(Debian) packaging and Git.

The big picture is as follows. In my view, the most natural way to work on a packaging project in version control [1] is to have an upstream branch which either tracks upstream Git/Hg/Svn, or imports of tarballs (or some combination thereof, and a Debian branch where both modifications to upstream source and commits to stuff in ./debian are added [2]. Deviations from this are mainly motivated by a desire to export source packages, a version control neutral interchange format that still preserves the distinction between upstream source and distro modifications. Of course, if you're happy with the distro modifications as one big diff, then you can stop reading now gitpkg $debian_branch $upstream_branch and you're done. The other easy case is if your changes don't touch upstream; then 3.0 (quilt) packages work nicely with ./debian in a separate tarball.

So the tension is between my preferred integration style, and making source packages with changes to upstream source organized in some nice way, preferably in logical patches like, uh, commits in a version control system. At some point we may be able use some form of version control repo as a source package, but the issues with that are for another blog post. At the moment then we are stuck with trying bridge the gap between a git repository and a 3.0 (quilt) source package. If you don't know the details of Debian packaging, just imagine a patch series like you would generate with git format-patch or apply with (surprise) quilt.

From Git to Quilt.

The most obvious (and the most common) way to bridge the gap between git and quilt is to export patches manually (or using a helper like gbp-pq) and commit them to the packaging repository. This has the advantage of not forcing anyone to use git or specialized helpers to collaborate on the package. On the other hand it's quite far from the vision of using git (or your favourite VCS) to do the integration that I started with.

The next level of sophistication is to maintain a branch of upstream-modifying commits. Roughly speaking, this is the approach taken by git-dpm, by gitpkg, and with some additional friction from manually importing and exporting the patches, by gbp-pq. There are some issues with rebasing a branch of patches, mainly it seems to rely on one person at a time working on the patch branch, and it forces the use of specialized tools or workflows. Nonetheless, both git-dpm and gitpkg support this mode of working reasonably well [3].

Lately I've been working on exporting patches from (an immutable) git history. My initial experiments with marking commits with git notes more or less worked [4]. I put this on the back-burner for two reasons, first sharing git notes is still not very well supported by git itself [5], and second Gitpkg maintainer Ron Lee convinced me to automagically pick out what patches to export. Ron's motivation (as I understand it) is to have tools which work on any git repository without extra metadata in the form of notes.

Linearizing History on the fly.

After a few iterations, I arrived at the following specification.

• The user supplies two refs upstream and head. upstream should be suitable for export as a .orig.tar.gz file [6], and it should be an ancestor of head.

• At source package build time, we want to construct a series of patches that

  1. Is guaranteed to apply to upstream
  2. Produces the same work tree as head, outside ./debian
  3. Does not touch ./debian
  4. As much as possible, matches the git history from upstream to head.

Condition (4) suggests we want something roughly like git format-patch upstream..head, removing those patches which are only about Debian packaging. Because of (3), we have to be a bit careful about commits that touch upstream and ./debian. We also want to avoid outputting patches that have been applied (or worse partially applied) upstream. git patch-id can help identify cherry-picked patches, but not partial application.

Eventually I arrived at the following strategy.

1. Use git-filter-branch to construct a copy of the history upstream..head with ./debian (and for technical reasons .pc) excised.

2. Filter these commits to remove e.g. those that are present exactly upstream, or those that introduces no changes, or changes unrepresentable in a patch.

3. Try to revert the remaining commits, in reverse order. The idea here is twofold. First, a patch that occurs twice in history because of merging will only revert the most recent one, allowing earlier copies to be skipped. Second, the state of the temporary branch after all successful reverts represents the difference from upstream not accounted for by any patch.

4. Generate a "fixup patch" accounting for any remaining differences, to be applied before any if the "nice" patches.

5. Cherry-pick each "nice" patch on top of the fixup patch, to ensure we have a linear history that can be exported to quilt. If any of these cherry-picks fail, abort the export.

Yep, it seems over-complicated to me too.

TL;DR: Show me the code.

