Aug 23, 2019 - Yorick Peterse    

Update: Why we're using a single codebase for GitLab Community and Enterprise editions

Dive deep into our decision to switch GitLab over to a single codebase as we review some of the benefits and challenges.

Join us at our first ever user conferences!

In "GitLab might move to a single Rails codebase" we announced that GitLab might move to using a single codebase for GitLab Community Edition (CE) and GitLab Enterprise Edition (EE). Since then we have decided to continue moving toward a single codebase. In this article I highlight some of the challenges, required work, and steps remaining to complete the switch.

Why use a single codebase?

For years, CE and EE used two different repositories for the Rails application. By using separate repositories we could separate proprietary code from code that is free software. On the surface this seems like a good idea for different reasons (e.g., licensing), but over the years the drawbacks began to outweigh the benefits.

We mention some of these drawbacks in a previous article, but more or less they all come down to the same core problem: It made the development process more complex than necessary. For example, we ended up with around 150 merge requests spread across CE and EE for a security release from several months ago. While the process of merging these merge requests is automated, we ran into a variety of issues (e.g. failing tests) that required manual intervention. We could have reduced the number of merge requests by half if we used a single repository, creating less work for developers and release managers.

Toward the end of 2018 I felt that we were running out of time and had to do something about the separation of CE and EE. We had always tried to avoid merging the two repositories due to the complexity and time involved, but it started to become more and more clear we had no other option. Marin Jankovski, Delivery engineering manager, and I made a plan to merge the two repositories. Marin wrote a design document that outlined the details of it all. The design document showed what challenges we faced, and gathered the critical support required for the largest engineering projects at GitLab to date.

Working toward a single codebase

Moving to a single codebase is not something we can do overnight for a project the size of GitLab. Workflows must be adapted, developers need to adjust to the new setup, and automation requires extensive changes.

One of the biggest challenges from an engineering perspective was to come up with a way to transparently remove proprietary code from GitLab when building a CE release. A naive approach might involve a script that removes known bits of proprietary code. While this might work for small projects that don't change often, this was not going to work for a project the size of GitLab.

Ruby provides us with a solution to this problem. In Ruby, you can create a module and inject it into another module or class. Once injected, the functionality of the module becomes available to the target module or class. This is best illustrated with a simple example:

class Person
  def initialize(name)
    @name = name
  end

  def name
    @name
  end
end

module Greet
  def greet
    "Hello #{name}"
  end
end

Person.include(Greet)

alice = Person.new('Alice')

alice.greet # => "Hello Alice"

Here we define a class Person, followed by a module that is used to create a message greeting a person. Next, we include it into the Person class, at which point we can use the module's methods for instances of the Person class. The result is the message "Hello Alice."

While this example is not exciting, using a setup like this allows us to move proprietary code to separate modules, and inject these modules when GitLab EE is used. For GitLab CE, we would remove these modules, and the code injecting these modules would have to disable itself transparently and automatically.

GitLab EE has been using this setup since late 2016 with all EE modules residing in a separate "ee" directory, but in a limited number of places. This meant that in some places EE and CE code got mixed together, while in other places the two are separate. For example, we had code like this:

 def lfs_upload_access?
   return false unless project.lfs_enabled?
   return false unless has_authentication_ability?(:push_code)
+  return false if project.above_size_limit? || objects_exceed_repo_limit?

   lfs_deploy_token? || can?(user, :push_code, project)
 end

Here EE added a line into an existing method without using a separate module, making it difficult to remove the EE-specific code when for CE.

Before we could move to a single codebase, we had to separate EE-specific code from code shared between CE and EE. Due to the amount of work necessary, we divided the work into two departments: backend and frontend. For every department we created issues outlining the work to do for the various parts of the codebase. We even included the exact lines of code that had to change directly in the created issues, making it simple to see what one had to do. Each department also had an engineer assigned as the lead engineer, responsible for taking on the most difficult challenges. Filipa Lacerda, senior frontend engineer of Verify (CI) and Delivery, was in charge of frontend code. As the Delivery backend engineer, I myself was in charge of backend code.

