SnowPro Specialty: Gen AI Question 42
Single answerUsing Snowpark Container ServicesA retail company wants to deploy an open-source large language model for internal product-support summarization. The model requires custom Python libraries and GPU resources, and the security team requires that inference stay inside Snowflake's governance boundary with access controlled through Snowflake roles. The engineering team also wants to expose the model as a scalable service that other Snowflake workloads can call. Which approach best meets these requirements?
- A
Package the model and dependencies into a container image, deploy it with Snowpark Container Services on a GPU-enabled compute pool, and expose inference through a Snowflake-managed service endpoint with role-based access controls.
- B
Create a Python UDF in Snowflake and install the model's system-level dependencies at runtime so the function can use attached GPUs from a virtual warehouse.
- C
Run the model on an external Kubernetes cluster and use external functions so Snowflake can invoke it while keeping all inference fully inside Snowflake's security boundary.
- D
Deploy the model as a Streamlit in Snowflake app because Streamlit sessions natively provide GPU-backed model serving and production-grade autoscaling for inference APIs.
Show answer and explanation
Correct answer: A
Explanation
The scenario points directly to Snowpark Container Services because the team needs all of the following at once: custom model runtime dependencies, GPU support, controlled in-platform execution, and service-style exposure for other Snowflake workloads. Snowpark Container Services supports deploying container images to compute pools, including GPU-enabled pools where available, which is the appropriate pattern for serving open-source Gen AI models that cannot run in simpler SQL or Python function abstractions. In contrast, Python UDFs and stored procedures are not a substitute for full containerized model serving, especially when OS-level packages or specialized inference stacks are required. External functions are useful for integration, but they execute against an external service and therefore do not satisfy the requirement to keep inference within Snowflake's governed boundary. Streamlit in Snowflake is valuable for user interfaces, not for scalable backend model serving. This aligns with Snowflake best practices for using Snowpark Container Services when workloads require custom containers, APIs/services, and specialized hardware such as GPUs.
- A. Correct.
Correct. Snowpark Container Services is designed for running containerized workloads inside Snowflake, including applications that need custom libraries, system packages, and GPU access. Using a GPU-enabled compute pool addresses the model's hardware requirements, while a Snowpark Container Services service can expose endpoints for inference. This approach keeps execution within Snowflake's governed environment and aligns with Snowflake's role-based access model.
- B. Incorrect.
Incorrect. Python UDFs are useful for in-database logic, but they are not the right mechanism for deploying a custom containerized LLM stack with system-level dependencies and direct GPU-backed model serving. You cannot rely on installing arbitrary OS-level dependencies at runtime in a UDF the way you would in a container image, and virtual warehouses are not how Snowpark Container Services GPU workloads are provisioned.
- C. Incorrect.
Incorrect. External functions can integrate Snowflake with services outside Snowflake, but that does not keep inference inside Snowflake's governance boundary. This option conflicts directly with the security requirement that inference remain internal to Snowflake-controlled execution.
- D. Incorrect.
Incorrect. Streamlit in Snowflake is intended for building interactive applications and dashboards, not as the primary mechanism for production model serving with custom containers and GPU-backed inference endpoints. It does not replace Snowpark Container Services for this use case.