You can clone my current version from

git://pivot.cs.unb.ca/gitpkg.git

This provides a script "git-debcherry" which does the history linearization discussed above. In order to test out how/if this works in your repository, you could run

git-debcherry --stat $UPSTREAM

For actual use, you probably want to use something like

git-debcherry -o debian/patches

There is a hook in hooks/debcherry-deb-export-hook that does this at source package export time.

I'm aware this is not that fast; it does several expensive operations. On the other hand, you know what Don Knuth says about premature optimization, so I'm more interested in reports of when it does and doesn't work. In addition to crashing, generating multi-megabyte "fixup patch" probably counts as failure.

Notes

1. This first part doesn't seem too Debian or git specific to me, but I don't know much concrete about other packaging workflows or other version control systems.

2. Another variation is to have a patched upstream branch and merge that into the Debian packaging branch. The trade-off here that you can simplify the patch export process a bit, but the repo needs to have taken this disciplined approach from the beginning.

3. git-dpm merges the patched upstream into the Debian branch. This makes the history a bit messier, but seems to be more robust. I've been thinking about trying this out (semi-manually) for gitpkg.

4. See e.g. exporting. Although I did not then know the many surprising and horrible things people do in packaging histories, so it probably didn't work as well as I thought it did.

5. It's doable, but one ends up spending about a bunch lines of code on duplicating basic git functionality; e.g. there is no real support for tags of notes.

6. Since as far as I know quilt has no way of deleting files except to list the content, this means in particular exporting upstream should yield a DFSG Free source tree.

I've been experimenting with a new packaging tool/workflow based on marking certain commits on my integration branch for export as quilt patches. In this post I'll walk though converting the package nauty to this workflow.

1. Add a control file for the gitpkg export hook, and enable the hook: (the package is already 3.0 (quilt))

   % echo ':debpatch: upstream..master' > debian/source/git-patches
   % git add debian/source/git-patches && git commit -m'add control file for gitpkg quilt export'
   % git config gitpkg.deb-export-hook /usr/share/gitpkg/hooks/quilt-patches-deb-export-hook
   

   This says that all commits reachable from master but not from upstream should be checked for possible export as quilt patches.

2. This package was previously maintained in the "recommend topgit style" with the patches checked in on a seperate branch, so grab a copy.

    % git archive --prefix=nauty/ build | (cd /tmp ; tar xvf -)
   

   More conventional git-buildpackage style packaging would not need this step.

3. Import the patches. If everything is perfect, you can use qit quiltimport, but I have several patches not listed in "series", and quiltimport ignores series, so I have to do things by hand.

   % git am  /tmp/nauty/debian/patches/feature/shlib.diff
   

4. Mark my imported patch for export.

   % git debpatch +export HEAD
   

5. git debpatch list outputs the following

   afb2c20 feature/shlib
   Export: true
   
   makefile.in |  241 +++++++++++++++++++++++++++++++++--------------------------
   1 files changed, 136 insertions(+), 105 deletions(-)
   

   The first line is the subject line of the patch, followed by any notes from debpatch (in this case, just 'Export: true'), followed by a diffstat. If more patches were marked, this would be repeated for each one.

   In this case I notice subject line is kindof cryptic and decide to amend.

    git commit --amend
   

6. git debpatch list still shows the same thing, which highlights a fundemental aspect of git notes: they attach to commits. And I just made a new commit, so

   git debpatch -export afb2c20
   git debpatch +export HEAD
   

7. Now git debpatch list looks ok, so we try git debpatch export as a dry run. In debian/patches we have

   0001-makefile.in-Support-building-a-shared-library-and-st.patch series

   That looks good. Now we are not going to commit this, since one of our overall goal is to avoid commiting patches. To clean up the export, rm -rf debian/patches

8. gitpkg master exports a source package, and because I enabled the appropriate hook, I have the following

    % tar tvf ../deb-packages/nauty/nauty_2.4r2-1.debian.tar.gz | grep debian/patches
    drwxr-xr-x 0/0               0 2012-03-13 23:08 debian/patches/
    -rw-r--r-- 0/0             143 2012-03-13 23:08 debian/patches/series
    -rw-r--r-- 0/0           14399 2012-03-13 23:08 debian/patches/0001-makefile.in-Support-building-a-shared-library-and-st.patch
   

   Note that these patches are exported straight from git.

