• Enable the container registry
  • Linux package installations
  • Helm Charts installations
  • Self-compiled installations
• Container registry domain configuration
  • Configure container registry under an existing GitLab domain
  • Configure container registry under its own domain
• Disable container registry site-wide
• Disable container registry for new projects site-wide
  • Increase token duration
• Configure storage for the container registry
  • Use file system
  • Use object storage
    • Configure s3 and gcs storage drivers for Linux package installations
    • Self-compiled installations
    • Migrate to object storage without downtime
    • Moving to Azure Object Storage
  • Disable redirect for storage driver
    • Encrypted S3 buckets
  • Storage limitations
• Change the registry’s internal port
• Disable container registry per project
• Use an external container registry with GitLab as an auth endpoint
  • Linux package installations
  • Self-compiled installations
• Configure container registry notifications
• Run the Cleanup policy now
  • Registry Disk Space Usage by Project
• Container registry metadata database
• Container registry garbage collection
  • Understanding the content-addressable layers
  • Remove unreferenced layers
  • Removing untagged manifests and unreferenced layers
  • Performing garbage collection without downtime
  • Running the garbage collection on schedule
  • Stop garbage collection
  • Continuous zero-downtime garbage collection
• Scaling by component
  • Database
  • Registry server
  • Redis Cache
  • Storage
  • Online garbage collection
• Configure GitLab and Registry to run on separate nodes (Linux package installations)
  • Configure Registry
  • Configure GitLab
• Architecture of GitLab container registry
  • Communication between GitLab and Registry
• Migrate from a third-party registry
  • Docker Distribution Registry
• Max retries for deleting container images

GitLab container registry administration

With the GitLab container registry, every project can have its own space to store Docker images.

For more details about the Distribution Registry:

• Configuration
• Storage drivers
• Deploy a registry server

This document is the administrator’s guide. To learn how to use the GitLab Container Registry, see the user documentation.

Enable the container registry

The process for enabling the container registry depends on the type of installation you use.

Linux package installations

If you installed GitLab by using the Linux package, the container registry may or may not be available by default.

The container registry is automatically enabled and available on your GitLab domain, port 5050 if you’re using the built-in Let’s Encrypt integration.

Otherwise, the container registry is not enabled. To enable it:

• You can configure it for your GitLab domain, or
• You can configure it for a different domain.

The container registry works under HTTPS by default. You can use HTTP but it’s not recommended and is beyond the scope of this document.

Helm Charts installations

For Helm Charts installations, see Using the Container Registry. in the Helm Charts documentation.

Self-compiled installations

If you self-compiled your GitLab installation:

1. You must deploy a registry using the image corresponding to the version of GitLab you are installing (for example: registry.gitlab.com/gitlab-org/build/cng/gitlab-container-registry:v3.15.0-gitlab)
2. After the installation is complete, to enable it, you must configure the Registry’s settings in gitlab.yml.
3. Use the sample NGINX configuration file from under lib/support/nginx/registry-ssl and edit it to match the host, port, and TLS certificate paths.

The contents of gitlab.yml are:

registry:
  enabled: true
  host: registry.gitlab.example.com
  port: 5005
  api_url: http://localhost:5000/
  key: config/registry.key
  path: shared/registry
  issuer: gitlab-issuer

Where:

A Registry init file is not shipped with GitLab if you install it from source. Hence, restarting GitLab does not restart the Registry should you modify its settings. Read the upstream documentation on how to achieve that.

At the absolute minimum, make sure your Registry configuration has container_registry as the service and https://gitlab.example.com/jwt/auth as the realm:

auth:
  token:
    realm: https://gitlab.example.com/jwt/auth
    service: container_registry
    issuer: gitlab-issuer
    rootcertbundle: /root/certs/certbundle

Container registry domain configuration

You can configure the Registry’s external domain in either of these ways:

• Use the existing GitLab domain. The Registry listens on a port and reuses the TLS certificate from GitLab.
• Use a completely separate domain with a new TLS certificate for that domain.

Because the container registry requires a TLS certificate, cost may be a factor.

Take this into consideration before configuring the container registry for the first time.

Configure container registry under an existing GitLab domain

If the container registry is configured to use the existing GitLab domain, you can expose the container registry on a port. This way you can reuse the existing GitLab TLS certificate.

