git clone downloads objects that are in GIT_OBJECT_DIRECTORY

Hi folks,

Doing a fresh git clone git://some.git.url/ foo seems to download the entire remote repository even if all the objects are already stored in GIT_OBJECT_DIRECTORY=/home/bcrl/.git/object . Is this a known bug? At 100MB for a kernel, this takes a *long* time.

		-ben (who needed to free up disk space)

Benjamin LaHaise <bcrl@kvack.org> wrote:

I believe it is a known missing feature. :-) git-clone doesn't prep HEAD to have some sort of starting point so the pull it uses to download everything literally downloads everything as nothing is in common.

One could work around it by running git-init-db to create the new clone locally, git-update-ref HEAD to some commit which you have in common with the remote, create a origin file, then perform a git-pull. This would only download the objects between the commit you put into HEAD and the current master of the remote... But that is actually some work.

I think Cogito's clone is capable of restarting a failed clone; I wonder if that logic would benefit you here?

Is using a common GIT_OBJECT_DIRECTORY across many clones actually pretty common? Maybe its time that git-clone gets some more smarts with regards to what it yanks from the origin.

Shawn Pearce <spearce@spearce.org> writes:

You would first 'clone -l -s' from your local repository and then clone into that from whatever remote, perhaps.

Junio C Hamano <junkio@cox.net> wrote:

Yea but that's about as much fun as creating a bare repository by hand. (Which I've been doing up until this thread prompted me to read git-clone.sh and learn the existance of --bare.)

It might be nicer if the user could place a list of locally (here locally being possibly remote but closer network-wise) available repositories which should be considered as sources for faster cloning. When cloning a remote repository git-clone would try to examine each of the designated repositories to see if any of them have commits in common with the remote; if so clone off that and then pull from the remote, but designating the remote as `origin'.

This could be ugly if you have a large number of locally available candidates or if the candidates are many (e.g. 1000s) commits behind the remote being cloned. But it would save the user from pulling down 100+MB of objects they already have while making it very convient to establish a new repository+working directory based on someone else's publically available repository.

Or we could just tell the user to create their own clone script, e.g. kernel-clone:

	#!/bin/sh
	git-clone -l -n -s ~/kernel-base "$2" &&
	cd "$2" &&
	echo "URL: $1" >.git/remotes/origin &&
	echo "Pull: master:origin" >>.git/remotes/origin &&
	git-pull

But it would be better if it was more integrated, and somehow slightly more automatic...

On Sun, 5 Mar 2006 21:57:02 -0500 Shawn Pearce <spearce@spearce.org> wrote:

It is already easy to start from a similar repo (eg. locally cloned) if you wish to conserve bandwidth.

However, it might be nice to have a command that allows you to change origin information for a repo without needing to know git internals; maybe something like:

$ git set-origin <URL>

Or maybe better:

$ git set-remote --pull master:origin origin <URL>

Sean

Hi,

On Sun, 5 Mar 2006, sean wrote:

FWIW, I once sent patches to make this easier by placing this information into the config file, but for reasons I did not understand, they were rejected. Sigh!

Ciao, Dscho