Some changes were small and took a short amount of time, with others were big and took weeks. One of my big challenges was to make sure CE and EE use the same database schema, changing just under 24,000 lines of code over a two-month period.

In total the work involved 55 different engineers submitting more than 600 merge requests, closing just under 400 issues, and changing nearly 1.5 million lines of code

Filipa spent a lot of time creating 168 frontend issues outlining specific tasks as well as submitting 124 merge requests to address the majority of these issues. Resolving some of these issues required getting rid of some technical debt first, such as breaking up large chunks of code into smaller chunks, and coming up with a way to create EE-specific Vue.js templates.

While Filipa and I took on the biggest challenges, in total the work involved 55 different engineers submitting more than 600 merge requests, closing just under 400 issues, and changing nearly 1.5 million lines of code.

Moving toward a single codebase

With most of the work done, we could start looking into what project setup we would use for a single codebase. We came up with three different approaches:

1. Moving all development into gitlab-ce

All code and development is moved into the gitlab-ce repository. The gitlab-ee repository is archived, and a separate repository is set up as a mirror of gitlab-ce, called gitlab-foss. Proprietary code is removed from this mirror automatically.

Since most of GitLab's development takes place in the current gitlab-ce repository, this setup would reduce the number of issues to move as well as merge requests to close. A downside of this approach is that clones of the gitlab-ce repository will include proprietary code.

2. Moving all development into gitlab-ee

All code and development is moved into the gitlab-ee repository. The gitlab-ce repository remains as is in terms of code, and will become a mirror of gitlab-ee. Like the first option, proprietary code is removed from this mirror automatically.

This setup means that users cloning gitlab-ce don't end up with proprietary code in their copy of gitlab-ce.

3. Moving all development into a new repository

We set up an entirely new repository called "gitlab," and move all code and development into this repository. The gitlab-ce and gitlab-ee repositories will become read-only. A mirror is set up (called "gitlab-foss") that mirrors the new "gitlab" repository, without including proprietary code.

Deciding which approach to take

Having evaluated all the benefits and drawbacks, we decided to go with option two: move development into gitlab-ee. This approach has several benefits:

  1. The code of the gitlab-ce repository remains as is, and won't include any proprietary code.
  2. We do not need a separate mirror repository that does not include proprietary code. Instead, we rename the gitlab-ce repository to "gitlab-foss." We are renaming the repository since having "gitlab" and "gitlab-ce" as project names could be confusing.
  3. Users building CE from source don't end up with proprietary code in their copy of the gitlab-ce repository.
  4. We keep the Git logs of both gitlab-ce and gitlab-ee, instead of losing the logs (this depends a bit on how we'd move repositories around).
  5. It requires the least amount of changes to our workflow and tooling.
  6. Using a single project and issue tracker for both CE and EE makes it easier to search for issues.

Issues created in the gitlab-ce project will move to the gitlab-ee project, which we will rename to just "gitlab" (or "gitlab-org/gitlab" if you include the group name). This project then becomes the single source of truth, and is used for creating issues for both the CE and EE distributions.

Moving merge requests across projects is not possible, so we will close any open merge requests. Authors of these merge requests will have to resubmit them to the "gitlab" (called "gitlab-ee" before the rename) project.

When moving issues or closing merge requests, a bot will also post a comment explaining why this is done, what steps the author of a merge request has to take, and where one might find more information about these procedures.

Prior to the single codebase setup, community contributions would be submitted to the gitlab-ce repository. In the single codebase, contributions are instead submitted to the new gitlab repository ("gitlab-org/gitlab"). EE-specific code resides in a "ee" directory in the repository. Code outside of this directory will be free and open source software, using the same license as the gitlab-ce repository currently uses. This means that as long as you do not change anything in this "ee" directory, the only change for community contributions is the use of a different repository.

Our current plan is to have a single codebase the first week of September. GitLab 12.3 will be the first release based on a single codebase.

Users that clone GitLab EE and/or GitLab CE from source should update their Git remote URLs after the projects are renamed. This is not strictly necessary as GitLab will redirect Git operations to the new repository. For users of our Omnibus packages and Docker images nothing changes.

Those interested in learning more about what went on behind the scenes can refer to the following resources:

Cover image from Unsplash