Advanced shared repository layouts

Bazaar is designed to give you flexibility in how you layout branches inside a shared repository. This flexibility allows users to tailor Bazaar to their workflow, but it also leads to questions about what is a “good” layout. We present some alternatives and give some discussion about the benefits of each.

One key point which should be mentioned is that any good layout should somehow highlight what branch a “general” user should grab. In SVN this is deemed the “trunk/” branch, and in most of the layouts this naming convention is preserved. Some would call this “mainline” or “dev”, and people from CVS often refer to this as “HEAD”.

“SVN-Style” (trunk/, branches/)

Most people coming from SVN will be familiar with their “standard” project layout. Which is to layout the repository as:

repository/       # Overall repository
 +- trunk/        # The mainline of development
 +- branches/     # A container directory
 |   +- foo/      # Branch for developing feature foo
 |     ...
 +- tags/         # Container directory
     +- release-X # A branch specific to mark a given release version

With Bazaar, that is a perfectly reasonable layout. It has the benefit of being familiar to people coming from SVN, and making it clear where the development focus is.

When you have multiple projects in the same repository, the SVN layout is a little unclear what to do.


The preferred method for SVN seems to be to give each project a top level directory for a layout like:

repository/            # Overall repository
 +- project1/          # A container directory
 |   +- trunk/         # The mainline of development of project1
 |   +- branches/      # A container directory
 |       +- foo/       # Branch for developing feature foo of project1
 |         ...
 +- project2/          # Container for project2
     +- trunk/         # Mainline for project2
     +- branches/      # Container for project2 branches

This also works with Bazaar. However, with Bazaar repositories are cheap to create (a simple bzr init-repo away), and their primary benefit is when the branches share a common ancestry.

So the preferred way for Bazaar would be:

project1/          # A repository for project1
 +- trunk/         # The mainline of development of project1
 +- branches/      # A container directory
     +- foo/       # Branch for developing feature foo of project1

project2/          # A repository for project2
 +- trunk/         # Mainline for project2
 +- branches/      # Container for project2 branches


There are also a few projects who use this layout in SVN:

repository/             # Overall repository
  +- trunk/             # A container directory
  |   +- project1       # Mainline for project 1
  |   +- project2       # Mainline for project 2
  |         ...
  +- branches/          # Container
      +- project1/      # Container (?)
      |   +- foo        # Branch 'foo' of project1
      +- project2/
          +- bar        # Branch 'bar' of project2

A slight variant is:

repository/             # Overall repository
  +- trunk/             # A container directory
  |   +- project1       # Mainline for project 1
  |   +- project2       # Mainline for project 2
  |         ...
  +- branches/          # Container
      +- project1-foo/  # Branch 'foo' of project1
      +- project2-bar/  # Branch 'bar' of project2

I believe the reason for this in SVN, is so that someone can checkout all of “trunk/” and get the all the mainlines for all projects.

This layout can be used for Bazaar, but it is not generally recommended.

  1. bzr branch/checkout/get is a single branch at a time. So you don’t get the benefit of getting all mainlines with a single command. [1]
  2. It is less obvious of whether repository/trunk/foo is the trunk of project foo or it is just the foo directory in the trunk branch. Some of this confusion is due to SVN, because it uses the same “namespace” for files in a project that it uses for branches of a project. In Bazaar, there is a clear distinction of what files make up a project, versus the location of the Branch. (After all, there is only one .bzr/ directory per branch, versus many .svn/ directories in the checkout).
[1]Note: NestedTreeSupport can provide a way to create “meta-projects” which aggregate multiple projects regardless of the repository layout. Letting you bzr checkout one project, and have it grab all the necessary sub-projects.

Nested Style (project/branch/sub-branch/)

Another style with Bazaar, which is not generally possible in SVN is to have branches nested within each-other. This is possible because Bazaar supports (and recommends) creating repositories with no working trees (--no-trees). With a --no-trees repository, because the working files are not intermixed with your branch locations, you are free to put a branch in whatever namespace you want.

