M2M, UML and the Defense Industry
Today I had a chance to take a look into a large defense contractor. Of course I got involved with that company because of my MDSD experience, so I got a chance to see how they use MDSD to develop naval battle management software for frigates. They are completeley committed to UML based modeling because it is essential that they are "standards-based" with everything they do ("in the worst case we'll go to the OMG and have our approach standardized" - you can only say this sentence if you're really big :-)). They use a very formal development process with various models at various level of abstraction. Most of the transformations between the models are manual. They also have a number of automated transformations towards the end of the process. Now, what was really interesting is that they have all their intermediate models (results of transformations) also in the UML tool. The tool was grinding heavily under the load. When I asked them what they actually do with these "intermediate" models, their answer was: "Noting.". They don't look at them much and they don't modify them.
So, I was asking myself: why keep these models persistent at all? It just results in huge amounts of "stuff" in the UML tool that you don't really care about. All in all, this really put me off and confirmed the approach I always use and proclaim, which is that intermediate models in a transformation chain are useful so you can test the transformation steps or use them for documentation, but you don't really work with them. You should consider them as basically a data structure to connect the various transformation stages. If you need to "annotate" these intermediate models in order to prepare them for the next transformation stage, you should use aspect models and have the transformer read those as well as the original input.
Also, it was quite obvious that the UML-only approach (profiles or not) isn't really a sustainable approach. UML was useful, in their case, as the modeling language (for the models people really worked with) but was completely unsuitable for the intermediate models which were very domain (and technology specific). Custom meta models should be used there. This results in simpler models as well as simpler transformations. And remember: since you don't actually look at or modify these intermediate models, it doens't matter that they are not "standard" UML. And after all, they are (indirect) instances of MOF, which is also a standard....
It was also interesting from another perspective. The last time I was involved with the defense industry was when I did my internship at EADS in Ulm in 1997. Back then, Ada was still an important language and Java was completely unimportant. Today, Ada is considered legacy, and mission critical applications (e.g. in the naval battle management domain) are built using Java. It was also interesting to see that they are using an iterative development process with iterations of approximately 4 weeks length and an intermediate release every three iterations. It is really good to see that the waterfall (and the V-Model) seems to give way to somewhat more agile development.