Question

A retail analytics team has trained a custom model in Vertex AI to score customer records for likelihood to purchase. Each night, they must score a batch of approximately 40 million records stored as CSV files in a Cloud Storage bucket. Latency is not important, but the job must run without maintaining any always-on serving infrastructure, and results should be written back to Cloud Storage. Which approach best meets these requirements?

Accepted Answer

Submit a Vertex AI batch prediction job that references the model and the Cloud Storage input files, writing predictions to a Cloud Storage output location.

Answer

Deploy the model to a Vertex AI online prediction endpoint and send the records via streaming requests using the SDK.

Answer

Export the model to a Compute Engine VM with a GPU and run a custom Python script that loops over each record and calls the local model.

Answer

Load the CSV files into BigQuery and use a scheduled query to invoke the online prediction endpoint for each row.