TensorFlow Serving Preflight and Predict

arazzo: 1.0.1
info:
  title: TensorFlow Serving Preflight and Predict
  summary: Confirm a model is loaded and its signature is known before running prediction inference.
  description: >-
    The canonical TensorFlow Serving integration pattern. Firing inference at a
    ModelServer blind is the most common source of confusing 400s and empty
    predictions, because the model may still be LOADING, may have failed to load,
    or may expose a different signature than the caller assumes. This workflow
    checks the serving state of the model, reads its metadata to confirm the
    signature the request will target, and only then submits the prediction.
    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: inferenceApi
  url: ../openapi/tensorflow-inference-api-openapi.yml
  type: openapi
- name: modelsApi
  url: ../openapi/tensorflow-models-api-openapi.yml
  type: openapi
workflows:
- workflowId: predict-preflight
  summary: Check model status, read the signature, then run prediction inference.
  description: >-
    Verifies the named model reports an AVAILABLE version, resolves its signature
    metadata, and submits a prediction request in either row format (instances)
    or column format (inputs).
  inputs:
    type: object
    required:
    - modelName
    - instances
    properties:
      modelName:
        type: string
        description: The name the model was registered under in the ModelServer (e.g. "half_plus_two").
      signatureName:
        type: string
        description: >-
          Optional signature to target. When omitted the ModelServer uses the
          model's default serving signature.
      instances:
        type: array
        description: >-
          Row format input. An array of prediction inputs, where each element is
          a scalar, a list, or an object of named input tensors.
        items: {}
  steps:
  - stepId: checkModelStatus
    description: >-
      Confirm the ModelServer has the model loaded and reports a version in the
      AVAILABLE state before any inference is attempted.
    operationId: getModelStatus
    parameters:
    - name: model_name
      in: path
      value: $inputs.modelName
    successCriteria:
    - condition: $statusCode == 200
    - context: $response.body
      condition: $.model_version_status[0].state == 'AVAILABLE'
      type: jsonpath
    outputs:
      servingState: $response.body#/model_version_status/0/state
      servingVersion: $response.body#/model_version_status/0/version
      errorCode: $response.body#/model_version_status/0/status/error_code
  - stepId: readModelMetadata
    description: >-
      Read the metadata of the latest available version to confirm the signature
      the prediction request will target, along with its expected input tensors.
    operationId: getModelMetadata
    parameters:
    - name: model_name
      in: path
      value: $inputs.modelName
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      modelSpecName: $response.body#/model_spec/name
      modelSpecVersion: $response.body#/model_spec/version
      signatureMetadata: $response.body#/metadata
  - stepId: runPrediction
    description: >-
      Submit the prediction request against the model's latest available
      version, returning one prediction per supplied instance.
    operationId: predictModel
    parameters:
    - name: model_name
      in: path
      value: $inputs.modelName
    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:
    servingVersion: $steps.checkModelStatus.outputs.servingVersion
    modelSpecVersion: $steps.readModelMetadata.outputs.modelSpecVersion
    predictions: $steps.runPrediction.outputs.predictions