One possibility is:

project/             # The overall repository, *and* the project's mainline branch
 + joe/              # Developer Joe's primary branch of development
 |  +- feature1/     # Developer Joe's feature1 development branch
 |  |   +- broken/   # A staging branch for Joe to develop feature1
 |  +- feature2/     # Joe's feature2 development branch
 |    ...
 + barry/            # Barry's development branch
 |  ...
 + releases/
    +- 1.0/
        +- 1.1.1/

The idea with this layout is that you are creating a hierarchical layout for branches. Where changes generally flow upwards in the namespace. It also gives people a little corner of the namespace to work on their stuff. One nice feature of this layout, is it makes branching “cheaper” because it gives you a place to put all the mini branches without cluttering up the global branches/ namespace.

The other power of this is that you don’t have to repeat yourself when specifying more detail in the branch name.

For example compare:

bzr branch http://host/repository/project/branches/joe-feature-foo-bugfix-10/


bzr branch http://host/project/joe/foo/bugfix-10

Also, if you list the repository/project/branches/ directory you might see something like:


Versus having these broken out by developer. If the number of branches are small, branches/ has the nice advantage of being able to see all branches in a single view. If the number of branches is large, branches/ has the distinct disadvantage of seeing all the branches in a single view (it becomes difficult to find the branch you are interested in, when there are 100 branches to look through).

Nested branching seems to scale better to larger number of branches. However, each individual branch is less discoverable. (eg. “Is Joe working on bugfix 10 in his feature foo branch, or his feature bar branch?”)

One other small advantage is that you can do something like:

 bzr branch http://host/project/release/1/1/1
 bzr branch http://host/project/release/1/1/2

To indicate release 1.1.1 and 1.1.2. This again depends on how many releases you have and whether the gain of splitting things up outweighs the ability to see more at a glance.

Sorted by Status (dev/, merged/, experimental/)

One other way to break up branches is to sort them by their current status. So you would end up with a layout something like:

project/               # Overall layout
 +- trunk/             # The development focus branch
 +- dev/               # Container directory for in-progress work
 |   +- joe-feature1   # Joe's current feature-1 branch
 |   +- barry-bugfix10 # Barry's work for bugfix 10
 |    ...
 +- merged/            # Container indicating these branches have been merged
 |   +- bugfix-12      # Bugfix which has already been merged.
 +- abandonded/        # Branches which are considered 'dead-end'

This has a couple benefits and drawbacks. It lets you see what branches are actively being developed on, which is usually only a small number, versus the total number of branches ever created. Old branches are not lost (versus deleting them), but they are “filed away”, such that the more likely you are to want a branch the easier it is to find. (Conversely, older branches are likely to be harder to find).

The biggest disadvantage with this layout, is that branches move around. Which means that if someone is following the project/dev/new-feature branch, when it gets merged into trunk/ suddenly bzr pull doesn’t mirror the branch for them anymore because the branch is now at project/merged/new-feature. There are a couple ways around this. One is to use HTTP redirects to point people requesting the old branch to the new branch. bzr >= 0.15 will let users know that http://old/path redirects to http://new/path. However, this doesn’t help if people are accessing a branch through methods other than HTTP (SFTP, local filesystem, etc).

It would also be possible to use a symlink for temporary redirecting (as long as the symlink is within the repository it should cause little trouble). However eventually you want to remove the symlink, or you don’t get the clutter reduction benefit. Another possibility instead of a symlink is to use a BranchReference. It is currently difficult to create these through the bzr command line, but if people find them useful that could be changed. This is actually how Launchpad allows you to bzr checkout Effectively a BranchReference is a symlink, but it allows you to reference any other URL. If it is extended to support relative references, it would even work over HTTP, SFTP, and local paths.

Sorted by date/release/etc (2006-06/, 2006-07/, 0.8/, 0.9)

Another method of allowing some scalability while also allowing the browsing of “current” branches. Basically, this works on the assumption that actively developed branches will be “new” branches, and older branches are either merged or abandoned.

