Provider

The Provider resource represents the provider specific configuration required to submit / update / delete ml resources with the given provider. e.g Vertex AI Platform. Providers configuration can be set using this resource and permissions for access can be configured via service accounts.

Note: changing the provider of a resource that was previously managed by another provider will result in a resource error.

Any referenced resources must always match the provider of the referencing resource (e.g. RunConfiguration to Pipeline) as updates are not propagated or checked and will result in runtime errors on the provider.

Common Fields

NameDescriptionExample
spec.serviceImageContainer image of the provider servicekfp-operator-kfp-provider-service:0.0.2
spec.executionModeKFP compiler execution mode Note: This field has been deprecated after v0.7.0 v1beta1 api specification changes, and will be removed in future versions.v1 (currently KFP) or v2 (Vertex AI)
spec.serviceAccountService Account name to be used for all provider-specific operations (see respective provider)kfp-operator-vertex-ai
spec.pipelineRootStorageThe storage location used by TFX (pipeline-root) to store pipeline artifacts and outputs - this should be a top-level directory and not specific to a single pipelinegcs://kubeflow-pipelines-bucket
spec.parametersParameters specific to each provider, i.e. VAIgcs://kubeflow-pipelines-bucket
spec.frameworks[]A list of frameworks supported by the provider.
spec.allowedNamespaces[]A list of namespaces that resources can reference this provider from. If a resource tries to reference this provider from a namespace not in the allowedNamespaces list, the resource will fail. If no allowedNamespaces list is configured, then resources can reference this provider from any namespace.- default

Framework

NameDescriptionExample
nameName of the framework.tfx
imageFramework image.ghcr.io/kfp-operator/kfp-operator-tfx-compiler:version-tag
patches[]A list of JSON patches, that will be applied to every Pipeline resource that uses this Provider before it’s passed (as JSON) to the corresponding Argo Workflow for processing.

Patch

NameDescriptionExample
typeThe type of patch to be applied to the pipeline resource definition JSON. Can be either json (RFC6902) or merge (RFC7396).json
payloadThe patch to be applied to the pipeline resource definition JSON.[{ "op": "add", "path": "/framework/parameters/beamArgs/0", "value": { "name": "newArg", "value": "newValue" } }]

Vertex AI

apiVersion: pipelines.kubeflow.org/v1beta1
kind: Provider
metadata:
  name: vai
  namespace: kfp-operator
spec:
  serviceImage: kfp-operator-vai-provider-service:<version>
  pipelineRootStorage: gs://<storage_location>
  serviceAccount: kfp-operator-vai
  parameters:
    eventsourcePipelineEventsSubscription: kfp-operator-vai-run-events-eventsource
    maxConcurrentRunCount: 1
    pipelineBucket: pipeline-storage-bucket
    vaiJobServiceAccount: kfp-operator-vai@<project>.iam.gserviceaccount.com
    vaiLocation: europe-west2
    vaiProject: <project>
  frameworks:
  - name: tfx
    image: ghcr.io/kfp-operator/kfp-operator-tfx-compiler:version-tag
    patches:
    - type: json
      patch: |
        [
          {
            "op": "add",
            "path": "/framework/parameters/beamArgs/0",
            "value": {
              "name": "project",
              "value": "<project>"
            }
          }
        ]

Vertex AI Specific Parameters

NameDescription
eventsourcePipelineEventsSubscriptionThe eventsource subscription used to capture run-completion events
maxConcurrentRunCountThe number of pipelines that may run concurrently
pipelineBucketThe output storage bucket for a trained pipeline model
vaiJobServiceAccountThe service account should be used by VAI when submitting a pipeline
vaiLocationThe region VAI should run a pipeline within
vaiProjectThe project VAI should run a pipeline within