Skip to content

Managing monorepos

Monorepos have become a regular part of development team workflows. While they have many advantages, monorepos can present performance challenges when using them in GitLab. Therefore, you should know:

  • What repository characteristics can impact performance.
  • Some tools and steps to optimize monorepos.

Impact on performance

Because GitLab is a Git-based system, it is subject to similar performance constraints as Git when it comes to large repositories that are gigabytes in size.

Monorepos can be large for many reasons.

Large repositories pose a performance risk when used in GitLab, especially if a large monorepo receives many clones or pushes a day, which is common for them.

Git itself has performance limitations when it comes to handling monorepos.

Monorepos can also impact notably on hardware, in some cases hitting limitations such as vertical scaling and network or disk bandwidth limits.

Gitaly is our Git storage service built on top of Git. This means that any limitations of Git are experienced in Gitaly, and in turn by end users of GitLab.

Optimize GitLab settings

You should use as many of the following strategies as possible to minimize fetches on the Gitaly server.


The most resource intensive operation in Git is the git-pack-objects process. It is responsible for figuring out all of the commit history and files to send back to the client.

The larger the repository, the more commits, files, branches, and tags that a repository has and the more expensive this operation is. Both memory and CPU are heavily utilized during this operation.

Most git clone or git fetch traffic (which results in starting a git-pack-objects process on the server) often come from automated continuous integration systems such as GitLab CI/CD or other CI/CD systems. If there is a high amount of such traffic, hitting a Gitaly server with many clones for a large repository is likely to put the server under significant strain.

Gitaly pack-objects cache

Turn on the Gitaly pack-objects cache, which reduces the work that the server has to do for clones and fetches.


The pack-objects cache caches the data that the git-pack-objects process produces. This response is sent back to the Git client initiating the clone or fetch. If several fetches are requesting the same set of refs, Git on the Gitaly server doesn't have to re-generate the response data with each clone or fetch call, but instead serves that data from an in-memory cache that Gitaly maintains.

This can help immensely in the presence of a high rate of clones for a single repository.

For more information, see Pack-objects cache.

Reduce concurrent clones in CI/CD

CI/CD loads tend to be concurrent because pipelines are scheduled during set times. As a result, the Git requests against the repositories can spike notably during these times and lead to reduced performance for both CI/CD and users alike.

Reduce CI/CD pipeline concurrency by staggering them to run at different times. For example, a set running at one time and another set running several minutes later.

Shallow cloning

In your CI/CD systems, set the --depth option in the git clone or git fetch call.

GitLab and GitLab Runner perform a shallow clone by default.

If possible, set the clone depth with a small number like 10. Shallow clones make Git request only the latest set of changes for a given branch, up to desired number of commits.

This significantly speeds up fetching of changes from Git repositories, especially if the repository has a very long backlog consisting of a number of big files because we effectively reduce amount of data transfer.

The following GitLab CI/CD pipeline configuration example sets the GIT_DEPTH.


    - ls -al

Git strategy

Use git fetch instead of git clone on CI/CD systems if it's possible to keep a working copy of the repository.

By default, GitLab is configured to use the fetch Git strategy, which is recommended for large repositories.


git clone gets the entire repository from scratch, whereas git fetch only asks the server for references that do not already exist in the repository. Naturally, git fetch causes the server to do less work. git-pack-objects doesn't have to go through all branches and tags and roll everything up into a response that gets sent over. Instead, it only has to worry about a subset of references to pack up. This strategy also reduces the amount of data to transfer.

Git clone path

GIT_CLONE_PATH allows you to control where you clone your repositories. This can have implications if you heavily use big repositories with a fork-based workflow.

A fork, from the perspective of GitLab Runner, is stored as a separate repository with a separate worktree. That means that GitLab Runner cannot optimize the usage of worktrees and you might have to instruct GitLab Runner to use that.

In such cases, ideally you want to make the GitLab Runner executor be used only for the given project and not shared across different projects to make this process more efficient.

The GIT_CLONE_PATH must be in the directory set in $CI_BUILDS_DIR. You can't pick any path from disk.

Git clean flags

GIT_CLEAN_FLAGS allows you to control whether or not you require the git clean command to be executed for each CI/CD job. By default, GitLab ensures that:

  • You have your worktree on the given SHA.
  • Your repository is clean.

GIT_CLEAN_FLAGS is disabled when set to none. On very big repositories, this might be desired because git clean is disk I/O intensive. Controlling that with GIT_CLEAN_FLAGS: -ffdx -e .build/ (for example) allows you to control and disable removal of some directories in the worktree between subsequent runs, which can speed-up the incremental builds. This has the biggest effect if you re-use existing machines and have an existing worktree that you can re-use for builds.

For exact parameters accepted by GIT_CLEAN_FLAGS, see the documentation for git clean. The available parameters are dependent on the Git version.

Git fetch extra flags

GIT_FETCH_EXTRA_FLAGS allows you to modify git fetch behavior by passing extra flags.

For example, if your project contains a large number of tags that your CI/CD jobs don't rely on, you could add --no-tags to the extra flags to make your fetches faster and more compact.

Also in the case where your repository does not contain a lot of tags, --no-tags can make a big difference in some cases. If your CI/CD builds do not depend on Git tags, setting --no-tags is worth trying.

For more information, see the GIT_FETCH_EXTRA_FLAGS documentation.

Configure Gitaly negotiation timeouts

You might experience a fatal: the remote end hung up unexpectedly error when attempting to fetch or archive:

  • Large repositories.
  • Many repositories in parallel.
  • The same large repository in parallel.