9. I'm done for now so

   git push 
   git debpatch push
   

the second command is needed to push the debpatch notes metadata to the origin. There is a corresponding fetch, merge, and pull commands.

More info

• Gitpkg bug

• Git Repo with git-debpatch (WIP)

• Example package: nauty

• Example package: bibutils In this package, I was already maintaining the upstream patches merged into my master branch; I retroactively added the quilt export.

As of version 0.17, gitpkg ships with a hook called quilt-patches-deb-export-hook. This can be used to export patches from git at the time of creating the source package.

This is controlled by a file debian/source/git-patches. Each line contains a range suitable for passing to git-format-patch(1). The variables UPSTREAM_VERSION and DEB_VERSION are replaced with values taken from debian/changelog. Note that $UPSTREAM_VERSION is the first part of $DEB_VERSION

An example is

 upstream/$UPSTREAM_VERSION..patches/$DEB_VERSION
 upstream/$UPSTREAM_VERSION..embedded-libs/$DEB_VERSION

This tells gitpkg to export the given two ranges of commits to debian/patches while generating the source package. Each commit becomes a patch in debian/patches, with names generated from the commit messages. In this example, we get 5 patches from the two ranges.

 0001-expand-pattern-in-no-java-rule.patch
 0002-fix-dd_free_global_constants.patch
 0003-Backported-patch-for-CPlusPlus-name-mangling-guesser.patch
 0004-Use-system-copy-of-nauty-in-apps-graph.patch
 0005-Comment-out-jreality-installation.patch

Thanks to the wonders of 3.0 (quilt) packages, these are applied when the source package is unpacked.

Caveats.

• Current lintian complains bitterly about debian/source/git-patches. This should be fixed with the next upload.

• It's a bit dangerous if you checkout such package from git, don't read any of the documentation, and build with debuild or something similar, since you won't get the patches applied. There is a proposed check that catches most of such booboos. You could also cause the build to fail if the same error is detected; this a matter of personal taste I guess.

I recently decided to try maintaining a Debian package (bibutils) without committing any patches to Git. One of the disadvantages of this approach is that the patches for upstream are not nicely sorted out in ./debian/patches. I decided to write a little tool to sort out which commits should be sent to upstream. I'm not too happy about the length of it, or the name "git-classify", but I'm posting in case someone has some suggestions. Or maybe somebody finds this useful.

#!/usr/bin/perl

use strict;

my $upstreamonly=0;

if ($ARGV[0] eq "-u"){
  $upstreamonly=1;
  shift (@ARGV);
}

open(GIT,"git log -z --format=\"%n%x00%H\" --name-only  @ARGV|");

# throw away blank line at the beginning.
$_=<GIT>;

my $sha="";
LINE: while(<GIT>){

  chomp();

  next LINE if (m/^\s*$/);

  if (m/^\x0([0-9a-fA-F]+)/){
    $sha=$1;
  } else {
    my $debian=0;
    my $upstream=0;

    foreach my $word  ( split("\x00",$_) ) {
      if  ($word=~m@^debian/@) {
        $debian++;
      } elsif (length($word)>0)  {
        $upstream++;
      }
    }

    if (!$upstreamonly){
      print "$sha\t";
      print "MIXED" if ($upstream>0  && $debian>0);
      print "upstream" if ($upstream>0  && $debian==0);
      print "debian" if ($upstream==0  && $debian>0);
      print "\n";
    } else {
      print "$sha\n" if ($upstream>0  && $debian==0);
    }

  }
}

=pod

=head1 Name
git-classify  - Classify commits as upstream, debian, or MIXED

=head1 Synopsis

=over

=item B<git classify> [I<-u>] [I<arguments for git-log>]

=back

=head1 Description

Classify a range of commits (specified as for git-log) as I<upstream>
(touching only files outside ./debian), I<debian> (touching files only
inside ./debian) or I<MIXED>. Presumably these last kind are to be
discouraged.

