking for.
All this convenience comes at a price, of course - Maypole is big, slow to load (can be a problem in CGI environment), and if you don't want it to auto-generate a whole CRUD application based on your database schema, it takes a lot more time to figure out than CGI::Application.
In ``traditional'' MVC applications, Views and Controllers register their interest in parts of the Model and will then be notified when those parts change. (Via the Observer pattern). This assumes a permanent, single model and persistent views, which is fine if your application is a single process on a single machine, but becomes cumbersome when you spread an application over multiple processes or machines. Most web based MVC frameworks use a passive Model instead, where the Controller and View request information from the Model when they need it.
If your Model's API closely matches its implementation (say, Class::DBI objects where each class corresponds to a database table), rearranging your implementation can result in Model API changes.
You could argue that having Class::DBI objects in the public API is not very good encapsulation anyway, but models can change for other reasons too.
CGI::Application - Best Practices Wiki
A discussion of many Perl templating systems.
Discussion of several Perl Object-Oriented Persistence systems.
An explanation of the MVC pattern in traditional GUI applications.
With discussion and links to related topics. Many other design patterns are also discussed here.
A clear explanation of the MVC pattern as used by the PHP based Web Application Toolkit.
Sun's explanation of the MVC architecture.
An alternative pattern that directly addresses the problems in designing a client-server application.
A series of articles describing an alternative to MVC, arguing that MVC is not object-oriented.
$Revision: 1.8 $ $Date: 2005/06/28 14:52:41 $