GCP Professional ML Engineer Practice Question 4: Notebook Prototype to Training Job
Google Cloud Professional Machine Learning Engineer · Test 1 · question 4 of 10
Your prototype trains a PyTorch model inside an Agent Platform Workbench notebook on a single attached GPU. The model must now be retrained nightly on fresh data as one step of an automated retraining workflow, the run must be reproducible from a pinned artifact months later, and you do not want to keep a notebook instance running between runs.
Which approach meets these requirements with the least operational overhead?
Show answer and explanation
Answer:
Custom training on Gemini Enterprise Agent Platform (formerly Vertex AI Training) is the serverless path from a notebook prototype to production: you submit a job, Google Cloud provisions the GPUs, runs the container, and releases the resources when it finishes, so nothing is billed or maintained between nightly runs. A pinned container image digest is the reproducible artifact the stem asks for, and a custom training job is a first-class step in an ML pipeline, which is how the nightly workflow gets orchestrated.
Option A is the tempting near-miss. Scheduled notebook execution is real and convenient, but a notebook is a poor production artifact (cell order and hidden kernel state undermine reproducibility), and keeping a Workbench instance sized for the job leaves idle GPU capacity attached between runs.
Option B and option C both work, and both hand you infrastructure you did not ask for. Managing an instance template, image updates and cron plumbing (B), or a GKE cluster with a GPU node pool, drivers and autoscaler (C), is substantially more operational overhead than submitting a job to a managed training service, and neither gives you native integration with pipeline orchestration or Model Registry.
Reference: Serverless training overview