Repository storage paths

Introduced in GitLab 8.10.

GitLab allows you to define multiple repository storage paths (sometimes called storage shards) to distribute the storage load between several mount points.

Notes:

You must have at least one storage path called default.

The paths are defined in key-value pairs. The key is an arbitrary name you can pick to name the file path.

The target directories and any of its sub-paths must not be a symlink.

No target directory may be a sub-directory of another; no nesting.

Example: this is OK:

default:
  path: /mnt/git-storage-1
storage2:
  path: /mnt/git-storage-2

This is not OK because it nests storage paths:

default:
  path: /mnt/git-storage-1
storage2:
  path: /mnt/git-storage-1/git-storage-2 # <- NOT OK because of nesting

Configure GitLab

Warning: In order for backups to work correctly, the storage path must not be a mount point and the GitLab user should have correct permissions for the parent directory of the path. In Omnibus GitLab this is taken care of automatically, but for source installations you should be extra careful.

The thing is that for compatibility reasons gitlab.yml has a different structure than Omnibus. In gitlab.yml you indicate the path for the repositories, for example /home/git/repositories, while in Omnibus you indicate git_data_dirs, which for the example above would be /home/git. Then, Omnibus will create a repositories directory under that path to use with gitlab.yml.

This little detail matters because while restoring a backup, the current contents of /home/git/repositories are moved to /home/git/repositories.old, so if /home/git/repositories is the mount point, then mv would be moving things between mount points, and bad things could happen. Ideally, /home/git would be the mount point, so then things would be moving within the same mount point. This is guaranteed with Omnibus installations (because they don't specify the full repository path but the parent path), but not for source installations.

Now that you've read that big fat warning above, let's edit the configuration files and add the full paths of the alternative repository storage paths. In the example below, we add two more mount points that are named nfs_1 and nfs_2 respectively.

NOTE: Note: This example uses NFS. We do not recommend using EFS for storage as it may impact GitLab's performance. See the relevant documentation for more details.

For installations from source

Edit gitlab.yml and add the storage paths:

repositories:
  # Paths where repositories can be stored. Give the canonicalized absolute pathname.
  # NOTE: REPOS PATHS MUST NOT CONTAIN ANY SYMLINK!!!
  storages: # You must have at least a 'default' storage path.
    default:
      path: /home/git/repositories
    nfs_1:
      path: /mnt/nfs1/repositories
    nfs_2:
      path: /mnt/nfs2/repositories

Restart GitLab for the changes to take effect.

Note: The gitlab_shell: repos_path entry in gitlab.yml will be deprecated and replaced by repositories: storages in the future, so if you are upgrading from a version prior to 8.10, make sure to add the configuration as described in the step above. After you make the changes and confirm they are working, you can remove the repos_path line.

For Omnibus installations

Edit /etc/gitlab/gitlab.rb by appending the rest of the paths to the default one:

git_data_dirs({
  "default" => { "path" => "/var/opt/gitlab/git-data" },
  "nfs_1" => { "path" => "/mnt/nfs1/git-data" },
  "nfs_2" => { "path" => "/mnt/nfs2/git-data" }
})

Note that Omnibus stores the repositories in a repositories subdirectory of the git-data directory.

Choose where new repositories will be stored

Once you set the multiple storage paths, you can choose where new repositories will be stored under Admin Area > Settings > Repository > Repository storage > Storage nodes for new repositories.

Beginning with GitLab 8.13.4, multiple paths can be chosen. New repositories will be randomly placed on one of the selected paths.