If the GitLab domain is https://gitlab.example.com and the port to the outside world is 5050, to configure the container registry:

• Edit gitlab.rb if you are using a Linux package installation.
• Edit gitlab.yml if you are using a self-compiled installation.

Ensure you choose a port different than the one that Registry listens to (5000 by default), otherwise conflicts occur.

registry_external_url 'https://registry-gitlab.example.com'
registry_nginx['redirect_http_to_https'] = true
registry_nginx['listen_port'] = 5678

Users should now be able to sign in to the container registry with their GitLab credentials using:

docker login gitlab.example.com:5050

Configure container registry under its own domain

When the Registry is configured to use its own domain, you need a TLS certificate for that specific domain (for example, registry.example.com). You might need a wildcard certificate if hosted under a subdomain of your existing GitLab domain. For example, *.gitlab.example.com, is a wildcard that matches registry.gitlab.example.com, and is distinct from *.example.com.

As well as manually generated SSL certificates (explained here), certificates automatically generated by Let’s Encrypt are also supported in Linux package installations.

Let’s assume that you want the container registry to be accessible at https://registry.gitlab.example.com.

registry_nginx['ssl_certificate'] = "/etc/gitlab/ssl/certificate.pem"
registry_nginx['ssl_certificate_key'] = "/etc/gitlab/ssl/certificate.key"

Users should now be able to sign in to the container registry using their GitLab credentials:

docker login registry.gitlab.example.com

Disable container registry site-wide

When you disable the Registry by following these steps, you do not remove any existing Docker images. Docker image removal is handled by the Registry application itself.

Disable container registry for new projects site-wide

If the container registry is enabled, then it should be available on all new projects. To disable this function and let the owners of a project to enable the container registry by themselves, follow the steps below.

Increase token duration

In GitLab, tokens for the container registry expire every five minutes. To increase the token duration:

1. On the left sidebar, at the bottom, select Admin.
2. Select Settings > CI/CD.
3. Expand Container Registry.
4. For the Authorization token duration (minutes), update the value.
5. Select Save changes.

Configure storage for the container registry

You can configure the container registry to use various storage backends by configuring a storage driver. By default the GitLab container registry is configured to use the file system driver configuration.

The different supported drivers are:

Although most S3 compatible services (like MinIO) should work with the container registry, we only guarantee support for AWS S3. Because we cannot assert the correctness of third-party S3 implementations, we can debug issues, but we cannot patch the registry unless an issue is reproducible against an AWS S3 bucket.

Use file system

If you want to store your images on the file system, you can change the storage path for the container registry, follow the steps below.

This path is accessible to:

• The user running the container registry daemon.
• The user running GitLab.

All GitLab, Registry, and web server users must have access to this directory.

Use object storage

If you want to store your images on object storage, you can change the storage driver for the container registry.

Read more about using object storage with GitLab.

Configure s3 and gcs storage drivers for Linux package installations

The following configuration steps are for the s3 and gcs storage drivers. Other storage drivers are supported.

To configure the s3 storage driver for a Linux package installation:

1. Edit /etc/gitlab/gitlab.rb:

   registry['storage'] = {
     's3' => {
       'accesskey' => 's3-access-key',
       'secretkey' => 's3-secret-key-for-access-key',
       'bucket' => 'your-s3-bucket',
       'region' => 'your-s3-region',
       'regionendpoint' => 'your-s3-regionendpoint'
     }
   }
   

   To avoid using static credentials, use an IAM role and omit accesskey and secretkey. Make sure that your IAM profile follows the permissions documented by Docker.

   registry['storage'] = {
     's3' => {
       'bucket' => 'your-s3-bucket',
       'region' => 'your-s3-region'
     }
   }
   

   If using with an AWS S3 VPC endpoint, then set regionendpoint to your VPC endpoint address and set pathstyle to false:

   registry['storage'] = {
     's3' => {
       'accesskey' => 's3-access-key',
       'secretkey' => 's3-secret-key-for-access-key',
       'bucket' => 'your-s3-bucket',
       'region' => 'your-s3-region',
       'regionendpoint' => 'your-s3-vpc-endpoint',
       'pathstyle' => false
     }
   }
   

   • regionendpoint is only required when configuring an S3 compatible service such as MinIO, or when using an AWS S3 VPC Endpoint.
   • your-s3-bucket should be the name of a bucket that exists, and can’t include subdirectories.
   • pathstyle should be set to true to use host/bucket_name/object style paths instead of bucket_name.host/object. Set to false for AWS S3.