=head2 Options

=over

=item B<-u> output only the SHA1 hashes of upstream commits (as
      defined above).

=back

=head1 Examples

Generate all likely patches to send upstream
   
     git classify -u $SHA..HEAD | xargs -L1 git format-patch -1

racket (previously known as plt-scheme) is an interpreter/JIT-compiler/development environment with about 6 years of subversion history in a converted git repo. Debian packaging has been done in subversion, with only the contents of ./debian in version control. I wanted to merge these into a single git repository.

The first step is to create a repo and fetch the relevant history.

TMPDIR=/var/tmp
export TMPDIR
ME=`readlink -f $0`
AUTHORS=`dirname $ME`/authors

mkdir racket && cd racket && git init
git remote add racket git://git.racket-lang.org/plt
git fetch --tags racket
git config  merge.renameLimit 10000
git svn init  --stdlayout svn://svn.debian.org/svn/pkg-plt-scheme/plt-scheme/ 
git svn fetch -A$AUTHORS
git branch debian

A couple points to note:

• At some point there were huge numbers of renames when then the project renamed itself, hense the setting for merge.renameLimit

• Note the use of an authors file to make sure the author names and emails are reasonable in the imported history.

• git svn creates a branch master, which we will eventually forcibly overwrite; we stash that branch as debian for later use.

Now a couple complications arose about upstream's git repo.

1. Upstream releases seperate source tarballs for unix, mac, and windows. Each of these is constructed by deleting a large number of files from version control, and occasionally some last minute fiddling with README files and so on.

2. The history of the release tags is not completely linear. For example,

rocinante:~/projects/racket  (git-svn)-[master]-% git diff --shortstat v4.2.4 `git merge-base v4.2.4 v5.0`
 48 files changed, 242 insertions(+), 393 deletions(-)

rocinante:~/projects/racket  (git-svn)-[master]-% git diff --shortstat v4.2.1 `git merge-base v4.2.1 v4.2.4`
 76 files changed, 642 insertions(+), 1485 deletions(-)

The combination made my straight forward attempt at constructing a history synched with release tarballs generate many conflicts. I ended up importing each tarball on a temporary branch, and the merges went smoother. Note also the use of "git merge -s recursive -X theirs" to resolve conflicts in favour of the new upstream version.

The repetitive bits of the merge are collected as shell functions.

import_tgz() { 
    if [ -f $1 ]; then 
        git clean -fxd; 
        git ls-files -z | xargs -0 rm -f; 
        tar --strip-components=1 -zxvf $1 ; 
        git add -A; 
        git commit -m'Importing '`basename $1`;
    else
        echo "missing tarball $1"; 
    fi; 
}

do_merge() {
    version=$1
    git checkout -b v$version-tarball v$version
    import_tgz ../plt-scheme_$version.orig.tar.gz
    git checkout upstream 
    git merge -s recursive -X theirs v$version-tarball
}

post_merge() {
    version=$1
    git tag -f upstream/$version
    pristine-tar commit ../plt-scheme_$version.orig.tar.gz
    git branch -d v$version-tarball
}

The entire merge script is here. A typical step looks like

do_merge 5.0
git rm collects/tests/stepper/automatic-tests.ss
git add `git status -s | egrep ^UA | cut -f2 -d' '`
git checkout v5.0-tarball doc/release-notes/teachpack/HISTORY.txt
git rm readme.txt
git add  collects/tests/web-server/info.rkt
git commit -m'Resolve conflicts from new upstream version 5.0'
post_merge 5.0

Finally, we have the comparatively easy task of merging the upstream and Debian branches. In one or two places git was confused by all of the copying and renaming of files and I had to manually fix things up with git rm.

cd racket || /bin/true
set -e

git checkout debian
git tag -f packaging/4.0.1-2 `git svn find-rev r98`
git tag -f packaging/4.2.1-1 `git svn find-rev r113`
git tag -f packaging/4.2.4-2 `git svn find-rev r126`

git branch -f  master upstream/4.0.1
git checkout master
git merge packaging/4.0.1-2
git tag -f debian/4.0.1-2

git merge upstream/4.2.1
git merge packaging/4.2.1-1
git tag -f debian/4.2.1-1

git merge upstream/4.2.4
git merge packaging/4.2.4-2
git rm collects/tests/stxclass/more-tests.ss && git commit -m'fix false rename detection'
git tag -f debian/4.2.4-2

git merge -s recursive -X theirs upstream/5.0
git rm collects/tests/web-server/info.rkt
git commit -m 'Merge upstream 5.0'

