Hi Nicolas,
Thanks for opening the discussion, and sorry for the late reply but I was a bit ill the last days.
I think the discussion is valid, but you are mixing two distinct concerns: the deep FAMIX inheritance model, and the traversals.
1. Inheritance-based modelling
This has indeed some historical roots and some constraints:
- We designed it in VW before Traits were usable
- We used the types capabilities in Fame to be able to generate code in multiple languages
We wanted since a long time to extend it to use Traits, but nobody got time to lo it. However, in the meantime, Alain Plantec did extend the Pharo version of Fame to use Traits. We need to integrate this into Fame, and then ideally take care that the same implementation happens in Java as well, so that VerveineJ can benefit from the code generation.
Using Fame Traits, we can nicely have the type of a property point to the Trait, rather than the MetaDescription.
Nevertheless, I think this is less of a problem in practice because you can easily extend your meta-model with new properties. What you cannot do, is change the meaning of one property, but that you will not be able to do with Traits either. However, what you could gain with Traits would be to compose a new entity faster. That would definitely be cool.
2. Traversals
In your example, allMethods is essentially a travesal of the system. First, this is one traversal of the system, and if you have another semantics, you should create another one. But, ideally, traversal engines should be separate from the model (Karl Lieberherr (
http://www.ccs.neu.edu/home/lieber/) is a prominent advocate of this). For example, to some extent this is what happened with Chef: it offered operators that could be composable and that reside outside of the model. For another example, the FAMIXGraphVisitor I created lets you crawl through FAMIX by defining a traversal that is independent of the model. It would definitely be great to have more of these.
However, fancy traversals also come with a cognitive cost, and we should always strive to strike a balance between what is generic and what is practical.
But, another thing to keep in mind is that the new focus of Moose is less on FAMIX as a generic meta-model, but on building dedicated new ones faster. We no longer are driven by large costs of the tools built on top of the meta-model, because we constantly decrease those costs with engines like Glamour or Roassal.
In summary, having:
- Traits in Fame and FAMIX would be interesting for enabling the composition of a new entity out of smaller pieces
- Building new kinds of traversals and aggregators (like Carrack) would be equally interesting for quickly putting together queries that need to deal with the newly created semantics
Cheers,
Doru