Google Professional Machine Learning Engineer Question 359
Select 3Google Cloud PlatformYou are tasked with designing an end-to-end ML pipeline for a retail company to predict customer churn using Google Cloud services. The pipeline must handle data ingestion, preprocessing, training, and deployment. The training data resides in BigQuery, and the model needs to serve predictions in real-time. What are the most appropriate components to include in your pipeline to meet these requirements?
- A
Use Dataflow to preprocess data and export the processed data to Cloud Storage for training.
- B
Use AI Platform Training to train the model with the preprocessed data stored in Cloud Storage.
- C
Use BigQuery ML to train the model directly on the data in BigQuery.
- D
Deploy the trained model to Vertex AI for real-time predictions.
- E
Use Pub/Sub for serving real-time predictions to end users.
Show answer and explanation
Correct answers: A, B, D
Explanation
The optimal pipeline for this scenario involves using Dataflow for preprocessing data and exporting it to Cloud Storage, training the model with AI Platform Training, and deploying it to Vertex AI for real-time predictions. BigQuery ML is excluded because it lacks support for advanced model deployment and high-performance real-time predictions. Pub/Sub, while useful for other applications, is not designed for serving real-time predictions.
- A. Correct.
Correct. Dataflow is an appropriate choice for preprocessing large datasets and moving processed data to Cloud Storage for training.
- B. Correct.
Correct. AI Platform Training (now part of Vertex AI) is a suitable service for training custom models using data stored in Cloud Storage.
- C. Incorrect.
Incorrect. While BigQuery ML is an excellent choice for training models directly on data in BigQuery, it does not meet the requirement for real-time predictions in this scenario.
- D. Correct.
Correct. Vertex AI is designed for deploying trained models with support for real-time predictions.
- E. Incorrect.
Incorrect. Pub/Sub is a messaging system and not suitable for serving real-time predictions. It is better suited for event-driven architectures or streaming data.