I'm thinking about distributed issue tracking systems that play nice with git. I don't care about other version control systems anymore :). I also prefer command line interfaces, because as commentators on the blog have mentioned, I'm a Luddite (in the imprecise, slang sense).

So far I have found a few projects, and tried to guess how much of a going concern they are.

Git Specific

• ticgit I don't know if this github at its best or worst, but the original project seems dormant and there are several forks. According the original author, this one is probably the best.

• git-issues Originally a rewrite of ticgit in python, it now claims to be defunct.

VCS Agnostic

• ditz Despite my not caring about other VCSs, ditz is VCS agnostic, just making files. Seems active.

• cil takes a similar approach to ditz, is written in Perl rather than Ruby, and should release again any day now (hint, hint).

• milli is a minimalist approach to the same theme.

Sortof VCS Agnostic

• bugs everywhere is written in python. Works with Arch, Bazaar, Darcs, Git, and Mercurial. There seems to some on-going development activity.

• simple defects has Git and Darcs integration. It seems active. It's written by bestpractical people, so no surprise it is written in Perl.

Updated

• 2010-10-01 Note activity for bugs everywhere
• 2012-06-22 Note git-issues self description as defunct. Update link for cil.

You have a gitolite install on host $MASTER, and you want a mirror on $SLAVE. Here is one way to do that. $CLIENT is your workstation, that need not be the same as $MASTER or $SLAVE.

1. On $CLIENT, install gitolite on $SLAVE. It is ok to re-use your gitolite admin key here, but make sure you have both public and private key in .ssh, or confusion ensues. Note that when gitolite asks you to double check the "host gitolite" ssh stanza, you probably want to change hostname to $SLAVE, at least temporarily (if not, at least the checkout of the gitolite-admin repo will fail) You may want to copy .gitolite.rc from $MASTER when gitolite fires up an editor.

2. On $CLIENT copy the "gitolite" stanza of .ssh/config to gitolite-mirror to a stanza called e.g. gitolite-slave fix the hostname of the gitolite stanza so it points to $MASTER again.

3. On $MASTER, as gitolite user, make passphraseless ssh-key. Probably you should call it something like 'mirror'

4. Still on $MASTER. Add a stanza like the following to $gitolite_user/.ssh/config

     host gitolite-mirror
       hostname $SLAVE
       identityfile ~/.ssh/mirror
   

   run ssh gitolite-mirror at least once to test and set up any "know_hosts" file.

5. On $CLIENT change directory to a checkout of gitolite admin from $MASTER. Make sure it is up to date with respect origin

    git pull
   

6. Edit .git/config (or, in very recent git, use git remote seturl --push --add) so that remote origin looks like

    fetch = +refs/heads/*:refs/remotes/origin/*
    url = gitolite:gitolite-admin
    pushurl = gitolite:gitolite-admin
    pushurl = gitolite-slave:gitolite-admin
   

7. Add a stanza

   repo @all
     RW+     = mirror
   

to the bottom of your gitolite.conf Add mirror.pub to keydir.

1. Now overwrite the gitolite-admin repo on $SLAVE

   git push -f

   Note that empty repos will be created on $SLAVE for every repo on $MASTER.

2. The following one line post-update hook to any repos you want mirrored (see the gitolite documentation for how to automate this) You should not modify the post update hook of the gitolite-admin repo.

   git push --mirror gitolite-mirror:$GL_REPO.git

3. Create repos as per normal in the gitolite-admin/conf/gitolite.conf. If you have set the auto post-update hook installation, then each repo will be mirrored. You should only push to $MASTER; any changes pushed to $SLAVE will be overwritten.

Fourth in a series (git-sync-experiments, git-sync-experiments2, git-sync-experiments3) of completely unscientific experiments to try and figure out the best way to sync many git repos.

