What CREATE MODEL costs in BigQuery

What it is

CREATE MODEL in BigQuery ML reads its training query like any other, but those bytes are not priced like a query. The model type sets the price, so the same SELECT can cost 50 times as much inside CREATE MODEL as it does on its own.

Why it costs money

• Linear and logistic regression, k-means, PCA, time series and contribution analysis cost $312.50 per TiB in the US, against $6.25 for a query. It is 50 times in every location.
• Auto-ARIMA fits candidate models and charges the bytes once for each. At the default AUTO_ARIMA_MAX_ORDER of 2, several series cost 12 candidates. One series costs 6 or 12, depending on the data.
• AutoML, DNN, wide-and-deep, boosted tree, random forest and autoencoder models bill the bytes at the query price, then add a training cost passed through from Google's AI platform. Bytes do not predict that part.
• Google counts evaluation, inspection and prediction in the free 1 TiB. It does not list model creation there, and each CREATE OR REPLACE MODEL bills again in full.

Worked example

CREATE MODEL `shop.churn`
OPTIONS (model_type = 'logistic_reg') AS
SELECT * FROM `shop.events`

CREATE MODEL `shop.churn`
OPTIONS (model_type = 'logistic_reg') AS
SELECT churned, visits, spend
FROM `shop.events` TABLESAMPLE SYSTEM (10 PERCENT)

How to fix it

1. 1Select only the columns the model uses. SELECT * pays the training price on every column.
2. 2Train on a partition range or a TABLESAMPLE, then widen it if the model needs more data.
3. 3For ARIMA, keep AUTO_ARIMA_MAX_ORDER low, or set AUTO_ARIMA to FALSE with NON_SEASONAL_ORDER on one series.
4. 4Retrain on a schedule that matches how fast the data changes.