Basically the date layout looks something like:

project/                # Overall project repository
 +- trunk/              # General mainline
 +- 2006-06/            # containing directory for branches created in this month
 |   +- feature1/       # Branch of "project" for "feature1"
 |   +- feature2/       # Branch of "project" for "feature2"
 +- 2005-05/            # Containing directory for branches create in a different month
     +- feature3/

This answers the question “Where should I put my new branch?” very quickly. If a feature is developed for a long time, it is even reasonable to copy a branch into the newest date, and continue working on it there. Finding an active branch generally means going to the newest date, and going backwards from there. (A small disadvantage is that most directory listings sort oldest to the top, which may mean more scrolling). If you don’t copy old branches to newer locations, it also has the disadvantage that searching for a branch may take a while.

Another variant is by release target:

project/          # Overall repository
 +- trunk/        # Mainline development branch
 +- releases/     # Container for release branches
 |   +- 0.8/      # The branch for release 0.8
 |   +- 0.9/      # The branch for release 0.9
 +- 0.8/          # Container for branches targeting release 0.8
 |   +- feature1/ # Branch for "feature1" which is intended to be merged into 0.8
 |   +- feature2/ # Branch for "feature2" which is targeted for 0.8
 +- 0.9/
     +- feature3/ # Branch for "feature3", targeted for release 0.9

Some possible variants include having the 0.9 directory imply that it is branched from 0.9 rather than for 0.9, or having the 0.8/release as the official release 0.8 branch.

The general idea is that by targeting a release, you can look at what branches are waiting to be merged. It doesn’t necessarily give you a good idea of what the state of the branch (is it in development or finished awaiting review). It also has a history-hiding effect, and otherwise has the same benefits and deficits as a date-based sorting.

Simple developer naming (project/joe/foo, project/barry/bar)

Another possibly layout is to give each developer a directory, and then have a single sub-directory for branches. Something like:

project/      # Overall repository
 +- trunk/    # Mainline branch
 +- joe/      # A container for Joe's branches
 |   +- foo/  # Joe's "foo" branch of "project"
 +- barry/
     +- bar/  # Barry's "bar" branch of "project"

The idea is that no branch is “nested” underneath another one, just that each developer has his/her branches grouped together.

A variant which is used by Launchpad is:

 +- joe/             # Joe's branches
 |   +- project1/    # Container for Joe's branches of "project1"
 |   |   +- foo/     # Joe's "foo" branch of "project1"
 |   +- project2/    # Container for Joe's "project2" branches
 |       +- bar/     # Joe's "bar" branch of "project2"
 |        ...
 +- barry/
 |   +- project1/    # Container for Barry's branches of "project1"
 |       +- bug-10/  # Barry's "bug-10" branch of "project1"
 |   ...
 +- group/
     +- project1/
         +- trunk/   # The main development focus for "project1"

This lets you easily browse what each developer is working on. Focus branches are kept in a “group” directory, which lets you see what branches the “group” is working on.

This keeps different people’s work separated from each-other, but also makes it hard to find “all branches for project X”. Launchpad compensates for this by providing a nice web interface with a database back end, which allows a “view” to be put on top of this layout. This is closer to the model of people’s home pages, where each person has a “~/public_html” directory where they can publish their own web-pages. In general, though, when you are creating a shared repository for centralization of a project, you don’t want to split it up by person and then project. Usually you would want to split it up by project and then by person.


In the end, no single naming scheme will work for everyone. It depends a lot on the number of developers, how often you create a new branch, what sort of lifecycles your branches go through. Some questions to ask yourself:

  1. Do you create a few long-lived branches, or do you create lots of “mini” feature branches (Along with this is: Would you like to create lots of mini feature branches, but can’t because they are a pain in your current VCS?)
  2. Are you a single developer, or a large team?
  3. If a team, do you plan on generally having everyone working on the same branch at the same time? Or will you have a “stable” branch that people are expected to track.