arazzo: 1.0.1
info:
  title: TensorFlow Serving Route Inference by Version Label
  summary: Resolve a version label such as stable or canary to a concrete version, then run inference pinned to it.
  description: >-
    TensorFlow ModelServer supports version labels, which let an operator assign
    names like "stable" and "canary" to specific version numbers and move those
    labels as models are promoted. Callers can then ask for a label instead of
    hard coding a version, and the routing changes underneath them without a code
    deploy. This workflow resolves the label to the concrete version currently
    behind it, records that version so the run is auditable, confirms the
    signature, and pins the inference to the resolved version number so a label
    move mid-flight cannot split a single logical request across two models.
    Every step spells out its request inline so the flow can be read and executed
    without opening the underlying OpenAPI description.
  version: 1.0.0
sourceDescriptions:
- name: tensorflowServingApi
  url: ../openapi/tensorflow-serving-openapi.yml
  type: openapi
workflows:
- workflowId: label-routed-inference
  summary: Resolve a version label to a version number and predict against that pinned version.
  description: >-
    Looks up the model version currently carrying the supplied label, reads that
    version's metadata, and submits a prediction pinned to the resolved version.
  inputs:
    type: object
    required:
    - modelName
    - label
    - instances
    properties:
      modelName:
        type: string
        description: The name the model was registered under in the ModelServer.
      label:
        type: string
        description: >-
          The version label to route by, as configured on the ModelServer
          (commonly "stable" or "canary").
      signatureName:
        type: string
        description: Optional signature to target. When omitted the model's default serving signature is used.
      instances:
        type: array
        description: Row format prediction inputs, one prediction returned per instance.
        items: {}
  steps:
  - stepId: resolveLabel
    description: >-
      Resolve the label to the model version currently behind it and confirm that
      version is AVAILABLE. The resolved version number is captured as an output
      so the rest of the flow is pinned and the run stays reproducible even if
      the label is moved while the flow is executing.
    operationId: getModelLabelStatus
    parameters:
    - name: model_name
      in: path
      value: $inputs.modelName
    - name: label
      in: path
      value: $inputs.label
    successCriteria:
    - condition: $statusCode == 200
    - context: $response.body
      condition: $.model_version_status[0].state == 'AVAILABLE'
      type: jsonpath
    outputs:
      resolvedVersion: $response.body#/model_version_status/0/version
      resolvedState: $response.body#/model_version_status/0/state
      resolvedErrorCode: $response.body#/model_version_status/0/status/error_code
  - stepId: readResolvedMetadata
    description: >-
      Read the signature metadata of the version the label resolved to, so the
      caller knows exactly which model shape it is about to send inputs to.
    operationId: getModelVersionMetadata
    parameters:
    - name: model_name
      in: path
      value: $inputs.modelName
    - name: version
      in: path
      value: $steps.resolveLabel.outputs.resolvedVersion
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      modelSpecVersion: $response.body#/model_spec/version
      signatureMetadata: $response.body#/metadata
  - stepId: predictResolvedVersion
    description: >-
      Submit the prediction pinned to the version number resolved from the label,
      rather than to the label itself, so every instance in this request is
      scored by the same model.
    operationId: predictModelVersion
    parameters:
    - name: model_name
      in: path
      value: $inputs.modelName
    - name: version
      in: path
      value: $steps.resolveLabel.outputs.resolvedVersion
    requestBody:
      contentType: application/json
      payload:
        signature_name: $inputs.signatureName
        instances: $inputs.instances
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      predictions: $response.body#/predictions
      outputs: $response.body#/outputs
  outputs:
    resolvedVersion: $steps.resolveLabel.outputs.resolvedVersion
    modelSpecVersion: $steps.readResolvedMetadata.outputs.modelSpecVersion
    predictions: $steps.predictResolvedVersion.outputs.predictions