JDJ's Bill Dudney (pictured) writes: With the popularity of Object Relational Mapping tools like Hibernate and Cayenne, developers are more often than other giving control of some of their code to models. Will this help raise MDA into the mainstream? Will MDA take its hoped-for place as the next level of abstraction for developers?

What about MDA

Model Driven Architecture, also known as MDA, started in late 2000 with a white paper. Basically the idea is that we define the software we want to build in sophisticated models that capture the detail of the application. Then from these abstract models a series of transformations is applied to turn that abstract model into a running application. The highest-level model is referred to as a Platform Independent Model (PIM). There is an even more abstract model called a Computational Independent Model but we won't discuss that model. The PIM is, as the name suggests, independent of the deployment platform (i.e. .NET Java EE 5 etc.) In this model the business is specified, classes that make up the domain are fleshed out and specified. This model is then transformed into one or more Platform Specific Models (PSM) that can be elaborated with more detail specific to the platform. From the PSM a running application can be generated.

Now of course there is a need to put in your own business logic. Most tools today provide a way for you to edit the 'business logic' apart from the fully generated code. For example AndroMDA (an open source MDA tool) will generate some files only once (where your business logic is written). Other tools like OptimalJ take a different approach giving you code you can edit and 'protected blocks' that are part of the generated code. I'm sure there are other approaches as well that are taken by other tools.

The Promise

• Productivity - the ability to raise the level of abstraction so that developers can become more productive is one of the greatest promises of MDA. Just as Java raised the level of abstraction from C/C++ so MDA raises the level abstraction for Java EE / .NET.
• Portability - greater ease of migration between various technologies (such as .NET to Java EE) or between different versions of the same technology (i.e. EJB 2.1 to EJB 3.0). Once developed your PIM is the repository of knowledge about your application, so moving to another underlying technology is 'easy'.
• Consistency - greater consistency of application architectural principals. This feature is hard to ignore. Many enterprise level projects have divergent architectures on the same project. Maintenance is very difficult in these projects to be sure.

The Problems

At its core MDA is about using meta-data to drive program creation. The idea is that if we can develop a sophisticated enough model to express software then we can fully generate the actual running program from the model (or even create a virtual machine that could execute the model). The problem with taking this idea too far of course is that we end up with just another platform. Probably even worse though is that it's 'programming with pictures' which was already tried at least once in the late 80s and failed miserably. Few are willing to try programming with pictures again.

Many proponents of the MDA approach like to say that given a PIM with sufficient detail one would be free to move between .NET, Java EE or to something like Hibernate & Spring assuming that you had the proper set of transformations for these other PSMs. While this is a great marketing pitch for MDA and the PIM it is just not that straightforward. A PSM has too much platform-specific knowledge buried in it to simply move between different technologies. After all the PSM is where the business logic that makes the application unique actually resides. In order to make this move all that logic must be changed to fit into the new target technology.

Finally, and probably most significantly, MDA must overcome the grass roots resistance to the idea. Developers like to develop. They don't like to have control taken from them. There is a fundamental distrust of code generation and a resistance to this type of abstraction.

Sea-Change?

There are a couple of moves afoot that make me think we might be on the verge of a change in perception about MDA. The first is a more pragmatic approach being taken by vendors. Instead of expecting developers to program in pictures, many vendors are taking a more pragmatic approach. Developers are expected to build more familiar UML class diagrams and annotate them. Few are expecting a full blown executable model.

Second and more significantly is the emergence of meta-data as a normal part of every day life for Java developers. Several years ago the XDoclet project started bringing meta-data into the mainstream. The 'killer-app' for XDoclet was that EJBs only needed the bean class, the rest of the required files (remote/local interfaces, deployment descriptor entries, value objects etc.) were updated/generated automatically by XDoclet. Many developers embraced this approach because of the reduced tedious work that had to be done. With XDoclet, developers no longer have to mess with keeping the remote and local interfaces in-sync with the implementation methods. Instead XDoclet automatically generates the remote and local interfaces.

At this point many MDA vendors and proponents should be saying to themselves, hey that is exactly what we have been doing for years! And it is, the difference is no visual model. The developers write the meta-data that would normally be in stereotypes and/or tagged values right into their code. For many this bridges a semantic gap that is missing in the visual modeling paradigm.

Back to the meta-data being more 'normal'. The other major change that has recently happened that brings meta-data front and center is the addition of Annotations to the Java SE 5 platform and especially the use of this meta-data in Java EE 5. Developers will be using meta-data on a daily basis.

Another thing that developers typically don't like about the MDA approach is the feeling of lack of control over what is generated. I have often heard the assertion that the developer could do better than the code generator and other such comments. While it is probably true that a hand-crafted piece of code would be 'better' in many respects, it is also true that the generated code can be done in a fraction of the time and is 'good enough'.

An area where developers have been resistant to adopt a meta-data driven approach in the past has been persistence. The same argument of 'I could do better' was used quite often. More recently though it seems that Object Relational Mapping frameworks like Hibernate and/or Cayenne have been gaining momentum. On the surface you might be thinking that I'm nuts to draw a comparison between R/O Mapping and MDA but the comparison is not that far off. Most MDA tools rely on code generation instead of a framework but basically it's the same kind of thing. Hibernate has a framework but could just as well generate code at run time (or imagine aspects being attached to your POJOs). Either way (framework or code generation) meta-data is driving the way objects are mapped to rows in tables.

With the current set of R/O mapping tools a lot of control is removed from the developer; what exact SQL is executed is no longer in the developers direct control. However many are willing to give up this control for the increased productivity gain allowed by using something like Hibernate. Who really wants to write all that JDBC code anyway?

So where does this leave us? Will MDA take its hoped -for place as the next level of abstraction for developers? Will MDA become the next best thing that is relegated to the dustbin of history? Hard to say for sure, one thing is for certain though, meta-data is becoming more a part of developers everyday lives.