   You can set a rate limit on connections to S3 to avoid 503 errors from the S3 API. To do this, set maxrequestspersecond to a number within the S3 request rate threshold:

   registry['storage'] = {
     's3' => {
       'accesskey' => 's3-access-key',
       'secretkey' => 's3-secret-key-for-access-key',
       'bucket' => 'your-s3-bucket',
       'region' => 'your-s3-region',
       'regionendpoint' => 'your-s3-regionendpoint',
       'maxrequestspersecond' => 100
     }
   }
   

2. Save the file and reconfigure GitLab for the changes to take effect.

To configure the gcs storage driver for a Linux package installation:

1. Edit /etc/gitlab/gitlab.rb:

   registry['storage'] = {
     'gcs' => {
       'bucket' => 'BUCKET_NAME',
       'keyfile' => 'PATH/TO/KEYFILE',
       # If you have the bucket shared with other apps beyond the registry, uncomment the following:
       # 'rootdirectory' => '/gcs/object/name/prefix'
     }
   }
   

   GitLab supports all available parameters.

2. Save the file and reconfigure GitLab for the changes to take effect.

Self-compiled installations

Configuring the storage driver is done in the registry configuration YAML file created when you deployed your Docker registry.

s3 storage driver example:

storage:
  s3:
    accesskey: 's3-access-key'                # Not needed if IAM role used
    secretkey: 's3-secret-key-for-access-key' # Not needed if IAM role used
    bucket: 'your-s3-bucket'
    region: 'your-s3-region'
    regionendpoint: 'your-s3-regionendpoint'
  cache:
    blobdescriptor: inmemory
  delete:
    enabled: true

your-s3-bucket should be the name of a bucket that exists, and can’t include subdirectories.

Migrate to object storage without downtime

To migrate storage without stopping the container registry, set the container registry to read-only mode. On large instances, this may require the container registry to be in read-only mode for a while. During this time, you can pull from the container registry, but you cannot push.

1. Optional: To reduce the amount of data to be migrated, run the garbage collection tool without downtime.

2. This example uses the aws CLI. If you haven’t configured the CLI before, you have to configure your credentials by running sudo aws configure. Because a non-administrator user likely can’t access the container registry folder, ensure you use sudo. To check your credential configuration, run ls to list all buckets.

   sudo aws --endpoint-url https://your-object-storage-backend.com s3 ls
   

   If you are using AWS as your back end, you do not need the --endpoint-url.

3. Copy initial data to your S3 bucket, for example with the aws CLI cp or sync command. Make sure to keep the docker folder as the top-level folder inside the bucket.

   sudo aws --endpoint-url https://your-object-storage-backend.com s3 sync registry s3://mybucket
   

   If you have a lot of data, you may be able to improve performance by running parallel sync operations.
4. To perform the final data sync, put the container registry in read-only mode and reconfigure GitLab.

5. Sync any changes dating from after the initial data load to your S3 bucket, and delete files that exist in the destination bucket but not in the source:

   sudo aws --endpoint-url https://your-object-storage-backend.com s3 sync registry s3://mybucket --delete --dryrun
   

   After verifying the command performs as expected, remove the --dryrun flag and run the command.

   The --delete flag deletes files that exist in the destination but not in the source. If you swap the source and destination, all data in the Registry is deleted.

6. Verify all container registry files have been uploaded to object storage by looking at the file count returned by these two commands:

   sudo find registry -type f | wc -l
   

   sudo aws --endpoint-url https://your-object-storage-backend.com s3 ls s3://mybucket --recursive | wc -l
   

   The output of these commands should match, except for the content in the _uploads directories and sub-directories.

7. Configure your registry to use the S3 bucket for storage.
8. For the changes to take effect, set the Registry back to read-write mode and reconfigure GitLab.

