Google Professional Machine Learning Engineer Question 383
Select 3Google Cloud PlatformYou are tasked with building an automated machine learning pipeline for training and deploying a model on Google Cloud. The pipeline should retrain the model whenever new data is uploaded to a Cloud Storage bucket. The training process involves pre-processing data, training the model, and evaluating its performance. Which components and triggers should you use to meet the requirements?
- A
Use Cloud Storage notifications to trigger a Cloud Function whenever new data is uploaded, which then invokes Cloud Build to execute the pipeline.
- B
Use a Pub/Sub topic to receive data upload events from Cloud Storage and trigger a Vertex AI Pipelines job.
- C
Use Cloud Scheduler to periodically trigger a Cloud Build job for retraining the model, regardless of new data availability.
- D
Use Dataflow to pre-process the data and send a notification to a Pub/Sub topic, which triggers Cloud Build to train and evaluate the model.
- E
Use Vertex AI Training to directly monitor the Cloud Storage bucket and train the model automatically.
Show answer and explanation
Correct answers: A, B, D
Explanation
The correct solution should dynamically respond to new data in the Cloud Storage bucket and execute the pipeline steps efficiently. Cloud Storage notifications, Pub/Sub, Cloud Build, and Dataflow are key components for building such a pipeline. Cloud Scheduler is unsuitable as it is time-based, and Vertex AI Training cannot directly monitor bucket changes without external triggers. Combining these tools ensures a robust and scalable pipeline for machine learning tasks.
- A. Correct.
This is a valid approach since Cloud Storage notifications can trigger a Cloud Function, which can orchestrate the pipeline by invoking Cloud Build for tasks such as data pre-processing, training, and evaluation.
- B. Correct.
This is a valid option because Pub/Sub can act as an intermediary to receive data upload notifications from Cloud Storage and trigger downstream Vertex AI Pipelines for training and evaluation.
- C. Incorrect.
This is not an optimal solution because Cloud Scheduler is time-based and does not respond dynamically to new data uploads, making it inefficient for this use case.
- D. Correct.
This is a valid approach because Dataflow can handle complex data pre-processing, and the Pub/Sub topic can be used to trigger further steps such as training and evaluation through Cloud Build.
- E. Incorrect.
This is incorrect because Vertex AI Training does not have a built-in mechanism to directly monitor a Cloud Storage bucket for changes. External triggers are required.