Petr Baudis [off-list ref] writes:

One Scott's concern that didn't occur to me was that a the time of
release, we could have broken links between the time tag is created and
tarballs are wrapped up. I *think* that in practice, this happens at the
same time, I wonder if Junio could confirm that.

Heh, and you did not Cc: me ;-)?

There is a mirroring process involved between the public machines and the
machine I push the tag into and place the tarballs.  I do not have control
over that mirroring.  But modulo that, the tarballs and RPMs are made
public before the tag and the tips of branches are pushed into the public
repository.

The release procedure goes like this (extend this as an addendum to
Documentation/howto/maintain-git.txt if somebody feels like it):

 * On the development machine outside k.org, create the tag, and prepare
   RPM for i386;

 * scp i386 RPM to a private staging area at k.org, and push the tag to a
   private building area also at k.org;

 * run the release procedure in the private building area at k.org, which:

   - builds x86_64 RPM and deposits it to the same private staging area
     i386 RPM were scp'ed to earlier;

   - builds the source tarball and documentation tarballs;

   - puts all of the above in /pub/software/scm/git/ to be mirrored out;

   - extracts html documentation tarball in /pub/software/scm/git/docs/v*
     to be mirrored out;

         http://www.kernel.org/pub/software/scm/git/docs/, the "current"
         documentation page, has links to these "documentation for
         released versions" and they point at these docs/v* areas.

 * push the tag and branch tips out to the public repository, so that it
   will be mirrored to /pub/scm/git/git.git/ (this updates the "current"
   documentation pages as a side effect);

 * send out the release announcement message to the list.

The 1.4.4.5 backport was an oddball.  I do not think I did anything other
than simply pushing the tag out.