Snowflake Feature Store¶

The Snowflake Feature Store lets data scientists and ML engineers create, maintain, and use ML features in data science and ML workloads, all within Snowflake.

Generically, features are data elements used as inputs to a machine learning model. Many columns in a dataset, such as temperature or attendance, can be used as features as-is. In other cases, a column can be made more useful for training via preprocessing and transformation. For example, you might derive a day-of-week feature from a timestamp to allow the model to detect weekly patterns. Other common feature transformations involve aggregating, differentiating, or time-shifting data. Feature engineering is the process of deciding what features are needed by your models and defining how they will be derived from the raw data.

A feature store lets you standardize commonly used feature transformations in a central repository, enabling reuse, helping to reduce duplication of data and effort, and improving productivity. It also helps maintain features by updating them on new source data, always providing correct, consistent, and fresh features in a single source of truth. By cultivating consistency in how features are extracted from raw data, a feature store can also help to make your production ML pipelines more robust.

The Snowflake Feature Store is designed to make creating, storing, and managing features for data science and machine learning workloads easier and more efficient. Hosted natively inside Snowflake, the Snowflake Feature Store provides the following advantages:

• Your data remains secure, completely under your control and governance, and never leaves Snowflake.

• The Snowsight Feature Store UI makes it easy to search for and discover features.

• Access is managed with fine-grained role-based access control.

Key benefits of the Snowflake Feature Store include support for:

• Both batch and streaming data, with efficient automatic updates as new data arrives

• Backfill and point-in-time correct features with ASOF JOIN

• Feature transformations authored in Python or SQL

• Automatic update and refresh of feature values from source data with Snowflake managed Feature Views

• Ability to use user-managed feature pipelines with external tools such as dbt

The Snowflake Feature Store is fully integrated with the Snowflake Model Registry and other Snowflake ML features for end-to-end production ML.

The following illustration shows how the Snowflake Feature Store fits into a machine learning pipeline:

• Raw data can be obtained in batch from tables or views or from streaming data sources.

• The raw data is then transformed by features defined by data engineers, resulting in a feature table.

• The feature table can be used to generate training datasets used for training models in Snowpark ML, or to enrich test data used by the model to make predictions.