Jonathan,
On Nov 17, 2006, at 9:20 PM, Jonathan Swartz wrote:
I agree it is uglier, however, I don't see it as a violation of DRY.
Most class hierarchies are not so shallow that you will have to
say extends 'My::Moose::Object'; more than a handful of times. In
my experience, having to do this helps to delineate the "base" of
your hierarchies, where as use My::Moose would be spread all over
your class hierarchy. This to me is a violation of DRY.
Sure, the number of classes that inherit directly from the base
class might make up a minority. But by that same logic, why
wouldn't you require people to inherit explicitly from
Moose::Object? :)
Actually Moose::Object is needed for Moose classes to function in the
expected way (providing a constructor, the does() method for checking
roles, and the means to run all BUILD and DEMOLISH methods). But I
actually don't require you to inherit from Moose::Object, I just make
it the default. If you want to inherit from another class, I don't
prevent it, nor do I try and force Moose::Object to be part of the
hierarchy.
It isn't so much the tedium of adding the 'extends' statement as
the consequence of someone *forgetting* to add it for some random
class. In that case, we'll have the wrong base class in one part of
the hierarchy, leading to obscure bugs.
True, your approach will solve this. But to play devils advocate,
something like this should get picked up in code reviews.
More generally, this isn't just about default base class - that was
just the initial motivating factor. I want to prepare for the
possibility that I'll want to tweak some other Moose behavior or
policy (in a perfectly external, API-respecting way) for the entire
project, after hundreds of classes have been created, and I don't
want to have to insert a new policy statement in all those classes.
I see the reasoning here as well, however (and this is just my
personal opinion, and development approach), this also ties you down
to /your/ Moose version. This is where I think a Log::Log4Perl style
config file would be more powerful since it would allow you to vary
metaclasses much more easily from a central location, and do it based
on package namespaces. I might try and sketch out this idea more this
weekend.
Though I suppose one could argue that most languages with strong OO
support never give you this much ability to control meta-OO
behavior, and people seem to do okay. And that allowing mass-
customization of Moose will tend to make classes less portable.
Well, Smalltalk and CLOS give you similar degrees of flexibility, but
in practice it is not used that much. Ruby gives you a lesser amount
of flexibility in it's meta-ness, and people seem to be really
experimenting with it a lot lately to do DSL-ish things. At some
point though, messing too deeply with the meta-layer is dangerous,
and ill-advised. Hell, I wrote the thing, and I try to keep my meta-
level-fiddling to a minimum :)
As for the mass-customization leading to less portability, you are
correct, LISP kinda suffered from this in the early 80s, and that is
why CLOS was standardized. However, because you can specialize
metaclasses (and sub-metaclasses like attributes, methods, etc) on
such a granular level (per-package, per-attribute, and even per-
method if you want), it is only an issue when you start subclasses
between heavily customized meta-schemes that you run the risk of
problems. This is basically the "metaclass incompatibility problem",
which is pretty well known (I link to some papers in some of the
POD), but with no agreed upon definitive solution. Moose takes a
"blow up if it gets too hairy" approach, although, you can override
this too, so if you *really* want the trouble, its all yours :)
- Stevan