Include a trained model in an app¶

Workflow - Add a model to an app¶

The following procedure outlines the typical workflow a provider follows create and add a Snowflake ML model to an app:

1. The provider develops a Snowflake ML model and logs it in the Snowflake Model Registry.

2. The provider exports the model artifacts from the Snowflake Model Registry and uploads them to a stage so that they are accessible to the application package.

3. The provider creates the model in the setup script of the app.

4. The app creates a model from these artifacts in the consumer account during installation or after an upgrade. Optionally, the app can grant access on the model to an application role.

5. The consumer uses the machine learning model if the provider configures the app to grant access to it.

Develop a machine learning model¶

Providers can develop new machine learning models or include existing models in an app.

• For information about developing models, see Snowflake ML Model Development.

• For information about managing models in a Snowflake Model Registry, see Snowflake Model Registry.

Export the model artifacts and upload to a stage¶

To include a model in an app, providers must export the model artifacts and upload them to a stage where they are accessible to the application package.

CREATE OR REPLACE PROCEDURE copy_model_artifacts_to_stage(src_registry_schema_fqn string, src_model string, src_model_version string, dst string)
  RETURNS STRING
  LANGUAGE python
  runtime_version = 3.11
  handler = 'copy_model_artifacts_to_stage'
  packages = ('snowflake-snowpark-python')
  execute as caller
as
$$

def copy_model_artifacts_to_stage(session, src_registry_schema_fqn, src_model, src_model_version, dst):
  session.use_schema(src_registry_schema_fqn)
  list_files = session.sql(f"list 'snow://model/{src_model}/versions/{src_model_version}/'")
  list_files.collect()
  for row in list_files.toLocalIterator():
     parts = row["name"].rsplit('/', 1)
     directory = parts[0]
     filename = parts[1]
     session.file.get(f"snow://model/{src_model}/{directory}/{filename}", f"/tmp/{directory}")
     session.file.put(f"/tmp/{directory}/{filename}", f"{dst}/{src_model}/{directory}", auto_compress=False, overwrite=True, source_compression="NONE")

  return f"Copied [snow://model/{src_model}/versions/{src_model_version}/*] to [{dst}/{src_model}/{directory}/]"
$$;

CALL copy_model_artifacts_to_stage('my_db.my_model_registry, 'my_model', 'V1', '@my_app_pkg.source_schema.source_stage/models');

Create the model objects in the consumer account¶

To create the model objects in the consumer account, the provider adds the necessary SQL commands to the setup script as shown in the following example:

CREATE APPLICATION ROLE IF NOT EXISTS app_user;

CREATE OR ALTER VERSIONED SCHEMA app_code;
GRANT USAGE ON SCHEMA app_code TO APPLICATION ROLE app_user;

CREATE OR REPLACE MODEL app_code.my_model FROM '/models/my_model/versions/V1;

Optionally, providers can grant access on the model to consumers by granting the USAGE privilege on the model to an application role:

GRANT USAGE ON MODEL app_code.my_model TO APPLICATION ROLE app_user;

Access the model within the app¶

To use the model internally as part of the app, providers add a SELECT statement to the setup script as shown in the following example:

SELECT app_code.my_model!predict(...);

Use the model as a consumer¶

If a provider grants privilege on the model to a consumer, the consumer can run the following command to access the model:

SELECT app_code.my_model!predict(...);

To run this command, consumers must use a role that has one of the following:

• The USAGE privilege granted on the model.

• The OWNERSHIP privilege on the application object.