You used Vertex AI Workbench notebooks to build a model in TensorFlow. The notebook i) loads data from Cloud Storage, ii) uses TensorFlow Transform to pre-process data, iii) uses built-in TensorFlow operators to define a sequential Keras model, iv) trains and evaluates the model with model.fit() on the notebook instance, and v) saves the trained model to Cloud Storage for serving. You want to orchestrate the model retraining pipeline to run on a weekly schedule while minimizing cost and implementation effort. What should you do?

Using the Kubeflow Pipelines SDK is the best practice to orchestrate AI pipelines with modular steps. Vertex AI Workbench does not provide alerts, and you would have to log in every week to check the pipeline run status. This does not minimize monitoring steps. Cloud Composer does not provide ML-specific monitoring capabilities. Also, unless many pipelines are hosted in Cloud Composer, this solution is not the most cost-efficient. Separating each cell in the notebook into a containerised application and using Cloud Workflows requires more effort and does not follow best practices given that Vertex AI pipelines is the more suitable product to run modular containerised AI pipeline steps.