9.2 Git and Other Systems - Migrating to Git

If you read the previous section about using git svn, you can easily use those instructions to git svn clone a repository; then, stop using the Subversion server, push to a new Git server, and start using that. If you want the history, you can accomplish that as quickly as you can pull the data out of the Subversion server (which may take a while).

However, the import isn’t perfect; and because it will take so long, you may as well do it right. The first problem is the author information. In Subversion, each person committing has a user on the system who is recorded in the commit information. The examples in the previous section show schacon in some places, such as the blame output and the git svn log. If you want to map this to better Git author data, you need a mapping from the Subversion users to the Git authors. Create a file called users.txt that has this mapping in a format like this:

To get a list of the author names that SVN uses, you can run this:

That generates the log output in XML format, then keeps only the lines with author information, discards duplicates, strips out the XML tags. Obviously this only works on a machine with grep, sort, and perl installed. Then, redirect that output into your users.txt file so you can add the equivalent Git user data next to each entry.

You can provide this file to git svn to help it map the author data more accurately. You can also tell git svn not to include the metadata that Subversion normally imports, by passing --no-metadata to the clone or init command. The metadata includes a git-svn-id inside each commit message that Git will generate during import. This can bloat your Git log and might make it a bit unclear.

This makes your import command look like this:

Now you should have a nicer Subversion import in your my_project directory. Instead of commits that look like this:

they look like this:

Not only does the Author field look a lot better, but the git-svn-id is no longer there, either.

You should also do a bit of post-import cleanup. For one thing, you should clean up the weird references that git svn set up. First you’ll move the tags so they’re actual tags rather than strange remote branches, and then you’ll move the rest of the branches so they’re local.

To move the tags to be proper Git tags, run:

This takes the references that were remote branches that started with refs/remotes/tags/ and makes them real (lightweight) tags.

Next, move the rest of the references under refs/remotes to be local branches:

It may happen that you’ll see some extra branches which are suffixed by @xxx (where xxx is a number), while in Subversion you only see one branch. This is actually a Subversion feature called “peg-revisions”, which is something that Git simply has no syntactical counterpart for. Hence, git svn simply adds the svn version number to the branch name just in the same way as you would have written it in svn to address the peg-revision of that branch. If you do not care anymore about the peg-revisions, simply remove them:

Now all the old branches are real Git branches and all the old tags are real Git tags.

There’s one last thing to clean up. Unfortunately, git svn creates an extra branch named trunk, which maps to Subversion’s default branch, but the trunk ref points to the same place as master. Since master is more idiomatically Git, here’s how to remove the extra branch:

The last thing to do is add your new Git server as a remote and push to it. Here is an example of adding your server as a remote:

Because you want all your branches and tags to go up, you can now run this:

All your branches and tags should be on your new Git server in a nice, clean import.

Since Mercurial and Git have fairly similar models for representing versions, and since Git is a bit more flexible, converting a repository from Mercurial to Git is fairly straightforward, using a tool called "hg-fast-export", which you’ll need a copy of:

The first step in the conversion is to get a full clone of the Mercurial repository you want to convert:

The next step is to create an author mapping file. Mercurial is a bit more forgiving than Git for what it will put in the author field for changesets, so this is a good time to clean house. Generating this is a one-line command in a bash shell:

This will take a few seconds, depending on how long your project’s history is, and afterwards the /tmp/authors file will look something like this:

In this example, the same person (Bob) has created changesets under four different names, one of which actually looks correct, and one of which would be completely invalid for a Git commit. Hg-fast-export lets us fix this by turning each line into a rule: "<input>"="<output>", mapping an <input> to an <output>. Inside the <input> and <output> strings, all escape sequences understood by the python string_escape encoding are supported. If the author mapping file does not contain a matching <input>, that author will be sent on to Git unmodified. If all the usernames look fine, we won’t need this file at all. In this example, we want our file to look like this:

The same kind of mapping file can be used to rename branches and tags when the Mercurial name is not allowed by Git.

The next step is to create our new Git repository, and run the export script:

The -r flag tells hg-fast-export where to find the Mercurial repository we want to convert, and the -A flag tells it where to find the author-mapping file (branch and tag mapping files are specified by the -B and -T flags respectively). The script parses Mercurial changesets and converts them into a script for Git’s "fast-import" feature (which we’ll discuss in detail a bit later on). This takes a bit (though it’s much faster than it would be over the network), and the output is fairly verbose:

That’s pretty much all there is to it. All of the Mercurial tags have been converted to Git tags, and Mercurial branches and bookmarks have been converted to Git branches. Now you’re ready to push the repository up to its new server-side home:

Bazaar is a DVCS tool much like Git, and as a result it’s pretty straightforward to convert a Bazaar repository into a Git one. To accomplish this, you’ll need to import the bzr-fastimport plugin.