You can attempt to mitigate this issue by increasing the default negotiation timeout values. For more information, see Configure the negotiation timeouts.

Optimize your repository

Another avenue to keeping GitLab scalable with your monorepo is to optimize the repository itself.

Profiling repositories

Large repositories generally experience performance issues in Git. Knowing why your repository is large can help you develop mitigation strategies to avoid performance problems.

You can use git-sizer to get a snapshot of repository characteristics and discover problem aspects of your monorepo.

For example:

Processing blobs: 1652370
Processing trees: 3396199
Processing commits: 722647
Matching commits to trees: 722647
Processing annotated tags: 534
Processing references: 539
| Name                         | Value     | Level of concern               |
| ---------------------------- | --------- | ------------------------------ |
| Overall repository size      |           |                                |
| * Commits                    |           |                                |
|   * Count                    |   723 k   | *                              |
|   * Total size               |   525 MiB | **                             |
| * Trees                      |           |                                |
|   * Count                    |  3.40 M   | **                             |
|   * Total size               |  9.00 GiB | ****                           |
|   * Total tree entries       |   264 M   | *****                          |
| * Blobs                      |           |                                |
|   * Count                    |  1.65 M   | *                              |
|   * Total size               |  55.8 GiB | *****                          |
| * Annotated tags             |           |                                |
|   * Count                    |   534     |                                |
| * References                 |           |                                |
|   * Count                    |   539     |                                |
|                              |           |                                |
| Biggest objects              |           |                                |
| * Commits                    |           |                                |
|   * Maximum size         [1] |  72.7 KiB | *                              |
|   * Maximum parents      [2] |    66     | ******                         |
| * Trees                      |           |                                |
|   * Maximum entries      [3] |  1.68 k   | *                              |
| * Blobs                      |           |                                |
|   * Maximum size         [4] |  13.5 MiB | *                              |
|                              |           |                                |
| History structure            |           |                                |
| * Maximum history depth      |   136 k   |                                |
| * Maximum tag depth      [5] |     1     |                                |
|                              |           |                                |
| Biggest checkouts            |           |                                |
| * Number of directories  [6] |  4.38 k   | **                             |
| * Maximum path depth     [7] |    13     | *                              |
| * Maximum path length    [8] |   134 B   | *                              |
| * Number of files        [9] |  62.3 k   | *                              |
| * Total size of files    [9] |   747 MiB |                                |
| * Number of symlinks    [10] |    40     |                                |
| * Number of submodules       |     0     |                                |

In this example, a few items are raised with a high level of concern. See the following sections for information on solving:

  • A large number of references.
  • Large blobs.

Large number of references

References in Git are branch and tag names that point to a particular commit. You can use the git for-each-ref command to list all references present in a repository. A large number of references in a repository can have detrimental impact on the command's performance. To understand why, we need to understand how Git stores references and uses them.

In general, Git stores all references as loose files in the .git/refs folder of the repository. As the number of references grows, the seek time to find a particular reference in the folder also increases. Therefore, every time Git has to parse a reference, there is an increased latency due to the added seek time of the file system.

To resolve this issue, Git uses pack-refs. In short, instead of storing each reference in a single file, Git creates a single .git/packed-refs file that contains all the references for that repository. This file reduces storage space while also increasing performance because seeking within a single file is faster than seeking a file within a directory. However, creating and updating new references is still done through loose files and are not added to the packed-refs file. To recreate the packed-refs file, run git pack-refs.

Gitaly runs git pack-refs during housekeeping to move loose references into packed-refs files. While this is very beneficial for most repositories, write-heavy repositories still have the problem that:

  • Creating or updating references creates new loose files.
  • Deleting references involves modifying the existing packed-refs file altogether to remove the existing reference.

These problems still cause the same performance issues.

In addition, fetches and clones from repositories includes the transfer of missing objects from the server to the client. When there are numerous references, Git iterates over all references and walks the internal graph structure for each reference to find the missing objects to transfer to the client. Iteration and walking are CPU-intensive operations that increase the latency of these commands.

In repositories with a lot of activity, this often causes a domino effect because every operation is slower and each operation stalls subsequent operations.

Mitigation strategies

To mitigate the effects of a large number of references in a monorepo:

  • Create an automated process for cleaning up old branches.

  • If certain references don't need to be visible to the client, hide them using the transfer.hideRefs configuration setting. Because Gitaly ignores any on-server Git configuration, you must change the Gitaly configuration itself in /etc/gitlab/gitlab.rb:

    gitaly['configuration'] = {
      # ...
      git: {
        # ...
        config: [
          # ...
          { key: "transfer.hideRefs", value: "refs/namespace_to_hide" },

In Git 2.42.0 and later, different Git operations can skip over hidden references when doing an object graph walk.

Large blobs

The presence of large files (called blobs in Git), can be problematic for Git because it does not handle large binary files efficiently. If there are blobs over 10 MB or instance in the git-sizer output, this probably means there is binary data in your repository.

Use LFS for large blobs

Store binary or blob files (for example, packages, audio, video, or graphics) as Large File Storage (LFS) objects. With LFS, the objects are stored externally, such as in Object Storage, which reduces the number and size of objects in the repository. Storing objects in external Object Storage can improve performance.

For more information, refer to the Git LFS documentation.

Reference architectures

Large repositories tend to be found in larger organisations with many users. The GitLab Quality Engineering and Support teams provide several reference architectures that are the recommended way to deploy GitLab at scale.

In these types of setups, the GitLab environment used should match a reference architecture to improve performance.