Moving to Azure Object Storage

registry['storage'] = {
  'azure' => {
    'accountname' => '<your_storage_account_name>',
    'accountkey' => '<base64_encoded_account_key>',
    'container' => '<container_name>',
    'trimlegacyrootprefix' => true
  }
}

storage:
  azure:
    accountname: <your_storage_account_name>
    accountkey: <base64_encoded_account_key>
    container: <container_name>
    trimlegacyrootprefix: true

By default, Azure Storage Driver uses the core.windows.net realm. You can set another value for realm in the azure section (for example, core.usgovcloudapi.net for Azure Government Cloud).

Disable redirect for storage driver

By default, users accessing a registry configured with a remote backend are redirected to the default backend for the storage driver. For example, registries can be configured using the s3 storage driver, which redirects requests to a remote S3 bucket to alleviate load on the GitLab server.

However, this behavior is undesirable for registries used by internal hosts that usually can’t access public servers. To disable redirects and proxy download, set the disable flag to true as follows. This makes all traffic always go through the Registry service. This results in improved security (less surface attack as the storage backend is not publicly accessible), but worse performance (all traffic is redirected via the service).

Encrypted S3 buckets

You can use server-side encryption with AWS KMS for S3 buckets that have SSE-S3 or SSE-KMS encryption enabled by default. Customer master keys (CMKs) and SSE-C encryption aren’t supported because this requires sending the encryption keys in every request.

For SSE-S3, you must enable the encrypt option in the registry settings. How you do this depends on how you installed GitLab. Follow the instructions here that match your installation method.

Storage limitations

There is no storage limitation, which means a user can upload an infinite amount of Docker images with arbitrary sizes. This setting should be configurable in future releases.

Change the registry’s internal port

The Registry server listens on localhost at port 5000 by default, which is the address for which the Registry server should accept connections. In the examples below we set the Registry’s port to 5010.

Disable container registry per project

If Registry is enabled in your GitLab instance, but you don’t need it for your project, you can disable it from your project’s settings.

Use an external container registry with GitLab as an auth endpoint

If you use an external container registry, some features associated with the container registry may be unavailable or have inherent risks.

For the integration to work, the external registry must be configured to use a JSON Web Token to authenticate with GitLab. The external registry’s runtime configuration must have the following entries:

auth:
  token:
    realm: https://gitlab.example.com/jwt/auth
    service: container_registry
    issuer: gitlab-issuer
    rootcertbundle: /root/certs/certbundle

Without these entries, the registry logins cannot authenticate with GitLab. GitLab also remains unaware of nested image names under the project hierarchy, like registry.example.com/group/project/image-name:tag or registry.example.com/group/project/my/image-name:tag, and only recognizes registry.example.com/group/project:tag.

Linux package installations

You can use GitLab as an auth endpoint with an external container registry.

1. Open /etc/gitlab/gitlab.rb and set necessary configurations:

   gitlab_rails['registry_enabled'] = true
   gitlab_rails['registry_api_url'] = "https://<external_registry_host>:5000"
   gitlab_rails['registry_issuer'] = "gitlab-issuer"
   

   • gitlab_rails['registry_enabled'] = true is needed to enable GitLab container registry features and authentication endpoint. The GitLab bundled container registry service does not start, even with this enabled.
   • gitlab_rails['registry_api_url'] = "http://<external_registry_host>:5000" must be changed to match the host where Registry is installed. It must also specify https if the external registry is configured to use TLS.

2. A certificate-key pair is required for GitLab and the external container registry to communicate securely. You need to create a certificate-key pair, configuring the external container registry with the public certificate (rootcertbundle) and configuring GitLab with the private key. To do that, add the following to /etc/gitlab/gitlab.rb:

   # registry['internal_key'] should contain the contents of the custom key
   # file. Line breaks in the key file should be marked using `\n` character
   # Example:
   registry['internal_key'] = "---BEGIN RSA PRIVATE KEY---\nMIIEpQIBAA\n"
   
   # Optionally define a custom file for a Linux package installation to write the contents
   # of registry['internal_key'] to.
   gitlab_rails['registry_key_path'] = "/custom/path/to/registry-key.key"
   

   Each time reconfigure is executed, the file specified at registry_key_path gets populated with the content specified by internal_key. If no file is specified, Linux package installations default it to /var/opt/gitlab/gitlab-rails/etc/gitlab-registry.key and populates it.

