[Yum-devel] Re: Speeding up depsolving

[Jeremy Katz]

On Wed, 2007-04-25 at 20:09 +0200, Gijs Hollestelle wrote:
> On 4/25/07, Gijs Hollestelle <gijs at gewis.nl> wrote:
> > On 4/25/07, Jeremy Katz <katzj at redhat.com> wrote:
> > > Okay, simpler case is cpio.
> > >
> > > What's happening is that I have multiple kernels installed.  We
> > > overwrite the kernel provs in goneprovs and thus only check with one of
> > > their provides, not both.  Which then leads to results which leave out
> > > the fact that libraw1394 depends on the newer of the two kernels I have
> > > installed.
> > Good catch!
> >
> > I've updated the patch. Now instead of constructing goneprovs and
> > gonefiles as dicts with goneprovs[name] = prov
> > I do goneprovs[prov] = 1 and when we need the name we get it as
> > prov[0] this avoids the issue you mentioned and probably saves a few
> > bytes of memory as well ;-)
> Looks like I jumped the gun there.. there was a reason the hash was
> built up as it was..
> 
> I did it the right way now, goneprovs[name] is now a list of provs. So
> in the kernel case goneprovs[kernel] will be ['prov1','prov2'].

This still looks like it's going to be overwriting previous values in
one or two cases.  And with dicts of lists and all the overhead of
checking for existence, etc I suspect that much of the speed gain (I
wasn't seeing a whole lot) is going to go away.  Given some of the other
places we're still making substantial speed gains, I suspect that this
isn't that productive of a path really...

Jeremy