I wanted to see how bundles worked, and if there was some potential for speedup versus mr. The following unoptimized script is about twice as fast as mr in updating 10 repos. Of course is not really doing exactly the right thing (since it only looks at HEAD), but it is a start maybe. Of course, maybe the performance difference has nothing to do with bundles. Anyway IPC::PerlSSH is nifty.

#!/usr/bin/perl

use strict;
use File::Slurp;
use IPC::PerlSSH;
use Git;

my %config;
eval(read_file('config.pl'));
die $@ if $@;


my $ips= IPC::PerlSSH->new(Host=>$config{host});

$ips->eval("use Git; use File::Temp qw (tempdir); use File::Slurp;");
$ips->eval('${main::tempdir}=tempdir();');

$ips->store( "bundle", 
             q{my $prefix=shift;
               my $name=shift;
               my $ref=shift;
                chomp($ref);
               my $repo=Git->repository($prefix.$name.'.git');
               my $bfile="${main::tempdir}/${name}.bundle";
               eval {$repo->command('bundle','create',
                             $bfile,
                                $ref.'..HEAD'); 1} 
                or do { return undef };
                my $bits=read_file($bfile);
                print STDERR ("got ",length($bits),"\n");
                return $bits;
}
);


foreach my $pair  (@{$config{repos}}){
    my ($local,$remote)=@{$pair};
    my $bname=$local.'.bundle';

    $bname =~ s|/|_|;
    $bname =~ s|^\.|@|;
    
    my $repo=Git->repository($config{localprefix}.$local);
    # force some commit to be bundled, just for testing
    my $head=$repo->command('rev-list','--max-count=1', 'origin/HEAD^');
    
    my $bits=$ips->call('bundle',$config{remoteprefix},$remote,$head);
    write_file($bname, {binmode => ':raw'}, \$bits);
    $repo->command_noisy('fetch',$ENV{PWD}.'/'.$bname,'HEAD');
}

The config file is just a hash

%config=(
    host=>'hostname',
    localprefix=>'/path/',
    remoteprefix=>'/path/git/'
    repos=>[
        [qw(localdir remotedir)],
    ]
)

So I have been getting used to madduck's workflow for topgit and debian packaging, and one thing that bugged me a bit was all the steps required to to build. I tend to build quite a lot when debugging, so I wrote up a quick and dirty script to

• export a copy of the master branch somewhere
• export the patches from topgit
• invoke debuild

I don't claim this is anywhere ready production quality, but maybe it helps someone.

Assumptions (that I remember)

• you use the workflow above
• you use pristine tar for your original tarballs
• you invoke the script (I call it tg-debuild) from somewhere in your work tree

Here is the actual script:

    #!/bin/sh

    set -x

    if [ x$1 = x-k ]; then
        keep=1
    else
        keep=0
    fi

    WORKROOT=/tmp
    WORKDIR=`mktemp -d $WORKROOT/tg-debuild-XXXX`
    # yes, this could be nicer
    SOURCEPKG=`dpkg-parsechangelog | grep ^Source: | sed 's/^Source:\s*//'`
    UPSTREAM=`dpkg-parsechangelog | grep ^Version: | sed -e 's/^Version:\s*//' -e s/-[^-]*//`
    ORIG=$WORKDIR/${SOURCEPKG}_${UPSTREAM}.orig.tar.gz

    pristine-tar checkout $ORIG

    WORKTREE=$WORKDIR/$SOURCEPKG-$UPSTREAM

    CDUP=`git rev-parse --show-cdup`
    GDPATH=$PWD/$CDUP/.git

    DEST=$PWD/$CDUP/../build-area

    git archive --prefix=$WORKTREE/ --format=tar master | tar xfP -
    GIT_DIR=$GDPATH make -C $WORKTREE -f debian/rules tg-export
    cd $WORKTREE && GIT_DIR=$GDPATH debuild 
    if [ $?==0 -a -d $DEST ]; then
        cp $WORKDIR/*.deb $WORKDIR/*.dsc $WORKDIR/*.diff.gz $WORKDIR/*.changes $DEST
    fi


    if [ $keep = 0 ]; then
        rm -fr $WORKDIR
    fi

Scenario

You are maintaining a debian package with topgit. You have a topgit patch against version k and it is has been merged into upstream version m. You want to "disable" the topgit branch, so that patches are not auto-generated, but you are not brave enough to just

   tg delete feature/foo

You are brave enough to follow the instructions of a random blog post.

