Question

A financial services company runs a fraud-detection application on Compute Engine. Data scientists manually export transaction logs to a VM, run Python scripts to train models weekly, and copy the resulting model files to a bucket for the serving application to load. Leadership asks the architect to envision future improvements that reduce operational toil, enable automated retraining as new data arrives, and provide reproducible, versioned model deployments — while minimizing the team's infrastructure management burden. Which evolution of the architecture best meets these goals?

Accepted Answer

Move the training scripts into a Vertex AI Pipelines workflow that reads from BigQuery, triggers retraining on schedule or on new data, and registers versioned models in the Vertex AI Model Registry for managed endpoint deployment.

Answer

Increase the training VM to a larger machine type with attached GPUs and add a cron job so the scripts run automatically each night.

Answer

Containerize the training scripts and deploy them as a Deployment on a self-managed Google Kubernetes Engine cluster, storing models in Persistent Disks.

Answer

Create a managed instance group of training VMs behind a load balancer so multiple training jobs can run concurrently during peak periods.