Model Inference in Snowflake¶

Snowflake uses two distinct compute engines:

• The warehouse (SQL Engine)

• Snowpark Container Service

The Snowflake Model Registry provides a unified interface to both engines. The optimal environment for your use cases depends on your latency, data type, and scaling requirements. Snowflake offers the following approaches to your inference workflows:

Real-time Inference (REST API): Designed for low-latency and real-time use cases. Requests are facilitated via HTTP endpoints and are ideal for powering external applications.

Snowflake Native Batch Inference (SQL): Designed for batch workloads that require integration with the Snowflake SQL ecosystem. For example, batch workloads can integrate with Dynamic Tables, Snowpark, DBT, and User Tasks. You can use a SQL function, you can embed intelligence directly into your existing data pipelines without moving data or managing external infrastructure.

Job-based Batch Inference: This approach is designed for high-throughput, distributed processing where inference is treated as a standalone compute stage. By decoupling inference from the SQL engine, you can optimize for both price and performance. You can use Batch Inference to help you handle massive datasets or navigate complex computational requirements. This is ideal for processing files—such as images, video, and audio—directly from Snowflake Stages.