On Mon, 12 Jan 2009, Shawn O. Pearce wrote: > Daniel Barkalow <barkalow@xxxxxxxxxxxx> wrote: > > On Sun, 11 Jan 2009, Junio C Hamano wrote: > > > Daniel Barkalow <barkalow@xxxxxxxxxxxx> writes: > > > > This supports a useful subset of the usual fetch logic, mostly in the > > > > config file. > > > > > > I like the direction this series is going. In the longer term, it would > > > be nice if we can have git-svn and git-cvsimport replaced with something > > > like this. > > The series seems to require that the foreign tool speak fast-import. I've > tried to get git-svn to use it, but its currently not possible because the > git-svn process needs to be able to read objects it has written during this > session. Those objects are stored in the output pack, where only the active > fast-import process can get to them. Thus git-svn can't use fast-import. > > As import as the git-p4 stuff is, git-svn is our "killer feature" when it > comes to foreign system integration. I think we need to make SVN work for > this foreign vcs thing to work. I think fast-import should be made sufficient for git-svn, rather than having the integration about to speak other things. As you say, git-svn is the killer feature, and I'd like to have it using (and therefore contributing to the testing of) all of the applicable tools. Maybe fast-import should be able to run with a bi-directional connection to its input, so it can acknowledge checkpoints? When the pipeline is set up in the main git code, it should be easy enough to do this sort of complicated stuff. > > > Is the current foreign vcs interface sufficiently rich to support git as a > > > foreign scm by using fast-import and fast-export? > > > > I think so, although it would be pretty strange, since it would generally > > have a whole lot of local data (a complete clone of any remote > > repository). > > It might not be that bad. If the foreign system is git and the > local system is git, we should have a massive amount of object reuse. > At least all of the blobs from the foreign system should be reused > by the local system, and that's like 80-90% of most project's > object state. The part I think is weird is that the helper would be first going to the network to get all of the data, and then extracting the history from it locally in a separate process (that is, passing data to it only through the filesystem). Other systems generally get the data on demand, which sort of makes more sense for the case where you will then run exactly one command (fast-export) to it. I suppose sticking the data into the destination object store would make this not so strange. > If the import is flawless (no mangling of commit messages or history) > then you should get 100% object reuse and the git-git import would > give you the same SHA-1, but just slower than going over git://. > > Its strange only because it would be sucking more CPU time on the > client than is necessary to do the fetch. But if something like > a filter-branch tool existed to massage a fast-import stream, it > might be useful to fetch someone else's tree, massage it a bit, > and then import it with a subtree merge. :-) That is true, and would be a good actual use for this. It would make it possible for Wine to have a remote for the upstream Kconfig repository (that being a filtering of the linux-2.6 repository). That is, a git repository can be "foreign" in that it disagrees with you over the project layout or contents. -Daniel *This .sig left intentionally blank* -- To unsubscribe from this list: send the line "unsubscribe git" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
