cat /dev/brain |

git: migrate SVN to git

published on Wednesday, June 28, 2017

Today's post covers the conversion process of large SVN repositories to git.

This is one of a series of posts where I want to share my experiences from converting the MAD-X SVN repository with roughly 6300 revisions and 370 branches and tags. So if you want to convert a large multi-branch/tag repository, keep reading.

If you're interested in methods to reduce the converted repository's weight before publishing the final version, you may also find git-du, git-unpack, git-dir2mod articles useful.

subgit vs svn2git

There are plenty of posts describing how to convert SVN repositories to git using git-svn or svn2git (which is based on the former). In my experience, however, these tools do not perform well for large SVN repositories, specifically if there are multiple branches.

I strongly recommend to use subgit instead. The main reasons to prefer this tool are:


git-svn and svn2git are extremely slow when working with multiple branches or tags: they read through the entire history multiple times. I cancelled the after about 6 hours into the process when it was still not finished.

On the other hand subgit went through the revisions only once and the whole import was finished in less than 5 minutes.

Branch structure

subgit seems to be better at correctly representing merge structures in git. For example, I had a branch that was merged like this:

o---o---o---o---o  trunk
     \       \
      A---o---B---o---o  next

for which the svn2git conversion failed to detect the merge and instead yielded a commit structure like this:

o---o---o---o---o  trunk
      A---o---B---o---o  next

The subgit generated commits looked as expected.


With git-svn, you have to manually create a list of authors by accessing the SVN log (as described in the linked posts and the svn2git README). subgit does this job for you and creates one in the REPO/subgit/authors.txt directory – one less thing to worry about.

Branch detection

subgit seems to be pretty good at auto-detecting branches in the branches/ folder – even for inconsistent/non-standard folder layout. For example, the repository that I converted had the actual source code living in trunk/madX and some of the branches had the code in branches/foo/madX others directly under branches/bar. subgit was able to correctly deduce the branches paths by using the --layout auto option.

Mirror the SVN repo

Let's talk business. Before importing the SVN repository to git, it is useful to create a mirror of the repository on the local machine in order to improve performance of subsequent accesses.

This can be done using at least three basic methods, with a strong preference on the third one:

1. dump using svnadmin

If you have direct access to the host from which the SVN repository is served, you can dump it and then import:

ssh HOST svnadmin PATH > dump
svnadmin create repo-svn
svnadmin load repo-svn < dump

2. dump using rsvndump

(not recommended)

If you don't have direct access, you can do the same using rsvndump with the root address of the repository:

rsvndump SVN_URL > dump
svnadmin create repo-svn
svnadmin load repo-svn < dump

However, note that for some reason the rsvndump operation needed about 16GiB in my /tmp when doing this. If you need to do this, be prepared to switch on additional swaps and remount your /tmp with more space or directly mount /tmp from disc.

3. synchronize using svnsync

This is the preferred option.

svnadmin create repo-svn
cd repo-svn
echo '#!/bin/sh' > hooks/pre-revprop-change
chmod +x hooks/pre-revprop-change
svnsync init file://`pwd` SVN_URL
svnsync sync file://`pwd`
cd -

Note that with this method you can easily pull new commits on the SVN upstream into your mirror by re-issueing the final svnsync sync command.


Finally, make the SVN mirror accessible on svn://localhost:3030:

svnserve --root repo-svn --listen-port 3030 -d

Subgit migration

Use subgit to convert your local SVN mirror into a git repository:

subgit configure svn://localhost:3030 repo-git --trunk trunk/madX --layout auto

Before continuing, inspect the file repo-git/subgit/config and fill repo-git/subgit/authors.txt with the correct names and email addresses corresponding to the SVN users.

We can now import the commits using the command:

subgit install repo-git

Note that subgit will continue to synchronize the resulting repo-git git repository. This means that before further working with this repository you should either subgit uninstall repo-git – or better clone it to a new location where your modifications will stick:

git clone --mirror repo-git repo-clone

You are now free to work on the repository in repo-clone. This is useful if you e.g. need to apply git filter-branch to make final adjustments before publishing the git repository.

Add revision numbers

You can add revision numbers to the commit messages like so:

cd repo-clone
git filter-branch --msg-filter '
    REV=$(git log --format="%N" $GIT_COMMIT -1 | cut -d" " -f1)
    echo -n "$REV: "
' -- --branches --tags

Polish before publish

In order to cut the repository size from about 380MiB to just over 100MiB, we performed a few additional operations on the MAD-X repository, making use of the git-filter-tree module to implement these operations efficiently:

git clone

For more info stick to the posts linked at the top of the page.

While you may have very different requirements this is to give you some ideas what can be done:

  • remove some PDF files using:
python ../git-filter-tree/git_filter_tree rm \
    doc/latexuguide/madxuguide.pdf \
    ... \
    doc/usrguide/reports/ \
    -- --branches --tags
  • unpack .gz ascii files:
python ../git-filter-tree/git_filter_tree unpack \
    -- --branches --tags
  • convert subdirectory into submodule:
./git-filter-tree/git-dir2mod \
    repo-clone subdir \
    repo-parent repo-submod

Migrating issues

I looked at a few alternative tools for migrating issues. In the end I settled for trac-hub for the following reasons:

  • support for attachments
  • support for labels
  • easy to modify

In fact, the last point turned out to be the key strength by far. I ended up patching most of the system for my own needs and completely changing the logic. The changes are now all integrated in the upstream so you don't have to worry about which one to choose.

The tool now uses githubs new issue import API which allows to perform the whole process…

  • …without hitting abuse detection warnings and getting blocked
  • …without sending email notifications
  • …without increasing your contribution count for each migrated issue
  • …much faster than with the normal issues API
  • …with correct creation/closed date set
  • …atomically without users being able to interfere in the creation of any single issue

For MAD-X, I decided to put all comments on an issue in the body of the initial post to make it more concise and readable and avoid putting my name on every message. For an example of how this looks, see e.g. here. This behaviour can be selected by using the -S flag.

In order to use the tool, you should first create a mapping of the revision numbers to commit ids as follows:

cd repo-clone
git update-ref refs/notes/commits refs/svn/map
git log --format="%H %s" --branches --tags \
    | cut -d':' -f1 | awk '{print $2 " " $1}' > revmap.txt

Now clone the tool from github:

git clone

You can use the script in tools/ as an example of how to download attachments from the trac system and then e.g. create a git repository containing all the attachments.

Next, create a config file that defines access to your trac database and your github user / token:

cd trac-hub
cp config.yaml.example config.yaml

When you're done, execute as follows:

bundle install --path vendor/bundle
bundle exec trac-hub \
    -c config.yaml -r ../revmap.txt -s 1 \

While importing, I recommend to temporarily limit repository interactions to collaborators only in the repositories settings.

This entry was tagged git, git-svn, migration, subgit, svn and svn2git

blog comments powered by Disqus