jfyne / dbt_ml

Package for dbt that allows users to train, audit and use BigQuery ML models.

BigQuery ML models in dbt

Package for dbt that allows users to train, audit and use BigQuery ML models. The package implements a model materialization that trains a BigQuery ML model from a select statement and a set of parameters. In addition to the model materialization a set of helper macros that assist with model audit and prediction are included.

Installation

To install the package add the package path to the packages.yml file in your dbt project

In order to use the model audit post-hook the following variables have to be set in your dbt_project.yml file.

Variable	Description
dbt_ml:audit_schema	Schema of the audit table.
dbt_ml:audit_table	Name of the audit table.

You will also need to specify the post-hook in your dbt_project.yml file^[1] as {{ dbt_ml.model_audit() }}. Optionally, you can use the dbt_ml.create_model_audit_table() macro to create the audit table automatically if it does not exist - for example in an on-run-start hook.

Example config for dbt_project.yml below:

vars:
  "dbt_ml:audit_schema": "audit"
  "dbt_ml:audit_table": "ml_models"
on-run-start:
  - '{% do adapter.create_schema(api.Relation.create(target.project, "audit")) %}'
  - "{{ dbt_ml.create_model_audit_table() }}"
models:
  <project>:
    ml:
      enabled: true
      schema: ml
      materialized: model
      post-hook: "{{ dbt_ml.model_audit() }}"

Usage

In order to use the model materialization, simply create a .sql file with a select statement and set the materialization to model. Additionaly, specify any BigQuery ML options in the ml_config key of the config dictionary.

# model.sql

{{
    config(
        materialized='model',
        ml_config={
            'model_type': 'logistic_reg',
            'early_stop': true,
            'ls_init_learn_rate': 0.1,
            ...
        }
    )
}}

select * from your_input

Note that the materialization should not be prefixed with dbt_ml, since dbt does not support namespaced materializations.

After training your model you can reference it in downstream dbt models using the included predict macro.

# downstream_model.sql

{{
    config(
        materialized='table'
    )
}}

with eval_data as (
    ...
)

select * from {{ dbt_ml.predict(ref('model'), 'eval_data') }}

If you're using a BQML matrix_factorization model, you can use the recommend macro in the same way.

# downstream_model.sql

with predict_features AS (
    ...
)

select * from {{ dbt_ml.recommend(ref('model'), 'predict_features') }}

The ML.DETECT_ANOMALIES function provides anomaly detection for BigQuery ML.

# detect_anomalies_model.sql

{{
    config(
        materialized='table'
    )
}}

with eval_data as (
    ...
)

select * from {{ dbt_ml.detect_anomalies(ref('model'), 'eval_data', threshold) }}

Tuning hyperparameters

BigQuery ML supports tuning model hyperparameters^[2], as does dbt_ml. In order to specify which hyperparameters to tune, and which parameterspace to use, one can use the dbt_ml.hparam_candidates and dbt_ml.hparam_range macros that map to the corresponding BigQuery ML methods.

The following example takes advantage of hyperparameter tuning:

{{
    config(
        materialized='model',
        ml_config={
            'model_type': 'dnn_classifier',
            'auto_class_weights': true,
            'learn_rate': dbt_ml.hparam_range(0.01, 0.1),
            'early_stop': false,
            'max_iterations': 50,
            'num_trials': 4,
            'optimizer': dbt_ml.hparam_candidates(['adam', 'sgd'])
        }
    )
}}

It is worth noting that one must set the num_trials parameter to a positive integer, otherwise BigQuery will return an error.

Overriding the package

If a user wishes to override/shim this package, instead of defining a var named dbt_ml_dispatch_list, they should now define a config in dbt_project.yml, for instance:

dispatch:
  - macro_namespace: dbt_ml
    search_order: ['my_project', 'dbt_ml']  # enable override

Reservations

Some BigQuery ML models, e.g. Matrix Factorization, cannot be run using the on-demand pricing model. In order to train such models, please set up a flex or regular reservation^[3] prior to running the model.

Footnotes

[1] The post-hook has to be specified in the dbt_project.yml instead of the actual model file because the relation is not available during parsing hence variables like {{ this }} are not properly templated.

[2] https://cloud.google.com/bigquery-ml/docs/reference/standard-sql/bigqueryml-hyperparameter-tuning

[3] https://cloud.google.com/bigquery/docs/reservations-tasks

References

• BigQuery ML Syntax and Options
• BigQuery ML Pricing