Checking your patch has really been merged upstream

This assumes that you tags upstream/j for version j.

git checkout feature/foo
git diff upstream/k

For each file foo.c modified in the output about, have a look at

git diff upstream/m foo.c

This kindof has to be a manual process, because upstream could easily have modified your patch (e.g. formatting).

The semi-destructive way

Suppose you really never want to see that topgit branch again.

git update-ref -d refs/topbases/feature/foo
git checkout master
git branch -M feature/foo merged/foo

The non-destructive way.

After I worked out the above, I realized that all I had to do was make an explicit list of topgit branches that I wanted exported. One minor trick is that the setting seems to have to go before the include, like this

TG_BRANCHES=debian/bin-makefile debian/libtoolize-lib debian/test-makefile
-include /usr/share/topgit/tg2quilt.mk

Conclusions

I'm not really sure which approach is best yet. I'm going to start with the non-destructive one and see how that goes.

Updated Madduck points to a third, more sophisticated approach in Debian BTS.

I wanted to report a success story with topgit which is a rather new patch queue managment extension for git. If that sounds like gibberish to you, this is probably not the blog entry you are looking for.

Some time ago I decided to migrate the debian packaging of bibutils to topgit. This is not a very complicated package, with 7 quilt patches applied to upstream source. Since I don't have any experience to go on, I decided to follow Martin 'madduck' Krafft's suggestion for workflow.

It all looks a bit complicated (madduck will be the first to agree), but it forced me to think about which patches were intended to go upstream and which were not. At the end of the conversion I had 4 patches that were cleanly based on upstream, and (perhaps most importantly for lazy people like me), I could send them upstream with tg mail. I did that, and a few days later, Chris Putnam sent me a new upstream release incorporating all of those patches. Of course, now I have to package this new upstream release :-).

The astute reader might complain that this is more about me developing half-decent workflow, and Chris being a great guy, than about any specific tool. That may be true, but one thing I have discovered since I started using git is that tools that encourage good workflow are very nice. Actually, before I started using git, I didn't even use the word workflow. So I just wanted to give a public thank you to pasky for writing topgit and to madduck for pushing it into debian, and thinking about debian packaging with topgit.

Third in a series (git-sync-experiments, git-sync-experiments2) of completely unscientific experiments to try and figure out the best way to sync many git repos.

If you want to make many ssh connections to a given host, then the first thing you need to do is turn on multiplexing. See the ControlPath and ControlMaster options in ssh config

Presuming that is not fast enough, then one option is to make many parallel connections (see e.g. git-sync-experiments2). But this won't scale very far.

In this week I consider the possibilities of running a tunneled socket to a remote git-daemon

ssh  -L 9418:localhost:9418 git-host.domain.tld git-daemon --export-all  

Of course from a security point of view this is awful, but I did it anyway, at least temporarily.

Running my "usual" test of git pull in 15 up-to-date repos, I get 3.7s versus about 5s with the multiplexing. So, 20% improvement, probably not worth the trouble. In both cases I just run a shell script like

  cd repo1 && git pull && cd ..
  cd repo2 && git pull && cd ..
  cd repo3 && git pull && cd ..
  cd repo4 && git pull && cd ..
  cd repo5 && git pull && cd ..

In a previous post I complained that mr was too slow. madduck pointed me to the "-j" flag, which runs updates in parallel. With -j 5, my 11 repos update in 1.2s, so this is probably good enough to put this project on the back burner until I get annoyed again.

I have the feeling that the "right solution" (TM) involves running either git-daemon or something like it on the remote host. The concept would be to set up a pair of file descriptors connected via ssh to the remote git-daemon, and have your local git commands talk to that pair of file descriptors instead of a socket. Alas, that looks like a bit of work to do, if it is even possible.

