cg-pull to pull branches, not the whole repository

Hello!

I believe cg-pull should be changed to pull only the objects relevant to the branch rather than the whole repository. This could be more expensive in terms of processor power (since the objects would be checked for parents), but it would save the bandwidth, especially for users tracking stable branches with little changes in presence of heavily changed development branches in the same repository.

To keep things simple, pulling the whole repository should be removed from cogito. Users don't need objects that are not reachable from any branch known to cogito. If I'm not tracking a branch, I don't want objects from it. To compensate for intermittently connected users, cg-pull could have an option to pull from all branches (or from all branches marked for automatic pull).

Ideally, cogito should not rely on the server being able to list directories. This would allow support more protocols and servers, e.g. http servers without directory listings. Unfortunately, listings are still needed to get refs/tags and refs/heads. Any fix for that would involve git core, and I don't see an elegant fix.

For git+ssh protocol, git-fetch-pack already provides a server side solution. Client side support would be needed for http and rsync.

I'd like to hear comments before I attempt any hacking in this direction.

Pavel Roskin <proski@gnu.org> writes:

All the necessary "core" support should already be there, except for the client side support. Please look at yesterday's thread with title "dumb transports not being welcomed".

Personally I think Cogito should be able to just call git-fetch to implement cg-pull -- I do not claim the current git-fetch has everything needed (incapable of using objects/info/alternates is an example), but we should be able to fix and enhance just one program to make everybody happy, not two.

On Wed, 2005-09-14 at 15:05 -0700, Junio C Hamano wrote:

Great. It looks like my concerns are fully understood and the work is underway to address them. Sorry for the noise.

Dear diary, on Wed, Sep 14, 2005 at 11:31:56PM CEST, I got a letter where Pavel Roskin <proski@gnu.org> told me that...

Hello,

well, that should be the case for all the methods but rsync. Actually, it's not the case for local pulls now either, because git-local-pull wasn't able to deal with packs, but I think that's fixed by now so I'll reenable that.

Well, there's that update-server-info thing. It'd be nice if git-send-pack could call that automagically or something, so it's kept up-to-date. But I have no idea about how expensive can the command actually be.