3. To change the container registry URL displayed in the GitLab Container Registry pages, set the following configurations:

   gitlab_rails['registry_host'] = "registry.gitlab.example.com"
   gitlab_rails['registry_port'] = "5005"
   

4. Save the file and reconfigure GitLab for the changes to take effect.

Self-compiled installations

1. Open /home/git/gitlab/config/gitlab.yml, and edit the configuration settings under registry:

   ## Container registry
   
   registry:
     enabled: true
     host: "registry.gitlab.example.com"
     port: "5005"
     api_url: "https://<external_registry_host>:5000"
     path: /var/lib/registry
     key: /path/to/keyfile
     issuer: gitlab-issuer
   

   Read more about what these parameters mean.

2. Save the file and restart GitLab for the changes to take effect.

Configure container registry notifications

You can configure the container registry to send webhook notifications in response to events happening in the registry.

Read more about the container registry notifications configuration options in the Docker Registry notifications documentation.

You can configure multiple endpoints for the container registry.

notifications:
  endpoints:
    - name: alistener
      disabled: false
      url: https://my.listener.com/event
      headers: <http.Header>
      timeout: 500
      threshold: 5 # DEPRECATED: use `maxretries` instead.
      maxretries: 5
      backoff: 1000

Run the Cleanup policy now

1. Configure the gitlab.rb file on the Sidekiq nodes to point to the correct registry URL.
2. Copy the registry.key file to each Sidekiq node.

For more information, see the Sidekiq configuration page.

To reduce the amount of Container Registry disk space used by a given project, administrators can set up cleanup policies and run garbage collection.

Registry Disk Space Usage by Project

To find the disk space used by each project, run the following in the GitLab Rails console:

projects_and_size = [["project_id", "creator_id", "registry_size_bytes", "project path"]]
# You need to specify the projects that you want to look through. You can get these in any manner.
projects = Project.last(100)

projects.each do |p|
   project_total_size = 0
   container_repositories = p.container_repositories

   container_repositories.each do |c|
       c.tags.each do |t|
          project_total_size = project_total_size + t.total_size unless t.total_size.nil?
       end
   end

   if project_total_size > 0
      projects_and_size << [p.project_id, p.creator&.id, project_total_size, p.full_path]
   end
end

# print it as comma separated output
projects_and_size.each do |ps|
   puts "%s,%s,%s,%s" % ps
end

To remove image tags by running the cleanup policy, run the following commands in the GitLab Rails console:

# Numeric ID of the project whose container registry should be cleaned up
P = <project_id>

# Numeric ID of a user with Developer, Maintainer, or Owner role for the project
U = <user_id>

# Get required details / objects
user    = User.find_by_id(U)
project = Project.find_by_id(P)
policy  = ContainerExpirationPolicy.find_by(project_id: P)

# Loop through each container repository
project.container_repositories.find_each do |repo|
  puts repo.attributes

  # Start the tag cleanup
  puts Projects::ContainerRepository::CleanupTagsService.new(container_repository: repo, current_user: user, params: policy.attributes.except("created_at", "updated_at")).execute
end

You can also run cleanup on a schedule.

To enable cleanup policies for all projects instance-wide, you need to find all projects with a container registry, but with the cleanup policy disabled:

# Find all projects where Container registry is enabled, and cleanup policies disabled

projects = Project.find_by_sql ("SELECT * FROM projects WHERE id IN (SELECT project_id FROM container_expiration_policies WHERE enabled=false AND id IN (SELECT project_id FROM container_repositories))")

# Loop through each project
projects.each do |p|

# Print project IDs and project full names
    puts "#{p.id},#{p.full_name}"
end

Container registry metadata database

• Generally available in GitLab 17.3.

The metadata database enables many new registry features, including online garbage collection, and increases the efficiency of many registry operations. See the Container registry metadata database page for details.

Container registry garbage collection

The container registry can use considerable amounts of storage space, and you might want to reduce storage usage. Among the listed options, deleting tags is the most effective option. However, tag deletion alone does not delete image layers, it only leaves the underlying image manifests untagged.

