Skip to main content

Model training and deployment life cycle

A model is trained and deployed if all the following chain of components are successful:

  1. Train pipeline will put the models in the following places:

    • Clustering: gs://{project}-jobs-{account}/{profile}/{job_name}/classifier/MLPTF_300_TfIdf_Splt_Agent/1
    • Matching: gs://{project}-jobs-{account}/{profile}/{job_name}/matching/MLPTF_300_TfIdf_Splt_Agent/1
    • URL: gs://{project}-jobs-{account}/{profile}/{job_name}/MLPTF_300_TfIdf_Splt_Agent/1

  2. Deployment component:

    • The new model is picked up from the above paths and deployed in the servable path: \ngs://{project}-servables-{account}/models/{profile}-{content_type}-[url|response]-recommender/{version}/\n(for URL models content_type = "default")

  3. Pending model monitor:

    • The latest model as deployed in the above path is picked up
    • The pending model deployment is correctly validating and promoting the model
      • For this, it is important to check that all the newly found redactions and URLs are reviewed, otherwise the validation will fail
    • This should lead to last model in "Admin/ML Models" to be set to "redacted"
    • Finally this model should appear in "Admin/Profiles/active [url|text] model"

  4. Finally the model should be working on staging.