Google Professional Machine Learning Engineer Question 378
Select 2Google Cloud PlatformYou are tasked with deploying a machine learning workflow built on MLflow to Google Cloud. The workflow includes experiment tracking, model storage, and serving predictions. Which of the following steps would you need to take to host the MLflow pipeline on Google Cloud while ensuring scalability and integration with GCP services?
- A
Deploy MLflow tracking server on Google Kubernetes Engine (GKE) and configure it to use Cloud SQL as the backend database for storing metadata.
- B
Use Cloud Storage as the artifact store for MLflow to store models and other artifacts.
- C
Set up Vertex AI Model Registry to replace MLflow's model registry for storing and serving models.
- D
Integrate MLflow with BigQuery to query experiment results directly from the database.
- E
Use Cloud Run to deploy the MLflow tracking server and configure it to scale automatically based on traffic.
Show answer and explanation
Correct answers: A, B
Explanation
To host third-party pipelines like MLflow on Google Cloud, you need to ensure scalability, integration with GCP services, and proper configurations for metadata and artifact storage. Deploying the MLflow tracking server on GKE enables scalability and flexibility, while using Cloud SQL and Cloud Storage ensures reliable and cost-effective storage solutions. These choices align with Google Cloud best practices for hosting third-party pipelines.
- A. Correct.
Deploying the MLflow tracking server on GKE allows you to run MLflow in a scalable and managed environment. Configuring Cloud SQL as the backend database ensures persistent and reliable storage of experiment metadata.
- B. Correct.
Using Cloud Storage as the artifact store is a recommended approach for storing large files (e.g., models, datasets) in a scalable and cost-effective manner that integrates well with other GCP services.
- C. Incorrect.
Vertex AI Model Registry is a GCP-native service and cannot be directly used as a replacement for MLflow's model registry. While Vertex AI offers similar functionality, it is not an extension of MLflow.
- D. Incorrect.
While integrating with BigQuery can be useful for analytics, MLflow does not natively support querying experiment results directly from BigQuery. Additional custom configurations would be needed.
- E. Incorrect.
Although Cloud Run is a serverless option, it may not be ideal for hosting MLflow tracking servers due to the need for persistent storage and handling complex request patterns. GKE is better suited for this use case.