To more effectively free up space, the container registry has a garbage collector that can delete unreferenced layers and (optionally) untagged manifests.

To start the garbage collector, use the registry-garbage-collect command provided by gitlab-ctl.

The time required to perform garbage collection is proportional to the container registry data size.

Prerequisites:

• You must have installed GitLab by using a Linux package or the GitLab Helm chart.

Understanding the content-addressable layers

Consider the following example, where you first build the image:

# This builds a image with content of sha256:111111
docker build -t my.registry.com/my.group/my.project:latest .
docker push my.registry.com/my.group/my.project:latest

Now, you do overwrite :latest with a new version:

# This builds a image with content of sha256:222222
docker build -t my.registry.com/my.group/my.project:latest .
docker push my.registry.com/my.group/my.project:latest

Now, the :latest tag points to manifest of sha256:222222. Due to the architecture of registry, this data is still accessible when pulling the image my.registry.com/my.group/my.project@sha256:111111, though it is no longer directly accessible via the :latest tag.

Remove unreferenced layers

Image layers are the bulk of the container registry storage. A layer is considered unreferenced when no image manifest references it. Unreferenced layers are the default target of the container registry garbage collector.

If you did not change the default location of the configuration file, run:

sudo gitlab-ctl registry-garbage-collect

If you changed the location of the container registry config.yml:

sudo gitlab-ctl registry-garbage-collect /path/to/config.yml

You can also remove all untagged manifests and unreferenced layers to recover additional space.

Removing untagged manifests and unreferenced layers

By default the container registry garbage collector ignores images that are untagged, and users can keep pulling untagged images by digest. Users can also re-tag images in the future, making them visible again in the GitLab UI and API.

If you do not care about untagged images and the layers exclusively referenced by these images, you can delete them all. Use the -m flag on the registry-garbage-collect command:

sudo gitlab-ctl registry-garbage-collect -m

If you are unsure about deleting untagged images, back up your registry data before proceeding.

Performing garbage collection without downtime

To do garbage collection while keeping the container registry online, put the registry in read-only mode and bypass the built-in gitlab-ctl registry-garbage-collect command.

You can pull but not push images while the container registry is in read-only mode. The container registry must remain in read-only for the full duration of the garbage collection.

By default, the registry storage path is /var/opt/gitlab/gitlab-rails/shared/registry.

To enable the read-only mode:

1. In /etc/gitlab/gitlab.rb, specify the read-only mode:

   registry['storage'] = {
     'filesystem' => {
       'rootdirectory' => "<your_registry_storage_path>"
     },
     'maintenance' => {
       'readonly' => {
         'enabled' => true
       }
     }
   }
   

2. Save and reconfigure GitLab:

   sudo gitlab-ctl reconfigure
   

   This command sets the container registry into the read-only mode.

3. Next, trigger one of the garbage collect commands:

   # Remove unreferenced layers
   sudo /opt/gitlab/embedded/bin/registry garbage-collect /var/opt/gitlab/registry/config.yml
   
   # Remove untagged manifests and unreferenced layers
   sudo /opt/gitlab/embedded/bin/registry garbage-collect -m /var/opt/gitlab/registry/config.yml
   

   This command starts the garbage collection. The time to complete is proportional to the registry data size.

4. Once done, in /etc/gitlab/gitlab.rb change it back to read-write mode:

   registry['storage'] = {
     'filesystem' => {
       'rootdirectory' => "<your_registry_storage_path>"
     },
     'maintenance' => {
       'readonly' => {
         'enabled' => false
       }
     }
   }
   

5. Save and reconfigure GitLab:

   sudo gitlab-ctl reconfigure
   

Running the garbage collection on schedule

Ideally, you want to run the garbage collection of the registry regularly on a weekly basis at a time when the registry is not being in-use. The simplest way is to add a new crontab job that it runs periodically once a week.

Create a file under /etc/cron.d/registry-garbage-collect:

SHELL=/bin/sh
PATH=/usr/local/sbin:/usr/local/bin:/sbin:/bin:/usr/sbin:/usr/bin

# Run every Sunday at 04:05am
5 4 * * 0  root gitlab-ctl registry-garbage-collect

You may want to add the -m flag to remove untagged manifests and unreferenced layers.

