Preparations

Run the following to enable image download from the Run:ai Container Registry on Google cloud:

kubectl create namespace runai-backend
kubectl apply -f runai-gcr-secret.yaml

To extract Run:ai files, replace <VERSION> in the command below and run:

tar xvf runai-<version>.tar.gz
cd deploy

kubectl create namespace runai-backend

Install the Run:ai Administrator Command-line Interface by following the steps here.

Use the Run:ai Administrator Command-line Interface located in the deploy folder. To allow running the binary, run:

chmod +x runai-adm

See https://helm.sh/docs/intro/install/ on how to install Helm. Run:ai works with Helm version 3 only (not helm 2).

The Helm installation image is under the deploy directory. Run:

tar xvf helm-<version>-linux-amd64.tar.gz
sudo mv linux-amd64/helm /usr/local/bin/

In previous versions, you set nodes that were dedicated to Run:ai software:

• Setting the nodes was mandatory.
• Run:ai was confined to these nodes. If the selected nodes failed or lacked resources, Run:ai would stop working.

In version 2.9,

• There is no need to set nodes for Run:ai.
• You can optionally set the Run:ai control plane to run on specific nodes.
• Run:ai is no longer confined to these notes. Instead, Kubernetes will attempt to schedule Run:ai pods to these nodes. If lacking resources, the Run:ai nodes will move to another, non-labeled node.

To set system worker nodes run:

kubectl label node <NODE-NAME> node-role.kubernetes.io/runai-system=true

If you want to preserve the previous functionality, please contact Run:ai customer support.

The Run:ai control plane should be installed on a set of dedicated Run:ai system worker nodes rather than GPU worker nodes. To set system worker nodes run:

kubectl label node <NODE-NAME> node-role.kubernetes.io/runai-system=true

To avoid single-point-of-failure issues, we recommend assigning more than one node in production environments.