Building Models Given the Task Considering Cost, Complexity, Latency, and Scalability PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 3)

Questions 11–15

1. 11expert · hard

   A logistics company wants to predict package delivery times. They have a model that is accurate but requires 500ms to run on a CPU. The company needs predictions in under 200ms to integrate with their tracking system. They are considering optimizing the model, using a GPU endpoint, or using batch predictions. The model is a complex ensemble. What is the most appropriate approach?

   • ADeploy to edge devices at each package center.
   • BUse a GPU endpoint to reduce latency below 200ms.
   • CSimplify the model to a linear regression to reduce latency.
   • DUse batch predictions every hour.
   Reveal Solution Discussion Select an answer first
2. 12expert · hard

   A weather forecasting company needs to predict temperature for the next 7 days for thousands of locations. They have historical data and need to update forecasts every hour. The team has limited ML expertise and wants to minimize cost. They are considering ARIMA, a DNN, and BigQuery ML. The DNN is the most accurate but requires significant expertise and compute. ARIMA is less accurate but simpler. BigQuery ML offers managed ARIMA. What is the most appropriate approach?

   • AUse a pre-trained LLM to forecast from historical text reports.
   • BTrain a custom DNN on Vertex AI and deploy as an online endpoint.
   • CUse a simple ARIMA model on a single VM.
   • DUse BigQuery ML's ARIMA_PLUS for managed, scalable forecasting.
   Reveal Solution Discussion Select an answer first
4. 13foundation · easy

   A bank must explain why a loan application was rejected, and regulators require a human-understandable reason. Which modeling approach best meets this interpretability requirement?

   • AA large language model with natural language explanations
   • BA logistic regression model
   • CA gradient-boosted tree with feature importance
   • DA deep neural network with SHAP values
   Reveal Solution Discussion Select an answer first
5. 14application · medium

   A ride-sharing company wants to predict rider demand for the next hour to pre-position drivers. The model must provide predictions every 10 minutes and handle spikes during events. The team has a trained model and wants to minimize serving cost while meeting latency requirements. Which deployment strategy is most appropriate?

   • ARun batch predictions every 10 minutes on a scheduled basis.
   • BDeploy to edge devices in each vehicle.
   • CUse BigQuery ML to run predictions on streaming data.
   • DDeploy as an online endpoint with autoscaling.
   Reveal Solution Discussion Select an answer first
6. 15expert · hard

   A hospital network wants to deploy a model that predicts patient deterioration from vital signs. The model must run on bedside monitors with no internet connection and provide predictions every second. The model is a complex ensemble that requires significant compute. The team is considering simplifying the model to a decision tree to fit on the devices. What is the most appropriate approach?

   • AUse a larger model on the edge device with a GPU.
   • BUse batch prediction on data collected every hour.
   • CSimplify the model to a decision tree and deploy to edge devices.
   • DDeploy the complex ensemble to a cloud endpoint and rely on Wi-Fi.
   Reveal Solution Discussion Select an answer first