The next system you’ll look at importing from is Perforce. As we discussed above, there are two ways to let Git and Perforce talk to each other: git-p4 and Perforce Git Fusion.

If your system isn’t one of the above, you should look for an importer online – quality importers are available for many other systems, including CVS, Clear Case, Visual Source Safe, even a directory of archives. If none of these tools works for you, you have a more obscure tool, or you otherwise need a more custom importing process, you should use git fast-import. This command reads simple instructions from stdin to write specific Git data. It’s much easier to create Git objects this way than to run the raw Git commands or try to write the raw objects (see Git Internals for more information). This way, you can write an import script that reads the necessary information out of the system you’re importing from and prints straightforward instructions to stdout. You can then run this program and pipe its output through git fast-import.

To quickly demonstrate, you’ll write a simple importer. Suppose you work in current, you back up your project by occasionally copying the directory into a time-stamped back_YYYY_MM_DD backup directory, and you want to import this into Git. Your directory structure looks like this:

In order to import a Git directory, you need to review how Git stores its data. As you may remember, Git is fundamentally a linked list of commit objects that point to a snapshot of content. All you have to do is tell fast-import what the content snapshots are, what commit data points to them, and the order they go in. Your strategy will be to go through the snapshots one at a time and create commits with the contents of each directory, linking each commit back to the previous one.

As we did in An Example Git-Enforced Policy, we’ll write this in Ruby, because it’s what we generally work with and it tends to be easy to read. You can write this example pretty easily in anything you’re familiar with – it just needs to print the appropriate information to stdout. And, if you are running on Windows, this means you’ll need to take special care to not introduce carriage returns at the end your lines – git fast-import is very particular about just wanting line feeds (LF) not the carriage return line feeds (CRLF) that Windows uses.

To begin, you’ll change into the target directory and identify every subdirectory, each of which is a snapshot that you want to import as a commit. You’ll change into each subdirectory and print the commands necessary to export it. Your basic main loop looks like this:

You run print_export inside each directory, which takes the manifest and mark of the previous snapshot and returns the manifest and mark of this one; that way, you can link them properly. “Mark” is the fast-import term for an identifier you give to a commit; as you create commits, you give each one a mark that you can use to link to it from other commits. So, the first thing to do in your print_export method is generate a mark from the directory name:

You’ll do this by creating an array of directories and using the index value as the mark, because a mark must be an integer. Your method looks like this:

Now that you have an integer representation of your commit, you need a date for the commit metadata. Because the date is expressed in the name of the directory, you’ll parse it out. The next line in your print_export file is:

where convert_dir_to_date is defined as:

That returns an integer value for the date of each directory. The last piece of meta-information you need for each commit is the committer data, which you hardcode in a global variable:

Now you’re ready to begin printing out the commit data for your importer. The initial information states that you’re defining a commit object and what branch it’s on, followed by the mark you’ve generated, the committer information and commit message, and then the previous commit, if any. The code looks like this:

You hardcode the time zone (-0700) because doing so is easy. If you’re importing from another system, you must specify the time zone as an offset. The commit message must be expressed in a special format:

The format consists of the word data, the size of the data to be read, a newline, and finally the data. Because you need to use the same format to specify the file contents later, you create a helper method, export_data:

All that’s left is to specify the file contents for each snapshot. This is easy, because you have each one in a directory – you can print out the deleteall command followed by the contents of each file in the directory. Git will then record each snapshot appropriately:

Note: Because many systems think of their revisions as changes from one commit to another, fast-import can also take commands with each commit to specify which files have been added, removed, or modified and what the new contents are. You could calculate the differences between snapshots and provide only this data, but doing so is more complex – you may as well give Git all the data and let it figure it out. If this is better suited to your data, check the fast-import man page for details about how to provide your data in this manner.

The format for listing the new file contents or specifying a modified file with the new contents is as follows:

Here, 644 is the mode (if you have executable files, you need to detect and specify 755 instead), and inline says you’ll list the contents immediately after this line. Your inline_data method looks like this:

You reuse the export_data method you defined earlier, because it’s the same as the way you specified your commit message data.

The last thing you need to do is to return the current mark so it can be passed to the next iteration:

That’s it. Here’s the script in its entirety:

If you run this script, you’ll get content that looks something like this:

To run the importer, pipe this output through git fast-import while in the Git directory you want to import into. You can create a new directory and then run git init in it for a starting point, and then run your script:

As you can see, when it completes successfully, it gives you a bunch of statistics about what it accomplished. In this case, you imported 13 objects total for 4 commits into 1 branch. Now, you can run git log to see your new history:

There you go – a nice, clean Git repository. It’s important to note that nothing is checked out – you don’t have any files in your working directory at first. To get them, you must reset your branch to where master is now:

You can do a lot more with the fast-import tool – handle different modes, binary data, multiple branches and merging, tags, progress indicators, and more. A number of examples of more complex scenarios are available in the contrib/fast-import directory of the Git source code.