Assorted concerns about git repository
Matt McCutchen
matt at mattmccutchen.net
Thu Nov 29 04:10:52 GMT 2007
Wayne,
Thanks for addressing the concerns. I have some comments:
On Wed, 2007-11-28 at 15:33 -0800, Wayne Davison wrote:
> The main purpose for the generated files being in CVS and/or git
> (including the magic "configure" script) is to make the build farm work.
> When a human checks out the files from git, I see no problem requiring
> them to build the generated files themselves [...]
That's reasonable.
> The latest code has been modified to only
> allow a build-farm system to [download the generated files ...]
Good.
> > 3. It looks like you removed patches/ from the history. I think it is
> > valuable to have the history of the branches from before they became
> > real branches.
>
> I didn't want all the patch (side-branch) commits cluttering up the
> master branch's release log.
That makes sense.
> This history can be accessed in the old
> CVS archive, which is going to remain available.
I would like to have it in my git repository. Should I import it
myself?
> > 4. Why do you recommend that people run "git-set-file-times"? I would
> > like to see more explanation before I mess with my source files' mtimes.
>
> The web page mentions that it sets the file times to their time of last
> modification. I want an INITIAL checkout (clone) to always set each
> file's time to the time of its last change (not the checkout time), and
> this script works-around git's deficiency in this area. This makes a
> fresh checkout the same no matter when it is cloned (though two
> checkouts would diverge in their mtime settings as they are
> incrementally updated after that point).
I understand that git-set-file-times will achieve this CVS-like effect.
My question is, why do you care about the mtimes of two fresh checkouts
being the same? I consider git's complete obliviousness to mtimes to be
a purity rather than a deficiency.
Finally, let me add another concern:
6. I notice that "support/patch-update" updates branches by rebasing.
This approach has a serious problem: when a branch with a long history
is rebased onto a new trunk, some of the intermediate revisions of the
branch may no longer make sense, and there is no good way to update the
branch while retaining a record of those revisions.
The obvious alternative, merging, is also deficient. Suppose branch B
depends on branch A (e.g., detect-renamed-lax and detect-renamed);
naturally, B would be set up to track A. Once B gets commits from A
woven into its history, there is no good way to remove the A-specific
material if the dependency later ceases to exist. Simply applying the
inverse of the latest trunk->A diff will trip up anyone who later tries
to merge B with another branch that actually does require A.
In light of this, I admit (red-faced) that the directory of patches
actually seems to be better than anything git supports natively. I
wonder if the git people have a recommended solution for projects like
rsync that have a trunk and branches with dependencies. I have some
ideas about how a solution could be added to (or scripted around) git,
but I will let them mature before proposing anything.
Matt
More information about the rsync
mailing list