Leveraging Gemini Enterprise Agent Platform for end-to-end ML workflows

Silverline Bank keeps customer transaction features in BigQuery tables that are refreshed hourly. A fraud model trained on Agent Platform must use the same feature values at training time and at prediction time, and the online scoring service needs feature lookups in milliseconds. The data team wants BigQuery to remain the single source of truth and does not want to build a custom copy pipeline.

How should you prepare the data for the model?

1. AExport the BigQuery tables to CSV files in Cloud Storage for training and have the scoring service query BigQuery directly for each predictionThe CSV export creates a second copy for training, and querying BigQuery on every prediction does not provide millisecond-latency lookups.
2. BWrite a Dataflow job that copies features from BigQuery into Memorystore for Redis for the scoring serviceThis can serve features quickly, but it is the custom copy pipeline the team wants to avoid, and keeping training and serving values consistent becomes the team's responsibility.
3. CRegister the BigQuery tables as feature data sources in Agent Platform Feature Store, use them for training, and sync them to an online store for low-latency servingFeature Store uses BigQuery as its offline store, so training reads the same source, and an online store serves the same features at low latency without a custom pipeline.
4. DLoad the feature tables into Cloud SQL for PostgreSQL and use them for both training and servingMoving the data to Cloud SQL breaks the requirement that BigQuery stay the source of truth and adds a database to size and maintain.

Tidewater Robotics needs to fine-tune a vision model on 16 NVIDIA H100 GPUs for about four days. The job checkpoints only once a day, so it must not be preempted, but its start can wait a few days for capacity, and the company does not want any long-term commitment. The resulting model will be served to a few internal users whose requests arrive sporadically during business hours, and the company wants to pay nothing for serving accelerators when there are no requests.

Which two actions should you perform? Each correct answer presents part of the solution.

1. ARun the fine-tuning job on GKE with a GPU node pool that uses flex-start with queued provisioning, powered by Dynamic Workload SchedulerFlex-start with queued provisioning queues the request and provisions all requested GPU nodes together when capacity is available, the nodes run for up to seven days without the reclamation risk of Spot VMs, and there is no commitment, which fits a flexible start time.
2. BDeploy the model to a Cloud Run service with an NVIDIA L4 GPU attached and minimum instances set to zeroCloud Run supports GPUs for inference and scales to zero, so the company pays for GPU time only while requests are being served.
3. CRun the fine-tuning job on Spot VMs with H100 GPUsSpot VMs are cheapest but can be preempted at any time, and with only daily checkpoints a preemption could lose up to a day of training.
4. DPurchase a three-year committed use discount for 16 H100 GPUsA three-year commitment is billed for the full term, which contradicts the requirement to avoid long-term commitments for a four-day job.
5. EServe the model from a Compute Engine A3 VM with H100 GPUs that runs continuouslyAn always-on GPU VM bills around the clock, which violates the requirement to pay nothing for accelerators when there are no requests.

Harborview Energy retrains a demand-forecasting model every week by running notebooks by hand: extract data from BigQuery, train, evaluate, and deploy if the new model beats the current one. Auditors now require every run to be reproducible with tracked inputs, parameters, and resulting artifacts. The ML team wants the workflow to run weekly on its own, without operating any orchestration servers or clusters.

What should you implement?

1. AA Managed Service for Apache Airflow (formerly Cloud Composer) environment with a DAG that runs the notebooks on a Compute Engine VMManaged Airflow can schedule the steps, but it adds an environment to run and does not record ML artifacts, parameters, and lineage for each run without extra work.
2. BA pipeline defined with the Kubeflow Pipelines SDK and run on Agent Platform Pipelines on a weekly schedule, with a condition step that deploys the model only when evaluation metrics improveAgent Platform Pipelines runs pipelines serverlessly, supports scheduled runs and conditional steps, and automatically records artifacts, parameters, and lineage in its metadata store for reproducibility.
3. CA cron job on a Compute Engine VM that runs the training script and copies the model file to Cloud StorageA cron job automates the schedule but leaves a server to maintain and provides no built-in tracking of inputs, parameters, or lineage.
4. DA Cloud Build trigger that runs the training code each time a commit is pushedCloud Build is suited to CI for pipeline code, but commit-driven runs do not meet the weekly retraining schedule or provide ML lineage tracking.