[PEAK] Trellis: setting @maintain attributes / initialization rollback

[Phillip J. Eby]

At 11:47 PM 10/8/2008 +0300, Sergey Schetinin wrote:
>Consider this example:
>
>class C(Component):
>     @maintain(optional=True)
>     def calc(self):
>         return 0
>
>     @maintain
>     def getx(self):
>         print 'get'
>         self.calc
>
>     @maintain
>     def set(self):
>         print 'set'
>         self.calc = 1
>
>c = C()
>
>Depending on luck (order of hashes of get / set attr names I suppose)
>we either get InputConflict or not.
>If the get rule runs first, set rule is a conflict because the Cell is
>set by the calc rule.
>If set gets to run first, there's no conflict because calc rule
>doesn't need to run.

Actually, in that last case, the calc rule does need to run -- and 
does.  It's just that its output is discarded, and not considered a conflict.


>Is it possible to have calc / get rules rollback even when set runs last?

Hm.  I suppose.  This seems to work (line 143 of trellis.py):

             if self._set_by is not ctrl.current_listener and 
self._set_by is not self:

It doesn't break any of the existing tests, and adding a test for the 
above (forcing the order using an __init__ method) seems to work:

     def testResetCalcValue(self):
         class C(trellis.Component):

             trellis.maintain(optional=True)
             def calc(self):
                 return 0

             trellis.maintain()
             def getx(self):
                 print 'get'
                 self.calc

             trellis.maintain()
             def set(self):
                 print 'set'
                 self.calc = 1

             def __init__(self):
                 self.getx
                 self.set

         c = C()


That is, it generates an InputConflict without the change, and prints 
"get", "set", "get" with the change.

I am a little suspicious of this, even though it *seems* 
correct.  Mainly I worry a little about certain corner cases for 
initialization, but feel free to try the patch above for now.  I want 
to think about this a bit more before I officially support it.

I agree, however, that it *should* be supported, because it's an 
order dependency of precisely the sort that the trellis should make 
impossible.  I just want to make sure it covers all the bases, and 
may need to think of some other test scenarios.