SnowPro Specialty: Gen AI Question 41
Single answerUsing Snowpark Container ServicesA GenAI team wants to deploy a custom inference service in Snowpark Container Services (SPCS) to serve a domain-specific embedding model that is not available as a managed Snowflake model. The service must read source documents from an internal stage, write generated embeddings back to Snowflake tables, and avoid any public internet exposure. Which design choice best meets these requirements with the least operational complexity?
- A
Deploy the model as a Snowpark Container Services service, grant it access to the required stage and tables through Snowflake roles and service configuration, and expose it only through Snowflake-managed internal connectivity.
- B
Run the model in an external Kubernetes cluster, expose a public HTTPS endpoint, and connect Snowflake to it through external functions so the service can read and write Snowflake data directly.
- C
Package the model as a Python UDF so it can download model files from the internet at runtime and write embeddings to internal stages without using a container service.
- D
Use a Snowflake Native App to host the model weights and execute GPU inference inside a standard virtual warehouse, since warehouses provide the same container runtime as Snowpark Container Services.
Show answer and explanation
Correct answer: A
Explanation
Snowpark Container Services is the appropriate Snowflake feature when teams need to run custom containerized applications or model inference services close to Snowflake data. In this scenario, the key requirements are: using a custom embedding model not provided as a managed Snowflake model, reading from internal stages, writing results back to Snowflake tables, and avoiding public internet exposure. SPCS fits because it supports containerized workloads within Snowflake's security and governance model, allowing controlled access to Snowflake resources and private service exposure patterns. By contrast, external Kubernetes plus external functions adds infrastructure and networking complexity; Python UDFs are not intended to replace containerized inference services with custom runtimes; and Native Apps and warehouses do not provide the same execution model as SPCS. Candidates should recognize SPCS as the best-practice pattern for secure, in-platform deployment of custom GenAI inference services that need tight integration with Snowflake data.
- A. Correct.
Correct. Snowpark Container Services is designed for running custom containerized workloads inside Snowflake, including model inference services. This approach lets the team keep the service close to Snowflake data, use Snowflake security constructs for controlled access to stages and tables, and avoid public internet exposure by using Snowflake-managed internal connectivity rather than publishing a public endpoint. It is the most direct and operationally simple option for a custom model that must interact with Snowflake data securely.
- B. Incorrect.
Incorrect. An external Kubernetes deployment can work architecturally, but it does not meet the goal of least operational complexity and introduces public endpoint management, network security, and additional infrastructure overhead. External functions are also not the primary pattern for a stateful custom inference service that needs close integration with Snowflake stages and tables. This option reflects a common misconception that any external endpoint is equivalent to running the workload natively in Snowflake.
- C. Incorrect.
Incorrect. Python UDFs are useful for certain in-database logic, but they are not a substitute for Snowpark Container Services when you need a custom containerized inference stack, specialized dependencies, or service-style model serving. The statement about downloading model files from the internet at runtime is also problematic in controlled Snowflake execution environments and does not align with the requirement to avoid public internet exposure. This distractor targets the misconception that UDFs can replace container services for all ML inference workloads.
- D. Incorrect.
Incorrect. A Snowflake Native App is a packaging and distribution mechanism, not a replacement for Snowpark Container Services runtime execution. Standard virtual warehouses do not provide the same container runtime as SPCS, and they are not how teams run arbitrary containerized GPU-backed inference services. This option confuses application packaging and warehouse compute with container orchestration capabilities.