ARA-C01 Question 194
Single answerSnowpark Container ServicesA financial services company wants to deploy a fraud-detection inference service using Snowpark Container Services. The service uses a custom Python model server packaged as a container image and must score requests from SQL and application clients with low latency. Security requirements state that model artifacts and reference data must remain inside Snowflake-managed governance boundaries, and the platform team wants to avoid managing external Kubernetes infrastructure. Which approach best meets these requirements?
- A
Deploy the model server as a Snowpark Container Services service, store model artifacts and reference data in Snowflake-managed storage such as stages or tables, and expose the service through an endpoint that can be invoked from supported clients within Snowflake's security model.
- B
Run the model server on an external Kubernetes cluster, connect it to Snowflake through an external function, and cache model artifacts locally on the cluster to reduce latency.
- C
Package the model as a Java UDF and store the container image in an internal stage so that SQL clients can call the container directly without creating a service.
- D
Create a Snowpark-optimized warehouse, install the model server into the warehouse runtime, and expose it as a permanent stored procedure for low-latency online inference.
Show answer and explanation
Correct answer: A
Explanation
The key architectural requirement is to run a custom containerized inference application inside Snowflake while keeping data and operational control within Snowflake boundaries. Snowpark Container Services addresses exactly this use case by allowing teams to deploy containerized services in Snowflake-managed infrastructure rather than external Kubernetes. This is appropriate when the workload requires custom runtimes, model servers, or APIs that go beyond standard UDF or stored procedure patterns. Best practice is to keep model artifacts and supporting data in Snowflake-managed storage such as stages and tables where possible, so governance, access control, and auditing remain centralized. External functions and external Kubernetes are valid integration patterns in other scenarios, but they are less suitable here because the company explicitly wants to avoid external infrastructure and retain governance inside Snowflake. Relevant Snowflake documentation areas include Snowpark Container Services concepts, services and endpoints, image repositories, and architecture guidance comparing container services with UDFs, stored procedures, and external functions.
- A. Correct.
Correct. Snowpark Container Services is designed for running containerized workloads directly in Snowflake, which fits the requirement to avoid external Kubernetes management. Keeping model artifacts and reference data in Snowflake stages or tables preserves Snowflake governance boundaries. Exposing the containerized application as a service aligns with low-latency inference patterns for internal platform use cases.
- B. Incorrect.
Incorrect. An external Kubernetes cluster violates the requirement to avoid managing external infrastructure and moves part of the solution outside Snowflake governance boundaries. External functions are useful for invoking external services, but they are not the best fit when the goal is to keep the model-serving stack and data governance within Snowflake-managed boundaries.
- C. Incorrect.
Incorrect. Java UDFs do not provide a way to run arbitrary container images directly from an internal stage. Snowpark Container Services requires defining and running a service, not calling a container image as if it were a UDF. This option reflects a common misconception that all executable logic in Snowflake can be turned into a UDF.
- D. Incorrect.
Incorrect. Warehouses execute SQL and Snowpark workloads but are not a mechanism for installing and hosting arbitrary long-running model servers. Snowpark-optimized warehouses are intended for memory-intensive Snowpark workloads, not for hosting containerized microservices. Permanent stored procedures also do not replace a network-accessible inference service.