Description
When deploying the AICR OCI bundle via ArgoCD on an EKS GB200 cluster using the `argocd-helm` deployer, the `gpu-operator` ArgoCD Application permanently shows `ComparisonError` / `Unknown` status. The app is functionally healthy (GPU Operator is running) but ArgoCD can never reconcile it.
Error Message
```
Failed to compare desired state to live state: failed to calculate diff:
error calculating structured merge diff: error building typed value from
config resource: .spec.kataSandboxDevicePlugin: field not declared in schema
```
Root Cause
The gpu-operator Helm chart (v26.3.1) defaults `kataSandboxDevicePlugin.enabled: true`, which renders `.spec.kataSandboxDevicePlugin` into the ClusterPolicy manifest. The ClusterPolicy CRD installed by `gpu-operator-pre` (v26.3.1) does not declare this field in its OpenAPI schema. ArgoCD's structured merge diff engine fails when the desired manifest contains a field that is absent from the live CRD schema.
The field was added to the CRD schema in v26.3.2, so upgrading the chart resolves the ComparisonError. However, even on v26.3.2+, the default `enabled: true` will deploy the kata sandbox device plugin on clusters that do not use kata containers.
Fix Required
Add `kataSandboxDevicePlugin.enabled: false` to the gpu-operator component values for EKS (or globally for deployments that don't use kata containers):
```yaml
kataSandboxDevicePlugin:
enabled: false
```
This prevents the field from being rendered into the ClusterPolicy, eliminating the ComparisonError on v26.3.1 and avoiding unnecessary kata plugin deployment on v26.3.2+.
Environment
- Cluster: EKS GB200, AWS us-east-1
- ArgoCD: v3.4.2
- gpu-operator chart: v26.3.1
- Deployer: `argocd-helm`
- Observed: `gpu-operator` ArgoCD Application stuck in `Unknown` health with `ComparisonError` condition indefinitely after fresh bundle deployment
Description
When deploying the AICR OCI bundle via ArgoCD on an EKS GB200 cluster using the `argocd-helm` deployer, the `gpu-operator` ArgoCD Application permanently shows `ComparisonError` / `Unknown` status. The app is functionally healthy (GPU Operator is running) but ArgoCD can never reconcile it.
Error Message
```
Failed to compare desired state to live state: failed to calculate diff:
error calculating structured merge diff: error building typed value from
config resource: .spec.kataSandboxDevicePlugin: field not declared in schema
```
Root Cause
The gpu-operator Helm chart (v26.3.1) defaults `kataSandboxDevicePlugin.enabled: true`, which renders `.spec.kataSandboxDevicePlugin` into the ClusterPolicy manifest. The ClusterPolicy CRD installed by `gpu-operator-pre` (v26.3.1) does not declare this field in its OpenAPI schema. ArgoCD's structured merge diff engine fails when the desired manifest contains a field that is absent from the live CRD schema.
The field was added to the CRD schema in v26.3.2, so upgrading the chart resolves the ComparisonError. However, even on v26.3.2+, the default `enabled: true` will deploy the kata sandbox device plugin on clusters that do not use kata containers.
Fix Required
Add `kataSandboxDevicePlugin.enabled: false` to the gpu-operator component values for EKS (or globally for deployments that don't use kata containers):
```yaml
kataSandboxDevicePlugin:
enabled: false
```
This prevents the field from being rendered into the ClusterPolicy, eliminating the ComparisonError on v26.3.1 and avoiding unnecessary kata plugin deployment on v26.3.2+.
Environment