I have been thinking about ways to speed multiple remote git on the same hosts. My starting point is mr, which does the job, but is a bit slow. I am thinking about giving up some generality for some speed. In particular it seems like it ought to be possible to optimize for the two following use cases:

• many repos are on the same host
• mostly nothing needs updating.

For my needs, mr is almost fast enough, but I can see it getting annoying as I add repos (I currently have 11, and mr update takes about 5 seconds; I am already running ssh multiplexing). I am also thinking about the needs of the Debian Perl Modules Team, which would have over 900 git repos if the current setup was converted to one git repo per module.

My first attempt, using perl module Net::SSH::Expect to keep an ssh channel open can be scientifically classified as "utter fail", since Net::SSH::Expect takes about 1 second to round trip "/bin/true".

Initial experiments using IPC::PerlSSH are more promising. The following script grabs the head commit in 11 repos in about 0.5 seconds. Of course, it still doesn't do anything useful, but I thought I would toss this out there in case there already exists a solution to this problem I don't know about.

 
#!/usr/bin/perl

use IPC::PerlSSH;
use Getopt::Std;
use File::Slurp;


my %config; 

eval( "\%config=(".read_file(shift(@ARGV)).")");
die "reading configuration failed: $@" if $@;


my $ips= IPC::PerlSSH->new(Host=>$config{host});

$ips->eval("use Git");

$ips->store( "ls_remote", q{my $repo=shift;
                       return Git::command_oneline('ls-remote',$repo,'HEAD');
                          } );

foreach $repo (@{$config{repos}}){
    print $ips->call("ls_remote",$repo);
}

P.S. If you google for "mr joey hess", you will find a Kiss tribute band called Mr. Speed, started by Joe Hess"

P.P.S. Hello planet debian!

To convert an svn repository containing only "/debian" to something compatible with git-buildpackage, you need to some work. Luckily zack already figured out how.

#

package=bibutils
version=3.40
mkdir $package
cd $package
git-svn init --stdlayout --no-metadata svn://svn.debian.org/debian-science/$package
git-svn fetch
# drop upstream branch from svn
git-branch -d -r upstream

# create a new upstream branch based on recipe from  zack
# 
git-symbolic-ref HEAD refs/heads/upstream
git rm --cached -r .
git commit --allow-empty -m 'initial upstream branch'
git checkout -f master
git merge upstream
git-import-orig --pristine-tar --no-dch ../tarballs/${package}_${version}.orig.tar.gz

If you forget to use --authors-file=file then you can fix up your mistakes later with something like the following. Note that after some has cloned your repo, this makes life difficult for them.

#!/bin/sh

project=vrr
name="David Bremner"
email="bremner@unb.ca"

git clone alioth.debian.org:/git/debian-science/packages/$project $project.new
cd $project.new
git branch upstream origin/upstream
git branch pristine-tar origin/pristine-tar
git-filter-branch --env-filter "export GIT_AUTHOR_EMAIL='bremner@unb.ca' GIT_AUTHOR_NAME='David Bremner'" master upstream pristine-tar

I am in the process of migrating (to git) some debian packages from a subversion repository created with svn-inject -l 2, namely

/trunk/package 
/branches/upstream/package/ 
/tags/package

Here is a script I wrote that seems to do the trick

#!/bin/sh

package=$1
stage=$1.from-svn
set -x
# my debian packages live under $SVNROOT/debian, with layout 2
mkdir $stage
cd $stage
git-svn init --no-metadata \
    --trunk $SVNROOT/debian/trunk/$package \
    --branches $SVNROOT/debian/branches/upstream/$package \
    --tags $SVNROOT/debian/tags/$package 
git-svn fetch 
git branch -r upstream current
 cd ..
# git clone --bare loses some gunk from git-svn. Anyway we need a bare repo
git clone --bare $stage $1.git

rm -rf $stage

Your mileage may vary of course.

UPDATED Apparently 'git branch -r upstream current' no longer works, if it ever did. If anyone can psychically figure out what I wanted to do there, I'm happy to translate that into git.