info: To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/product/ux/technical-writing/#assignments
---
# Dependency Proxy
The Dependency Proxy is a pull-through-cache for registry images from DockerHub. This document describes how this
feature is constructed in GitLab.
## Container registry
The Dependency Proxy for the container registry acts a stand-in for a remote container registry. In our case,
the remote registry is the public DockerHub registry.
From the user's perspective, the GitLab instance is just a container registry that they are interacting with to
pull images by using `docker login gitlab.com`
When you use `docker login gitlab.com`, the Docker client uses the [v2 API](https://docs.docker.com/registry/spec/api/)
to make requests.
To support authentication, we must include one route:
- [API Version Check](https://docs.docker.com/registry/spec/api/#api-version-check)
To support `docker pull` requests, we must include two additional routes:
- [Pulling an image manifest](https://docs.docker.com/registry/spec/api/#pulling-an-image-manifest)
- [Pulling an image layer (blob)](https://docs.docker.com/registry/spec/api/#pulling-a-layer)
These routes are defined in [`gitlab-org/gitlab/config/routes/group.rb`](https://gitlab.com/gitlab-org/gitlab/-/blob/3f76455ac9cf90a927767e55c837d6b07af818df/config/routes/group.rb#L164-175).
In its simplest form, the Dependency Proxy manages three requests:
- Logging in / returning a JWT
- Fetching a manifest
- Fetching a blob
Here is what the general request sequence looks like for the Dependency Proxy:
```mermaid
sequenceDiagram
Client->>+GitLab: Login? / request token
GitLab->>+Client: JWT
Client->>+GitLab: request a manifest for an image
GitLab->>+ExternalRegistry: request JWT
ExternalRegistry->>+GitLab : JWT
GitLab->>+ExternalRegistry : request manifest
ExternalRegistry->>+GitLab : return manifest
GitLab->>+GitLab : store manifest
GitLab->>+Client : return manifest
loop request image layers
Client->>+GitLab: request a blob from the manifest
GitLab->>+ExternalRegistry: request JWT
ExternalRegistry->>+GitLab : JWT
GitLab->>+ExternalRegistry : request blob
ExternalRegistry->>+GitLab : return blob
GitLab->>+GitLab : store blob
GitLab->>+Client : return blob
end
```
### Authentication and authorization
When a Docker client authenticates with a registry, the registry tells the client where to get a JSON Web Token
(JWT) and to use it for all subsequent requests. This allows the authentication service to live in a separate
application from the registry. For example, the GitLab container registry directs Docker clients to get a token
from `https://gitlab.com/jwt/auth`. This endpoint is part of the `gitlab-org/gitlab` project, also known as the
rails project or web service.
When a user tries to sign in to the dependency proxy with a Docker client, we must tell it where to get a JWT. We
can use the same endpoint we use with the container registry: `https://gitlab.com/jwt/auth`. But in our case,
we tell the Docker client to specify `service=dependency_proxy` in the parameters so can use a separate underlying
service to generate the token.
This sequence diagram shows the request flow for logging into the Dependency Proxy.
```mermaid
sequenceDiagram
autonumber
participant C as Docker CLI
participant R as GitLab (Dependency Proxy)
Note right of C: User tries `docker login gitlab.com` and enters username/password
C->>R: GET /v2/
Note left of R: Check for Authorization header, return 401 if none, return 200 if token exists and is valid
R->>C: 401 Unauthorized with header "WWW-Authenticate": "Bearer realm=\"http://gitlab.com/jwt/auth\",service=\"registry.docker.io\""
Note right of C: Request Oauth token using HTTP Basic Auth
C->>R: GET /jwt/auth
Note left of R: Token is returned
R->>C: 200 OK (with Bearer token included)
Note right of C: original request is tested again
C->>R: GET /v2/ (this time with `Authorization: Bearer [token]` header)
Note right of C: Login Succeeded
R->>C: 200 OK
```
The dependency proxy uses its own authentication service, separate from the authentication managed by the UI
(`ApplicationController`) and API (`ApiGuard`). Once the service has created a JWT, the `DependencyProxy::ApplicationController`
manages authentication and authorization for the rest of the requests. It manages the user by using `GitLab::Auth::Result` and
is similar to the authentication implemented by the Git client requests in `GitHttpClientController`.
### Caching
Blobs are cached artifacts with no logic around them. We cache them by digest. When we receive a request for a new blob,
we check to see if we have a blob with the requested digest, and return it. Otherwise we fetch it from the external
registry and cache it.
Manifests are more complicated, partially due to [rate limiting on DockerHub](https://www.docker.com/increase-rate-limits/).
A manifest is essentially a recipe for creating an image. It has a list of blobs to create a certain image. So
`alpine:latest` has a manifest associated with it that specifies the blobs needed to create the `alpine:latest`
image. The interesting part is that `alpine:latest` can change over time, so we can't just cache the manifest and
Once we put it all together, the sequence for requesting an image file looks like this:
```mermaid
sequenceDiagram
Client->>Workhorse: GET /v2/*group_id/dependency_proxy/containers/*image/manifests/*tag
Workhorse->>Rails: GET /v2/*group_id/dependency_proxy/containers/*image/manifests/*tag
Rails->>Rails: Check DB. Is manifest persisted in cache?
alt In Cache
Rails->>Workhorse: Respond with send-url injector
Workhorse->>Client: Send the file to the client
else Not In Cache
Rails->>Rails: Generate auth token and download URL for the manifest in upstream registry
Rails->>Workhorse: Respond with send-dependency injector
Workhorse->>External Registry: Request the manifest
External Registry->>Workhorse: Download the manifest
Workhorse->>Rails: GET /v2/*group_id/dependency_proxy/containers/*image/manifest/*tag/authorize
Rails->>Workhorse: Respond with upload instructions
Workhorse->>Client: Send the manifest file to the client with original headers
Workhorse->>Object Storage: Save the manifest file with some of it's header values
Workhorse->>Rails: Finalize the upload
end
```
### Cleanup policies
The cleanup policies for the Dependency Proxy work as time-to-live policies. They allow users to set the number
of days a file is allowed to remain cached if it has been unread. Since there is no way to associate the blobs
with the images they belong to (to do this, we would need to build the metadata database that the container registry
folks built), we can set up rules like "if this blob has not been pulled in 90 days, delete it". This means that
any files that are continuously getting pulled will not be removed from the cache, but if, for example,
`alpine:latest` changes and one of the underlying blobs is no longer used, it will eventually get cleaned up
because it has stopped getting pulled. We use the `read_at` attribute to track the last time a given
`dependency_proxy_blob` or `dependency_proxy_manifest` was pulled.
These work using a cron worker, [DependencyProxy::CleanupDependencyProxyWorker](https://gitlab.com/gitlab-org/gitlab/-/blob/7359d23f4e078479969c872924150219c6f1665f/app/workers/dependency_proxy/cleanup_dependency_proxy_worker.rb#L4),
that will kick off two [limited capacity](../sidekiq/limited_capacity_worker.md) workers: one to delete blobs,
and one to delete manifests. The capacity is set in an [application setting](settings.md#container-registry).
### Historic reference links
- [Dependency proxy for private groups](https://gitlab.com/gitlab-org/gitlab/-/merge_requests/46042) - initial authentication implementation
