Re: Last minute feature for 1.0.0

On Fri, May 04, 2007 at 03:38:14PM +0100, Richard Boulton
wrote:
> Olly Betts wrote:
> >I understand your motivation for wanting this in
1.0.0, but I'm afraid
> >the time really has passed for crowbarring random
new features into the
> >release.  I've already said "not until after
1.0.0" to patches submitted
> >weeks or months ago!
> 
> Fair enough - I just thought I'd ask, since the feature
requires an 
> incompatible database change.

I didn't look at the patch itself, but presumably it just
adds extra
information to the end of the metadata record, so this
change can at
least quietly upgrade existing flint databases?

Flint databases without the metainfo would be the same
before and after
anyway, and an application relying on this feature would
need to require
a particular Xapian version anyway.

> I don't _think_ there's a bug here, but I understand
your worries.  I 
> think I'll pull the patch back out and merge it with
the patch in bug 
> #143 for later application.

I don't actually think there is either, but I worry there
might be.

Removing the unused transaction method declarations is
certainly fine
(but they're non-virtual, so I doubt it actually makes any
different to
the generated code).

> >Also, we now always retrieve and parse the metainfo
record, even if we
> >never use the information - if you're using a
weighting scheme which
> >ignores document length, I think that means that
the first search on a
> >cold database will now be slower.  At least that's
not the default
> >for BM25Weight.
> 
> On the other hand, we only retrieve and parse the
information once now, 
> rather than every time we calculate the average
document length.

Yes, that is certainly an improvement (though I think the
layers
generally don't reread such values).

Cheers,
    Olly

_______________________________________________
Xapian-devel mailing list
Xapian-devellists.xapian.org
http://lists.xapian.org/mailman/listinfo/xapian-devel

Olly Betts wrote:
> I didn't look at the patch itself, but presumably it
just adds extra
> information to the end of the metadata record, so this
change can at
> least quietly upgrade existing flint databases?

Ah, no.  The problem is that pack_uint_last() is used for
the lastdocid 
value, instead of pack_uint(), so we don't know where the
end of the 
metadata record is.  If we made the tiny change to change 
pack_uint_last() to pack_uint() in the set method (and the
corresponding 
change to the unpack methods), we'd be able to make the
change without 
breaking old databases.

-- 
Richard

_______________________________________________
Xapian-devel mailing list
Xapian-devellists.xapian.org
http://lists.xapian.org/mailman/listinfo/xapian-devel