​
Configuration

First we have to add a declarative Snowflake configuration in Python.

​
Credentials with Account and Organization

snowflake_config.py
import featureform as ff

ff.register_snowflake(
    name = "snowflake_docs",
    description = "Example offline store",
    username = snowflake_username,
    password = snowflake_password,
    account = snowflake_account,
    organization = snowflake_org,
    database = snowflake_database,
    schema = snowflake_schema,
    warehouse = snowflake_wh,
    catalog = ff.SnowflakeCatalog(
        external_volume = snowflake_external_volume,
        base_location = snowflake_base_location,
        table_config=ff.SnowflakeDynamicTableConfig(
            target_lag='1 days',
            refresh_mode=ff.RefreshMode.INCREMENTAL,
            initialize=ff.Initialize.ON_CREATE,
        ),
    )
)

​
Legacy Credentials

Older Snowflake accounts may have credentials that use an Account Locator rather than an account and organization to connect. Featureform provides a separate registration function to support these credentials.

snowflake_config.py
import featureform as ff

ff.register_snowflake_legacy(
    name = "snowflake_docs",
    description = "Example offline store",
    username = snowflake_username,
    password =  snowflake_password,
    account_locator = snowflake_account_locator,
    database = snowflake_database,
    schema = snowflake_schema,
    warehouse = snowflake_wh,
    catalog = ff.SnowflakeCatalog(
        external_volume = snowflake_external_volume,
        base_location = snowflake_base_location,
        table_config=ff.SnowflakeDynamicTableConfig(
            target_lag='1 days',
            refresh_mode=ff.RefreshMode.INCREMENTAL,
            initialize=ff.Initialize.ON_CREATE,
        ),
    )
)

​
Dynamic Apache Iceberg™ Tables

Featureform’s Snowflake integration leverages dynamic Apache Iceberg™ tables to offer automated data transformation pipelines from SQL transformations to training sets.

This requires users to create an external volume prior to registering Snowflake as a provider. See the repo terraforming-snowflake-external-volumes for an example of creating an external volume backed by an AWS S3 bucket using Terraform.

Additionally, users may have to manually enable change tracking on source tables for incremental refreshes to work.

The name of the external volume, as well as a folder name in the bucket (i.e. base_location), is required to instantiate ff.SnowflakeCatalog. Currently, the only option for catalog is to use Snowflake as the Iceberg catalog.

Additionally, users can set “provider-level” values for target lag and refresh mode that will be used by all resources (i.e. SQL transformations, features, labels and training sets) unless otherwise specified at the resource level.

​
Session Parameters

Users may optionally add session parameters when configuring Snowflake as an offline store. The parameters will be used when Featureform creates a connection with Snowflake.

For example, providing a value for query_tag at configuration ensures all queries issues by Featureform will have the value of query_tag applied to them, which will be visible in the Snowflake UI.

​
Applying

Once our config file is complete, we can apply it to our Featureform deployment

featureform apply snowflake_config.py --host $FEATUREFORM_HOST

We can re-verify that the provider is created by checking the Providers tab of the Feature Registry.

​
Mutable Configuration Fields

Many of the values provided at the time of registration can be changed if users want to re-apply the same Python file (e.g. snowflake_config.py) again with different values. The following list of configurations fields are mutable:

  • description

  • username

  • password

  • role

  • warehouse

  • database

  • schema

  • session_params

​
Implementation

​
Datasets

​
Primary Datasets

Primary datasets (i.e. tables) are read-only and are only used as inputs in transformations, features and/or labels.

​
Transformation Sources

SQL transformations are used to create a dynamic Apache Iceberg™ table. The refresh mode (i.e. full or incremental) depends on the user-defined SQL query, so users should be aware of query expressions, keywords, etc. that might prevent incremental refreshes.

​
Offline to Inference Store Materialization

When a feature is registered, Featureform creates an internal transformation to get the newest value of every feature and its associated entity. A Kubernetes job is then kicked off to sync this up with the inference store.

​
Training Set Generation

Training set are created from one or more features and a label. Label are registered separately as a dynamic table to ensure the latest label values are used in the training set. When users define a timestamp column for a feature and label combination, Featureform uses Snowflake’s ASOF JOIN to combine feature values with their labels.