BigQuery ML

The step this removes

A conventional machine learning project on warehouse data begins with an export.

Data is queried out, written to files, moved to an environment with the right libraries, loaded into memory and only then trained on. Afterwards the model exists somewhere else, and applying it to warehouse data means moving data again.

That movement is frequently the slowest part of the project, and it is often the part that raises questions about where data is allowed to go.

BigQuery ML removes it. The model is trained where the data already sits.

Models created with SQL

A model is created with a statement. You name the model type and supply a query that returns the training data.

Prediction is another statement. It returns predictions alongside the input rows, so the result is a table like any other.

That has an immediate consequence: predictions can be joined to other warehouse tables in the same query. A prediction next to the customer record and their order history is one statement rather than an integration.

Who this opens machine learning to

The audience is people who already write SQL against the warehouse.

That is a large group, and conventionally they cannot train a model without learning a framework, a language and an environment. Each of those is a real obstacle, and together they are usually enough to prevent the attempt.

Expressing the model in SQL removes all three. An analyst who can write a query can train a model on data they already understand, which is frequently a bigger advantage than any technical property of the model.

What it is well suited to

The strongest use is establishing whether a signal exists.

Before committing to a machine learning project, the question is whether the data supports the prediction at all. Answering that conventionally takes an export, an environment and several days.

Answering it in SQL takes an afternoon. If the result is promising, a proper project is justified. If it is not, a great deal of effort has been saved.

The model types

The documented types cover common tasks: classification, regression, clustering, forecasting and others.

These are the shapes most business prediction problems take, and the documentation lists what is available along with the options for each.

Who uses it

BigQuery ML is used by analysts and data teams working in BigQuery, particularly where the data already lives there and the alternative involves exporting it.

Points to consider

It is specific to BigQuery, so it applies where that is your warehouse.

Model types are those provided. Work needing custom architecture or unusual preprocessing belongs in a dedicated framework, and the documentation is clear about what is covered.

Cost is warehouse compute. Training consumes it, prediction consumes it, and repeated training on large tables consumes a good deal of it. The pricing documentation is the reference, and understanding it before scheduling frequent retraining is sensible.

Getting started

The documentation includes an introduction and a first example that creates a model from a query and produces predictions, followed by a page per model type. Training one model on an existing table demonstrates the whole approach in a few statements.