[news.eclipse.modeling.mdt.uml2] Re: Question about extending classes wi

["Christian W. Damus" <cdamus@xxxxxxxxxxxxx>]

Hi, Kenn,

Thanks for your response.  It is good to know that the OMG is actively 
looking at this, but also a little disconcerting to think that such a 
user-friendly mechanism as profiles have been might disappear!  Yes, 
stereotypes are fully equivalent to MOF models, but I'm not sure that 
the concept of ProfileApplication has any equivalent in metamodeling. 
It's a very definite relationship indicating some reasonably 
well-defined notions of domain-specific extension.  Of course, if 
"profile application" just takes another form, then that's another matter.

On the subject of the best practice for extensions:  I understand the 
practicality of defining extensions only on leaf stereotypes, but that 
does raise problems, too:

  - all leaf stereotypes in a hierarchy may end up with redundant
    extensions of the same metaclasses
  - how do I ask "is any stereotype conforming to Foo applied to my
    element" using standard association navigation semantics in OCL?
  - I don't necessarily know what is a leaf stereotype.  Another
    profile may well need to specialize my stereotypes

Perhaps the Classifier-ness of metaclass extensions can help address 
some of these issues.  Extensions (like other assocations) are 
classifiers, and thus can be abstract.  Let the super-stereotype in a 
hierarchy define an abstract extension that specializing stereotypes can 
specialize and redefine its ends, as necessary.  They can narrow or 
widen the redefined extension end's metaclass type for their own needs, 
while letting clients of the more general stereotype still make 
practical use of it.

I think we can marry some of these pragmatic best practices with 
standard metamodeling practices.  ;-)

In the mean-time, I think I shall also continue to mind an old 
best-practice that we haven't mentioned, yet:  don't make extensions 
required!  :-D  Or, is that not still a best practice?

Cheers,

Christian


Kenn Hussey wrote:
> Christian,
>
> I agree that the autmatic application of required stereotypes is unnecessary 
> and inconsistent, at least from a purely structural perspective. Indeed, the 
> bounds on extension ends would ideally be evaluated only during validation, 
> and a case where a required stereotype has not been applied, valdation would 
> fail.
>
> Unfortunately, much of the intended/expected semantics in UML aren't 
> strictly structural and are specified in a rather informal way; 
> incidentally, there is a working group at the OMG wrestling with this very 
> issue. The intended semantics of profile application (as an operation), in 
> cases where the profile contains required stereotypes, are such that all 
> required stereotypes are automatically applied. This is evidently 
> problematic in cases where there is a stereotype hierarchy involved, and 
> unsurprisingly is something that isn't explicitly addressed by the 
> specification. In the absence of formal guidance for cases like these, we 
> are left with best practices (which may eventually make their way into the 
> specification) as our guide. Based on the way the UML2 implementation works 
> (for better or worse), the "best practice" is to define extensions only on 
> non-abstract stereotypes and on leaf stereotypes in a hierarchy; using such 
> a practice offers the benefits of proeperty inheritence while at the same 
> time avoiding ambguous semantics.
>
> Based on some of the discussions that have been taking place at the OMG, it 
> may well be the case that the notion of "profiles", at least in its current 
> incarnation, will go away (to be replaced by MOF modeling, which is 
> essentially what profiles, originally introduced as a "lightweight" 
> extension mechanism, have become in practice).
>
> Kenn

-----8<-----