Moving to Subversion

CVS has a number of limitations that have been eliminated by Subversion. For the development of Python, the most notable improvements are:

• the ability to rename files and directories, and to remove directories, while keeping the history of these files.
• support for change sets (sets of correlated changes to multiple files) through global revision numbers. Change sets are transactional.
• atomic, fast tagging: a cvs tag might take many minutes; a Subversion tag (svn cp) will complete quickly, and atomically. Likewise, branches are very efficient.
• support for offline diffs, which is useful when creating patches.

Moving to python.org

SourceForge has kindly provided an important infrastructure for the past years. Unfortunately, the attention that SF received has also caused repeated overload situations in the past, to which the SF operators could not always respond in a timely manner. In particular, for CVS, they had to reduce the load on the primary CVS server by introducing a second, read-only CVS server for anonymous access. This server is regularly synchronized, but lags behind the the read-write CVS repository between synchronizations. As a result, users without commit access can see recent changes to the repository only after a delay.

On python.org, it would be possible to make the repository accessible for anonymous access.

Collect SSH keys

After some discussion, svn+ssh was selected as the best method for write access to the repository. Developers can continue to use their SSH keys, but they must be installed on python.org.

In order to avoid having to create a new Unix user for each developer, a single account should be used, with command= attributes in the authorized_keys files.

The lines in the authorized_keys file should read like this (wrapped for better readability):

command="/usr/bin/svnserve --root=/svnroot -t 
--tunnel-user='<username>'",no-port-forwarding,
no-X11-forwarding,no-agent-forwarding,no-pty
ssh-dss <key> <comment>

As the usernames, the real names should be used instead of the SF account names, so that people can be better identified in log messages.

Administrator Access

Administrator access to the pythondev account should be granted to all current admins of the Python SF project. To distinguish between shell login and svnserve login, admins need to maintain two keys. Using OpenSSH, the following procedure can be used to create a second key:

cd .ssh
ssh-keygen -t DSA -f pythondev -C <user>@pythondev
vi config

In the config file, the following lines need to be added:

Host pythondev
  Hostname dinsdale.python.org
  User pythondev
  IdentityFile ~/.ssh/pythondev

Then, shell login becomes possible through "ssh pythondev".

Downloading the CVS Repository

The CVS repository can be downloaded from

http://cvs.sourceforge.net/cvstarballs/python-cvsroot.tar.bz2

Since this tarball is generated only once a day, some time must pass after the repository freeze before the tarball can be picked up. It should be verified that the last commit, as recorded on the python-commits mailing list, is indeed included in the tarball.

After the conversion, the converted CVS tarball should be kept forever on www.python.org/archive/python-cvsroot-<date>.tar.bz2

Converting the CVS Repository

The Python CVS repository contains two modules: distutils and python. The python module is further structured into dist and nondist, where dist only contains src (the python code proper). nondist contains various subdirectories.

These should be reorganized in the Subversion repository to get shorter URLs, following the <project>/{trunk,tags,branches} structure. A project will be created for each nondist directory, plus for src (called python), plus distutils. Reorganizing the repository is best done in the CVS tree, as shown below.

The fsfs backend should be used as the repository format (which requires Subversion 1.1). The fsfs backend has the advantage of being more backup-friendly, as it allows incremental repository backups, without requiring any dump commands to be run.

The conversion should be done using the cvs2svn utility, available e.g. in the cvs2svn Debian package. As cvs2svn does not currently support the project/trunk structure, each project needs to be converted separately. To get each conversion result into a separate directory in the target repository, svnadmin load must be used.

Subversion has a different view on binary-vs-text files than CVS. To correctly carry the CVS semantics forward, svn:eol-style should be set to native on all files that are not marked binary in the CVS.

In summary, the conversion script is:

#!/bin/sh
rm cvs2svn-*
rm -rf python py.new
tar xjf python-cvsroot.tar.bz2
rm -rf python/CVSROOT
svnadmin create --fs-type fsfs py.new
mv python/python python/orig
mv python/orig/dist/src python/python
mv python/orig/nondist/* python
# nondist/nondist is empty
rmdir python/nondist
rm -rf python/orig
for a in python/*
do
  b=`basename $a`
  cvs2svn -q --dump-only --encoding=latin1 --force-branch=cnri-16-start \
  --force-branch=descr-branch --force-branch=release152p1-patches \
  --force-tag=r16b1 $a
  svn mkdir -m"Conversion to SVN" file:///`pwd`/py.new/$b
  svnadmin load -q --parent-dir $b py.new < cvs2svn-dump
  rm cvs2svn-dump
done

Sample results of this conversion are available at

http://www.dcl.hpi.uni-potsdam.de/pysvn/

Publish the Repository

The repository should be published at http://svn.python.org/projects. Read-write access should be granted to all current SF committers through svn+ssh://pythondev@svn.python.org/; read-only anonymous access through WebDAV should also be granted.

As an option, websvn (available e.g. from the Debian websvn package) could be provided. Unfortunately, in the test installation, websvn breaks because it runs out of memory.

The current SF project admins should get write access to the authorized_keys2 file of the pythondev account.

Disable CVS

It appears that CVS cannot be disabled entirely. Only the user interface can be removed from the project page; the repository itself remains available. If desired, write access to the python and distutils modules can be disabled through a CVS commitinfo entry.