TensorFlow Serving Compare a Candidate Version Against the Default

arazzo: 1.0.1
info:
  title: TensorFlow Serving Compare a Candidate Version Against the Default
  summary: Score the same instances against a pinned candidate version and the default version to measure rollout drift.
  description: >-
    The standard safeguard before promoting a newly trained model. TensorFlow
    ModelServer can hold several versions of the same model in memory at once,
    which makes it possible to send identical inputs to a candidate version and
    to the version currently serving default traffic, then compare the two
    responses before shifting any real traffic. This workflow confirms the
    candidate version has finished loading, reads its signature metadata to be
    sure it is input compatible with the version in production, and then runs the
    same instances through both. 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: version-canary-compare
  summary: Run identical instances through a candidate model version and the default version.
  description: >-
    Verifies the candidate version is AVAILABLE, confirms its signature, then
    predicts against the pinned candidate and against the model's default
    version so the two sets of predictions can be diffed.
  inputs:
    type: object
    required:
    - modelName
    - candidateVersion
    - instances
    properties:
      modelName:
        type: string
        description: The name the model was registered under in the ModelServer.
      candidateVersion:
        type: integer
        description: The version number of the newly trained candidate to evaluate (e.g. 2).
      signatureName:
        type: string
        description: Optional signature to target on both versions. Must be the same on each side for the comparison to be meaningful.
      instances:
        type: array
        description: >-
          The evaluation instances sent unchanged to both versions. Holding the
          input constant is what makes the two responses comparable.
        items: {}
  steps:
  - stepId: checkCandidateStatus
    description: >-
      Confirm the candidate version has finished loading and reports AVAILABLE.
      A candidate still in the LOADING state would otherwise fail the comparison
      for reasons that have nothing to do with model quality.
    operationId: getModelVersionStatus
    parameters:
    - name: model_name
      in: path
      value: $inputs.modelName
    - name: version
      in: path
      value: $inputs.candidateVersion
    successCriteria:
    - condition: $statusCode == 200
    - context: $response.body
      condition: $.model_version_status[0].state == 'AVAILABLE'
      type: jsonpath
    outputs:
      candidateState: $response.body#/model_version_status/0/state
      candidateErrorCode: $response.body#/model_version_status/0/status/error_code
  - stepId: readCandidateMetadata
    description: >-
      Read the candidate version's signature metadata to confirm it accepts the
      same inputs as the version currently serving default traffic.
    operationId: getModelVersionMetadata
    parameters:
    - name: model_name
      in: path
      value: $inputs.modelName
    - name: version
      in: path
      value: $inputs.candidateVersion
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      candidateSpecVersion: $response.body#/model_spec/version
      candidateSignatureMetadata: $response.body#/metadata
  - stepId: predictCandidate
    description: >-
      Score the evaluation instances against the pinned candidate version.
    operationId: predictModelVersion
    parameters:
    - name: model_name
      in: path
      value: $inputs.modelName
    - name: version
      in: path
      value: $inputs.candidateVersion
    requestBody:
      contentType: application/json
      payload:
        signature_name: $inputs.signatureName
        instances: $inputs.instances
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      candidatePredictions: $response.body#/predictions
  - stepId: predictDefault
    description: >-
      Score the identical instances against the model's default version, the one
      currently answering unpinned production traffic, to produce the baseline
      the candidate is judged against.
    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:
      defaultPredictions: $response.body#/predictions
  outputs:
    candidateSpecVersion: $steps.readCandidateMetadata.outputs.candidateSpecVersion
    candidatePredictions: $steps.predictCandidate.outputs.candidatePredictions
    defaultPredictions: $steps.predictDefault.outputs.defaultPredictions