[PEAK] Drop Python 2.3 support for the trellis?

[Phillip J. Eby]

One possibility I'm looking at for decreasing trellis memory usage is 
to use a specialized subclass of weakref.  By adding pairs of forward 
and backwards pointers, it would be possible to make a linked list of 
weakrefs and have them automatically unlink themselves when their 
target is deleted.  This would be ideal for managing notification 
links in the trellis, especially under the new algorithm.

The downside, of course, is that weakrefs aren't subclassable in 
Python 2.3.  You have to have 2.4 at least.  But currently, the 
target version for pretty much all PEAK projects is Python 2.3, as it 
has been for almost 3 years now.  (We dropped 2.2 support in December '04.)

At the time, a big argument for keeping 2.3 was that it was what Mac 
OS shipped, but by now Apple is shipping 2.5 (with setuptools, 
even!).  In my day-to-day work I use 2.5 mostly, as that is what 
Chandler's built on, and the build folks at OSAF are planning to use 
the system Python installations on Ubuntu and Mac OS Leopard.  So 2.5 
support seems pretty ubiquitous.

Looking at the PyPI download stats, there seems to be similar 
interest levels between 2.3 and 2.4 for less-popular PEAK packages 
like SymbolType and BytecodeAssembler.  But for more-popular packages 
like DecoratorTools and setuptools, interest in 2.4 and 2.5 versions 
is nearly *two orders of magnitude* higher than interest in 2.3.

That suggests that the availability of 2.4 is quite good, and that it 
might be reasonable to start dropping support for 2.3 in newer 
packages.  Going to 2.4 would certainly be a win for decorator 
syntax, built-in sets, and the enhanced 'doctest' package (which in 
some PEAK packages is kludged in as a backport to support 2.3).

So, I'm thinking it might be well worth dropping 2.3 support for the 
Trellis.  The Trellis' own download stats show 60 2.3 downloads vs. 
90 for 2.4 and 110 for 2.5, so interest in the Trellis appears 
somewhat correlated to the Python version in use.  :)

However, that's still a pretty good number of 2.3 downloads.  Is 
anybody here in that group?

If worst comes to worst, I could probably figure out a way to kludge 
up the weakref trick to work on 2.3, by not subclassing weakref and 
instead wrapping one.  It would, however, have degraded performance 
and increased memory usage under 2.3.  And of course it'd be more 
work for me, so I probably wouldn't do it until the 2.4+ version was 
stabilized.  (i.e., the SVN trunk would be 2.4+ for a while.)

(Another alternative would be to create the specialized link type 
directly in Pyrex or C, but that introduces its own headaches, and 
still wouldn't be sane to implement prior to stabilization of the new 
algorithm.  But since link manipulation is likely to be where the 
trellis spends most of its time, it would probably be worth doing at 
some point.)

Any thoughts?