On Fri, Aug 26, 2005 at 12:54:47AM -0700, Bob Ippolito wrote:
On Aug 25, 2005, at 11:30 PM, Andrew Bennetts wrote:
[...]
That's a pretty reasonable expectation. I think emitting a warning if this doesn't hold would be a sane thing to do, and it wouldn't break any code.
It would certainly slow it down, though. I think this is a "consenting adults" kind of scenario. If someone registers an adapter, they probably didn't do it by accident!
Right. The bug the OP has here isn't that the adapter shouldn't have been registered, but that the adapter fails to properly declare what interfaces it provides, so that the result of IBar(IFoo(x)) will fail even though x -> IFoo and IFoo -> IBar adapters are registered.
The performance question is important... benchmarks of any changes here would be a good idea.
A weaker solution would be to only check at registration time, but that's uglier, because it only works for adapters that are classes, and that don't have creative __new__ methods. That's almost certainly 99% of adapters, but ugly and incomplete solutions make me nervous.
Anyway, the underlying zope.interface machinery allows this -- and includes doctest documentation that relies on this (it registers *ints* as adapters, iirc!). I'm inclined to agree with your "consenting adults" approach, but I don't feel strongly either way.
-Andrew.