May be delegation is better than subclassing when the class library is 
evolving?
--------------------------------------------------------------------------------

The root of the problem is, IMO, the very large surface area between class 
library and its subclasses in some NewStuff. An evolving class library will 
always be a problem, but almost impossible to tackle properly when there is 
a lot of subclassing. Automated tests help. Then somebody has to go through 
all tests for all available packages whenever there is a new version of the 
library and arrange for a bug fix when anything breaks. And what about 
interference between NewStuff-1 and NewStuff-x?

Being badly burnt, I have come to believe that subclassing often is evil 
because of the large surface areas involved. An alternative seems to be 
delegation. NewStuff classes will then defines facades to instances of 
library classes (e.g. Morphs). We then at least will declare the interfaces 
we use. Incidentally, it will also make us aware of the interfaces we rely 
upon without knowing it. The library could offer components with relatively 
stable external interfaces and evolving internals. May be we should promote 
interfaces to become first class citizens like in Java. (The current method 
categories are grossly underused).

An additional solution could be to provide bounded object spaces, something 
of an object parallel to the UML 2.0 merge relation -- a bounded object 
space with controlled visibility to other spaces. But even that won't help 
when needed library functionality changes.

--Trygve


At 05.11.2003 08:32, you wrote:
>Hi Guys,
>
>Here's a weird and somewhat OT thought that was triggered by a discussion I
>had today. One of the things I am noticing when looking at SqueakMap (and my
>use of it) over the last year is that many of the packages that are
>available from SM tend to break in the face of later Squeak versions. What
>happens? Effectively, these packages have been written under the assumptions
>of a certain environment in which they exist (namely that of Squeak X.Y) and
>if these assumptions are ever (even partially) invalidated they break.
>
>In the discussion I had today Alan (again) mentioned that one of the things
>that could be done for making sure an object (or a package) can survive in
>some environment is to model (at least some) of the assumptions about the
>environment explicitly. We do this internally by, for example, using unit
>tests for the package but typically we have no model whatsoever about our
>environment. This is (at least I think so) an even larger problem for a
>dynamically typed environment than it is for a statically typed environment.
>In the statically typed environment you can at least be assured that the
>compiler will do certain checks on your types (and messages) so unless the
>semantics of some message (or type) changes you should be safe (in theory).
>
>What I'm curious about here is: Has anyone ever seen a (hopefully even
>usable ;) approach to model "environmental assumptions"? Are there any
>common rules/guidelines on which "environment tests" could be based?
>Anything else in this area?
>
>Cheers,
>   - Andreas


-- 

Trygve Reenskaug      mailto: trygver at ifi.uio.no
Morgedalsvn. 5A       http://heim.ifi.uio.no/~trygver
N-0378 Oslo           Tel: (+47) 22 49 57 27
Norway