ARA-C01 Question 191
Single answerOutline basic AI/ML pipelines and applicationsA retail company stores clickstream, product catalog, and order history data in Snowflake. The data science team wants to build a churn prediction pipeline while minimizing data movement and operational complexity. They need a solution that supports SQL-based feature engineering in Snowflake, model training with external ML libraries, and batch scoring results written back to Snowflake for downstream dashboards. Which approach BEST meets these requirements?
- A
Use Snowpark to prepare features in Snowflake, use Snowpark ML or Snowpark Container Services to train with Python-based ML libraries close to the data, and write batch prediction results back into Snowflake tables.
- B
Export all Snowflake data daily to flat files in object storage, train the model on a separate server, and keep predictions only in the external environment to avoid loading them back into Snowflake.
- C
Use materialized views as the primary method to train and deploy the churn model because materialized views can execute ML algorithms directly inside Snowflake.
- D
Replicate operational tables to a second Snowflake account and use database replication as the main mechanism for model training and inference.
Show answer and explanation
Correct answer: A
Explanation
For a practical Snowflake-centered AI/ML pipeline, the recommended pattern is to engineer features in Snowflake using SQL and/or Snowpark, train models close to the data using Snowpark capabilities, and persist predictions back into Snowflake tables for reporting and operational consumption. This reduces data egress, improves governance, and simplifies architecture. Snowpark enables developers to use familiar languages such as Python while pushing processing closer to Snowflake data. Snowpark ML supports common machine learning workflow tasks, and Snowpark Container Services can be used when workloads require custom containers or broader ML frameworks. In contrast, exporting data to external systems increases operational complexity and data movement, materialized views are not ML tools, and replication is a data distribution capability rather than an ML pipeline component. These patterns are consistent with Snowflake guidance around Snowpark, Snowpark ML, and in-platform data application design for AI/ML use cases.
- A. Correct.
Correct. This approach aligns with Snowflake best practices for AI/ML pipelines: keep data in Snowflake for feature engineering, use Snowpark for Python to work close to the data, and leverage Snowpark ML and, when needed, Snowpark Container Services for more flexible model training and inference workloads using external ML libraries. Writing predictions back to Snowflake supports downstream analytics, BI, and operational use cases while minimizing unnecessary data movement.
- B. Incorrect.
Incorrect. While exporting data to object storage and training externally is possible, it increases data movement, introduces operational overhead, and conflicts with the requirement to minimize complexity. Keeping predictions only outside Snowflake also makes it harder to serve downstream dashboards and data consumers that already rely on Snowflake as the central data platform.
- C. Incorrect.
Incorrect. Materialized views improve performance for repeated query patterns, but they do not train or deploy machine learning models. A candidate might choose this option because feature pipelines often use derived datasets, but materialized views are a query optimization/storage feature, not an ML execution framework.
- D. Incorrect.
Incorrect. Database replication is designed for data availability, disaster recovery, and cross-region or cross-account distribution scenarios. It is not the main mechanism for building an ML pipeline. Replication does not provide model training logic or inference capabilities, so it does not satisfy the stated requirements.