Skip to content

Setup Researcher Access Control

Introduction

The following instructions explain how to complete the configuration of access control for Researchers. Run:ai access control is at the Project level. When you assign Users to Projects, only these users are allowed to submit Jobs and access Jobs details.

This requires several steps:

  • Assign users to their Projects.
  • (Mandatory) Modify the Kubernetes entry point (called the Kubernetes API server) to validate credentials of incoming requests against the Run:ai Authentication authority.
  • (Command-line Interface usage only) Modify the Kubernetes profile to prompt the Researcher for credentials when running runai login (or oc login for OpenShift).

Administration User Interface Setup

Assign Users to Projects

Assign Researchers to Projects:

  • Open the Run:ai user interface and navigate to Users. Add a Researcher and assign it a Researcher role.
  • Navigate to Projects. Edit or create a Project. Use the Access Control tab to assign the Researcher to the Project.
  • If you are using Single Sign-On, you can also assign Groups. For more information see the Single Sign-On documentation.

Kubernetes Configuration

Important

As of Run:ai version 2.15, you only need to perform this step when accessing Run:ai from the command-line interface or sending YAMLs directly to Kubernetes

As described in authentication overview, you must direct the Kubernetes API server to authenticate via Run:ai. This requires adding flags to the Kubernetes API Server. The flags show in the Run:ai user interface under Settings | General | Researcher Authentication | Server configuration.

Modifying the API Server configuration differs between Kubernetes distributions:

  • Locate the Kubernetes API Server configuration file. The file's location may differ between different Kubernetes distributions. The location for vanilla Kubernetes is /etc/kubernetes/manifests/kube-apiserver.yaml
  • Edit the document, under the command tag, add the server configuration text described above.
  • Verify that the kube-apiserver-<master-node-name> pod in the kube-system namespace has been restarted and that changes have been incorporated. Run the below and verify that the oidc flags you have added:
kubectl get pods -n kube-system kube-apiserver-<master-node-name> -o yaml

No configuration is needed. Instead, Run:ai assumes that an Identity Provider has been defined at the OpenShift level and that the Run:ai Cluster installation has set the OpenshiftIdp flag to true. For more information see the Run:ai OpenShift control-plane setup.

Edit Rancher cluster.yml (with Rancher UI, follow this). Add the following:

cluster.yml
kube-api:
    always_pull_images: false
    extra_args:
        oidc-client-id: runai  # (1)
        ...
  1. These are example parameters. Copy the actual parameters from Settings | General | Researcher Authentication as described above.

You can verify that the flags have been incorporated into the RKE cluster by following the instructions here and running docker inspect <kube-api-server-container-id>, where <kube-api-server-container-id> is the container ID of api-server via obtained in the Rancher document.

If working via the RKE2 Quickstart, edit /etc/rancher/rke2/config.yaml. Add the parameters provided in the server configuration section as described above in the following fashion:

/etc/rancher/rke2/config.yaml
kube-apiserver-arg:
- "oidc-client-id=runai" # (1)
...
  1. These are example parameters. Copy the actual parameters from Settings | General | Researcher Authentication as described above.

If working via Rancher UI, need to add the flag as part of the cluster provisioning.

Under Cluster Management | Create, turn on RKE2 and select a platform. Under Cluster Configuration | Advanced | Additional API Server Args. Add the Run:ai flags as <key>=<value> (e.g. oidc-username-prefix=-).

Install Anthos identity service by running:

gcloud container clusters update <gke-cluster-name> \
    --enable-identity-service --project=<gcp-project-name> --zone=<gcp-zone-name>

Install the yq utility and run:

For username-password authentication, run:

kubectl get clientconfig default -n kube-public -o yaml > login-config.yaml
yq -i e ".spec +={\"authentication\":[{\"name\":\"oidc\",\"oidc\":{\"clientID\":\"runai\",\"issuerURI\":\"$OIDC_ISSUER_URL\",\"kubectlRedirectURI\":\"http://localhost:8000/callback\",\"userClaim\":\"sub\",\"userPrefix\":\"-\"}}]}" login-config.yaml
kubectl apply -f login-config.yaml

For single-sign-on, run:

kubectl get clientconfig default -n kube-public -o yaml > login-config.yaml
yq -i e ".spec +={\"authentication\":[{\"name\":\"oidc\",\"oidc\":{\"clientID\":\"runai\",\"issuerURI\":\"$OIDC_ISSUER_URL\",\"groupsClaim\":\"groups\",\"kubectlRedirectURI\":\"http://localhost:8000/callback\",\"userClaim\":\"email\",\"userPrefix\":\"-\"}}]}" login-config.yaml
kubectl apply -f login-config.yaml

Where the OIDC flags are provided in the Run:ai server configuration section as described above.

Then update runaiconfig with the Anthos endpoint - gke-oidc-envoy. Get the externel IP of the service in the Anthos namespace.

kubectl get svc -n anthos-identity-service
NAME               TYPE           CLUSTER-IP    EXTERNAL-IP     PORT(S)              AGE
gke-oidc-envoy     LoadBalancer   10.37.3.111   39.201.319.10   443:31545/TCP        12h

Add the IP to runaiconfig

kubectl -n runai patch runaiconfig runai -p '{"spec": {"researcher-service": {"args": {"gkeOidcEnvoyHost": "35.236.229.19"}}}}'  --type="merge"

To create a kubeconfig profile for Researchers run:

kubectl oidc login --cluster=CLUSTER_NAME --login-config=login-config.yaml \
    --kubeconfig=developer-kubeconfig

(this will require installing the kubectl oidc plug-in as described in the Anthos document above gcloud components install kubectl-oidc)

Then modify the developer-kubeconfig file as described in the Command-line Inteface Access section below.

  • In the AWS Console, under EKS, find your cluster.
  • Go to Configuration and then to Authentication.
  • Associate a new identity provider. Use the parameters provided in the server configuration section as described above. The process can take up to 30 minutes.

Please follow the "Vanilla Kubernetes" instructions

Please contact Run:ai customer support.

See specific instructions in the documentation of the Kubernetes distribution.

Command-line Interface Access

To control access to Run:ai (and Kubernetes) resources, you must modify the Kubernetes configuration file. The file is distributed to users as part of the Command-line interface installation.

When making changes to the file, keep a copy of the original file to be used for cluster administration. After making the modifications, distribute the modified file to Researchers.

  • Under the ~/.kube directory edit the config file, remove the administrative user, and replace it with text from Settings | General | Researcher Authentication | Client Configuration.
  • Under contexts | context | user change the user to runai-authenticated-user.

Test via Command-line interface

  • Run: runai login (in OpenShift environments use oc login rather than runai login).
  • You will be prompted for a username and password. In a single sign-on flow, you will be asked to copy a link to a browser, log in and return a code.
  • Once login is successful, submit a Job.
  • If the Job was submitted with a Project to which you have no access, your access will be denied.
  • If the Job was submitted with a Project to which you have access, your access will be granted.

You can also submit a Job from the Run:ai User interface and verify that the new job shows on the job list with your user name.

Test via User Interface

  • Open the Run:ai user interface, go to Jobs.
  • On the top-right, select Submit Job.

Tip

If you do not see the button or it is disabled, then you either do not have Researcher access or the cluster has not been set up correctly. For more information, refer to user interface overview.