[MERGE] knit._PackAccess

[John Arbash Meinel]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Robert Collins wrote:

...

> 
>> I'm okay with a tuple interface, as long as it is encapsulated well (so 
>> we
>> don't have to do tuple gymnastics in the bulk of our code). I've already 
>> stated
>> that I think _iter_changes should really return objects, since by then 
>> we've
>> culled the bulk. If we know that we'll having page-based indicies, we 
>> may also
>> have already culled the bulk of the data.
> 
> Initial pull and push will always be accessing every single key; we're
> dealing with very large key counts. I'd be loath to add an object
> interface to this code. Indexing is so low level, and so performance
> sensitive :(.
> 

Yeah. We should use tuple/dict/list interfaces where they are
particularly performance sensitive, and I'm fine with that. Well, I'm
okay with that, I really wish that objects weren't quite as slow as they
are. Have we checked if compiled objects are just as slow? I'm guessing
they would be, since you still have the getattr() overhead as they are
meant to be flexible.

John
=:->


-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFGudZjJdeBCYSNAAMRAu9+AJ4kHlxA2EsCpud/Yi119+z5Q+vlZgCeIpgQ
yiM3/V2qPEHhtkLN7zl0Ovk=
=laOA
-----END PGP SIGNATURE-----