Databricks ML Associate Practice Question 30: Batch vs Realtime Deployment
Databricks Certified Machine Learning Associate · Test 3 · question 10 of 10
A retailer scores all 20 million customers once a week to choose who receives a printed catalogue. The mailing house collects the scores from a table each Monday morning. No application queries the model interactively, and predictions computed on Sunday night are perfectly acceptable.
Which deployment approach fits these requirements at the lowest cost?
Show answer and explanation
Answer:
The workload has every characteristic of batch inference: the full population is known ahead of time, predictions are consumed from a table rather than requested one at a time, and a weekly refresh meets the freshness requirement. A scheduled job spins compute up, scores 20 million rows in a distributed pass, writes to Delta and shuts down, so the retailer pays only for the minutes the job runs.
Option B pays for an always-on endpoint all week to serve one burst of traffic, and 20 million sequential HTTP requests is far slower than a distributed scan of the same data.
Option C keeps compute running continuously to process a dataset that changes once a week. Streaming earns its cost when records arrive continuously and must be handled as they land, which is not the case here.
Option D reduces the idle cost of option B but keeps its shape. The mailing run still makes 20 million individual requests, and the first of them additionally pays a cold-start delay while the endpoint scales up from zero.
Reference: Deploy models with Mosaic AI Model Serving