H2O Driverless AI Helm - Error when deploying on ICP 3.1.1 / Kubernetes version 1.11.3 . Need Helm Chart Update

[bmarolleau]

Hello,
Here is the error I get when deploying on ICP 3.1.1 (no pb with ICP 2.x) due to the GPU management by K8s which changed with the latest versions of K8s. Symptom: POD scheduling error at helm install. With ICP 3.1 & 3.1.1 (and K8s version >= 11) nvidia.com/gpu should be used instead of alpha.kubernetes.io/nvidia-gpu, Here is a modified helm chart that works on my environment:
The critical part is in the Helm templates/deployment.yaml file in the requests/limits lines:

resources:
            limits:
            {{- if and (eq (.Capabilities.KubeVersion.Major|int) 1) (lt (.Capabilities.KubeVersion.Minor|int) 11) }}
              alpha.kubernetes.io/nvidia-gpu: {{ .Values.resources.limits.gpu }}
            {{- else }}
              nvidia.com/gpu: {{ .Values.resources.limits.gpu }}
            {{- end }}
              memory: {{ .Values.resources.limits.memory }}
            requests:
            {{- if and (eq (.Capabilities.KubeVersion.Major|int) 1) (lt (.Capabilities.KubeVersion.Minor|int) 11) }}
              alpha.kubernetes.io/nvidia-gpu: {{ .Values.resources.requests.gpu }}
            {{- else }}
              nvidia.com/gpu: {{ .Values.resources.requests.gpu }}
            {{- end }}
              memory: {{ .Values.resources.requests.memory }}

Here is the modified file to be placed in the templates folder of the helm chart, as an example: deployment.zip

[CreatureDev]

This issue will be fixed in the latest release