PrimeHub is a kubernetes-based multi-user machine learning platform. For multi-user requirements, we fully integrate with keycloak as the identity provider (IdP) solution.
Here is a high-level diagram of PrimeHub
- Keycloak: Identity provider. Provide user databases and authentication/authorization services.
- Console: User interface to use and manage PrimeHub platform. It is a rich web application (by react) and sends the user command to graphql server.
- GraphQL: API Server to manage PrimeHub. The API may create/update the resources in kubernetes by kubernetes api or update users/groups by keycloak admin api.
- Controllers: Controllers are a group of components to watch and reconcile the state of the kubernetes and keycloak resources. The basic concept is described in the kubernetes official document
- Custom Resources: Kubernetes provide powerful extensibility for API. We can define the custom resources and allow us to store these resources in kubernetes.
- UI Components: For some features, we integrate existing third-party solutions (e.g. jupyterhub). We would customize them and integrate our PrimeHub graphql API and configure the OIDC client in our keycloak.
|Identify server for PrimeHub. It is responsible for managing users, groups, authentication.
|PrimeHub UI for users and administrators.
|Third-party multiple users jupyter project. We integrate it to spawn the jupyter servers.
|A special jupyter server for operation purposes.
|The primary API Server for PrimeHub.
|A metacontroller-based controller. It is responsible for provisioning an NFS server for a shared volume. (metacontroller is a general-purpose controller to implement a controller.)
|A metacontroller-based controller. It is responsible to manage the gitsync volume.
|A metacontroller-based controller. It is responsible to manage the volume upload server.
|The single process controller to manage jobs, image builder, license, etc. This component is relatively new and we hope to include all metacontroller-based controllers to this component in the future.
|It implements the admission controller to intercept the request the creation of resources. Currently, we use it to guarantee a pod does not request more resources than the quota.
|It monitors images, instance types, volumes custom resources and generates corresponding keycloak roles.
PrimeHub in the core does not have its database. The persistence state is stored in the keycloak and the kubernetes cluster.
The data respectively are
- Keycloak: Store the users, groups, user/group binding (member), roles, and group/role binding.
- Kubernetes: Store the common resources among groups, like
secrets. Or user-created items, like
In PrimeHub design, the common resources (e.g.
image) can be associated with groups. There is a corresponding role of this resource defined in keycloak. And the relationship between resource and group is implemented by role binding. The following diagram depicts the relationship.
For more information, please refer to data model documentation
- Graphql is the primary API server. We use it to control the resources in keycloak and cluster.
- For end-users request, graphql requires an id-token (defined in OpenID Connect) to access the graphql endpoint.
- For server-to-server requests, graphql requires a shared secret to access the graphql endpoint with full permission.
- Controllers watch the cluster/keycloak states and reconcile the current state to the desired state.
- Controllers may call GraphQL to get the user/group configuration. However, graphql should not know the controllers' existence.