Sync Configuration¶

The ArgoCD ApplicationSets use production-grade sync options validated through testing.

Sync Options¶

Option	Purpose
`ServerSideApply=true`	Handles large CRDs (DSC, InferenceService, ClusterPolicy) and prevents annotation size limits
`SkipDryRunOnMissingResource=true`	Allows retry-based convergence when CRDs don't exist yet
`CreateNamespace=true`	ArgoCD manages namespace lifecycle for use cases
`RespectIgnoreDifferences=true`	Honors configured `ignoreDifferences` rules

ignoreDifferences Rules¶

OLM Subscriptions -- Prevents perpetual drift on .status and .spec.startingCSV:

ignoreDifferences:
  - group: operators.coreos.com
    kind: Subscription
    jsonPointers:
      - /spec/startingCSV
      - /status

InferenceService, ServingRuntime, Route (model and service ApplicationSets) -- Prevents drift from KServe controllers and OpenShift router:

ignoreDifferences:
  - group: serving.kserve.io
    kind: InferenceService
    jsonPointers:
      - /status
      - /metadata/annotations
  - group: serving.kserve.io
    kind: ServingRuntime
    jsonPointers:
      - /status
  - group: route.openshift.io
    kind: Route
    jsonPointers:
      - /spec/host
      - /status

ClusterPolicy, NodeFeatureDiscovery (instances ApplicationSet) -- Prevents drift from operator-managed status:

ignoreDifferences:
  - group: nvidia.com
    kind: ClusterPolicy
    jsonPointers:
      - /status
  - group: nfd.openshift.io
    kind: NodeFeatureDiscovery
    jsonPointers:
      - /status

DataScienceCluster -- Prevents drift on /status and operator-managed component fields:

ignoreDifferences:
  - group: datasciencecluster.opendatahub.io
    kind: DataScienceCluster
    jsonPointers:
      - /status
      - /spec/components/dashboard
      - /spec/components/kserve
      - /spec/components/ray
      - /spec/components/trainingoperator
      - /spec/components/modelmeshserving
      - /spec/components/codeflare
      - /spec/components/datasciencepipelines
      - /spec/components/workbenches
      - /spec/components/modelregistry
      - /spec/components/trustyai

Retry Policies¶

Layer	Max Retries	Backoff	Max Duration
Operators	5	30s (factor 2)	5 min
Instances	10	60s (factor 2)	10 min
Models	10	60s (factor 2)	10 min
Services	10	60s (factor 2)	10 min

The higher retry count for instances and use cases gives operators time to install their CRDs before ArgoCD attempts to apply instance resources.

RHOAI 3.3 Specifics¶

Setting	Value	Notes
RHOAI channel	`fast-3.x`	Required for 3.x releases
DSC API	`datasciencecluster.opendatahub.io/v1`	Stable API for 3.3
Kueue	`Unmanaged` in DSC	Red Hat Build of Kueue Operator manages it separately
JobSet	Standalone operator	Required for Kubeflow Trainer v2
GPU Operator	Requires `spec.daemonsets` and `spec.dcgm`	Validated with v25.x
Kueue instance	Requires `spec.config.integrations.frameworks`	List of supported job frameworks