Stop garbage collection

If you anticipate stopping garbage collection, you should manually run garbage collection as described in Performing garbage collection without downtime. You can then stop garbage collection by pressing Control+C.

Otherwise, interrupting gitlab-ctl could leave your registry service in a down state. In this case, you must find the garbage collection process itself on the system so that the gitlab-ctl command can bring the registry service back up again.

Also, there’s no way to save progress or results during the mark phase of the process. Only once blobs start being deleted is anything permanent done.

Continuous zero-downtime garbage collection

You can run garbage collection in the background without the need to schedule it or require read-only mode, if you migrate to the metadata database.

Scaling by component

This section outlines the potential performance bottlenecks as registry traffic increases by component. Each subsection is roughly ordered by recommendations that benefit from smaller to larger registry workloads. The registry is not included in the reference architectures, and there are no scaling guides which target number of seats or requests per second.

Database

1. Move to a separate database: As database load increases, scale vertically by moving the registry metadata database to a separate physical database. A separate database can increase the amount of resources available to the registry database while isolating the traffic produced by the registry.
2. Move to a HA PostgreSQL third-party solution: Similar to Praefect, moving to a reputable provider or solution enables HA and is suitable for multi-node registry deployments. You must pick a provider that supports native Postgres partitioning, triggers, and functions, as the registry makes heavy use of these.

Registry server

1. Move to a separate node: A separate node is one way to scale vertically to increase the resources available to the container registry server process.
2. Run multiple registry nodes behind a load balancer: While the registry can handle a high amount of traffic with a single large node, the registry is generally intended to scale horizontally with multiple deployments. Configuring multiple smaller nodes also enables techniques such as autoscaling.

Redis Cache

Enabling the Redis cache improves performance, but also enables features such as renaming repositories.

1. Redis Server: A single Redis instance is supported and is the simplest way to access the benefits of the Redis caching.
2. Redis Sentinel: Redis Sentinel is also supported and enables the cache to be HA.
3. Redis Cluster: Redis Cluster can also be used for further scaling as deployments grow.

Storage

1. Local file system: A local file system is the default and is relatively performant, but not suitable for multi-node deployments or a large amount of registry data.
2. Object storage: Use object storage to enable the practical storage of a larger amount of registry data. Object storage is also suitable for multi-node registry deployments.

Online garbage collection

1. Adjust defaults: If online garbage collection is not reliably clearing the review queues, you can adjust the interval settings in the manifests and blobs sections under the gc configuration section. The default is 5s, and these can be configured with milliseconds as well, for example 500ms.
2. Scale horizontally with the registry server: If you are scaling the registry application horizontally with multi-node deployments, online garbage collection automatically scales without the need for configuration changes.

Configure GitLab and Registry to run on separate nodes (Linux package installations)

By default, package assumes that both services are running on the same node. To get GitLab and Registry to run on a separate nodes, separate configuration is necessary for Registry and GitLab.

Configure Registry

Below you can find configuration options you should set in /etc/gitlab/gitlab.rb, for Registry to run separately from GitLab:

• registry['registry_http_addr'], default set programmatically. Needs to be reachable by web server (or LB).
• registry['token_realm'], default set programmatically. Specifies the endpoint to use to perform authentication, usually the GitLab URL. This endpoint needs to be reachable by user.
• registry['http_secret'], random string. A random piece of data used to sign state that may be stored with the client to protect against tampering.
• registry['internal_key'], default automatically generated. Contents of the key that GitLab uses to sign the tokens. They key gets created on the Registry server, but it is not used there.
• gitlab_rails['registry_key_path'], default set programmatically. This is the path where internal_key contents are written to disk.
• registry['internal_certificate'], default automatically generated. Contents of the certificate that GitLab uses to sign the tokens.
• registry['rootcertbundle'], default set programmatically. Path to certificate. This is the path where internal_certificate contents are written to disk.
• registry['health_storagedriver_enabled'], default set programmatically. Configure whether health checks on the configured storage driver are enabled.
• gitlab_rails['registry_issuer'], default value. This setting needs to be set the same between Registry and GitLab.

Configure GitLab

Below you can find configuration options you should set in /etc/gitlab/gitlab.rb, for GitLab to run separately from Registry:

• gitlab_rails['registry_enabled'], must be set to true. This setting signals to GitLab that it should allow Registry API requests.
• gitlab_rails['registry_api_url'], default set programmatically. This is the Registry URL used internally that users do not need to interact with, registry['registry_http_addr'] with scheme.
• gitlab_rails['registry_host'], for example, registry.gitlab.example. Registry endpoint without the scheme, the address that gets shown to the end user.
• gitlab_rails['registry_port']. Registry endpoint port, visible to the end user.
• gitlab_rails['registry_issuer'] must match the issuer in the Registry configuration.
• gitlab_rails['registry_key_path'], path to the key that matches the certificate on the Registry side.
• gitlab_rails['internal_key'], contents of the key that GitLab uses to sign the tokens.

Architecture of GitLab container registry

The GitLab registry is what users use to store their own Docker images. Because of that the Registry is client facing, meaning that we expose it directly on the web server (or load balancers, LB for short).

The flow described by the diagram above:

1. A user runs docker login registry.gitlab.example on their client. This reaches the web server (or LB) on port 443.
2. Web server connects to the Registry backend pool (by default, using port 5000). Since the user didn’t provide a valid token, the Registry returns a 401 HTTP code and the URL (token_realm from Registry configuration) where to get one. This points to the GitLab API.
3. The Docker client then connects to the GitLab API and obtains a token.
4. The API signs the token with the registry key and hands it to the Docker client
5. The Docker client now logs in again with the token received from the API. It can now push and pull Docker images.

Reference: https://distribution.github.io/distribution/spec/auth/token/

Communication between GitLab and Registry

Registry doesn’t have a way to authenticate users internally so it relies on GitLab to validate credentials. The connection between Registry and GitLab is TLS encrypted. The key is used by GitLab to sign the tokens while the certificate is used by Registry to validate the signature. By default, a self-signed certificate key pair is generated for all installations. This can be overridden as needed.

GitLab interacts with the Registry using the Registry private key. When a Registry request goes out, a new short-living (10 minutes) namespace limited token is generated and signed with the private key. The Registry then verifies that the signature matches the registry certificate specified in its configuration and allows the operation. GitLab background jobs processing (through Sidekiq) also interacts with Registry. These jobs talk directly to Registry to handle image deletion.

Migrate from a third-party registry

Using external container registries in GitLab was deprecated in GitLab 15.8 and the end of support occurred in GitLab 16.0. See the deprecation notice for more details.

The integration is not disabled in GitLab 16.0, but support for debugging and fixing issues is no longer provided. Additionally, the integration is no longer being developed or enhanced with new features. Third-party registry functionality might be completely removed after the new GitLab container registry version is available for self-managed (see epic 5521). Only the GitLab container registry is planned to be supported.

This section has guidance for administrators migrating from third-party registries to the GitLab container registry. If the third-party container registry you are using is not listed here, you can describe your use cases in the feedback issue.

For all of the instructions provided below, you should try them first on a test environment. Make sure everything continues to work as expected before replicating it in production.

Docker Distribution Registry

The Docker Distribution Registry was donated to the CNCF and is now known as the Distribution Registry. This registry is the open source implementation that the GitLab container registry is based on. The GitLab container registry is compatible with the basic functionality provided by the Distribution Registry, including all the supported storage backends. To migrate to the GitLab container registry you can follow the instructions on this page, and use the same storage backend as the Distribution Registry. The GitLab container registry should accept the same configuration that you are using for the Distribution Registry.

Max retries for deleting container images

Errors could happen when deleting container images, so deletions are retried to ensure the error is not a transient issue. Deletion is retried up to 10 times, with a back off delay between retries. This delay gives more time between retries for any transient errors to resolve.

Setting a maximum number of retries also helps detect if there are any persistent errors that haven’t been solved in between retries. After a deletion fails the maximum number of retries, the container repository status is set to delete_failed. With this status, the repository no longer retries deletions.

You should investigate any container repositories with a delete_failed status and try to resolve the issue. After the issue is resolved, you can set the repository status back to delete_scheduled so images can start to be deleted again. To update the repository status, from the rails console:

container_repository = ContainerRepository.find(<id>)
container_repository.update(status: 'delete_scheduled')