bigquery

Apply column pruning, predicate pushdown, early aggregation before joins, and common subexpression reuse. Mandatory rewrites include replacing WHERE col IN (SELECT ...) with WHERE EXISTS, and choosing VIEW versus TABLE materialization based on freshness, cost, and reuse.

Call AI.FORECAST as a table-valued function in the FROM clause with data_col and timestamp_col as required arguments. It uses the TimesFM model, supports id_cols for multiple series, horizon from 1 to 10000, and optional confidence_level and context_window settings.

BigFrames executes operations in BigQuery rather than local memory, so you should avoid to_pandas() and read_gbq() for SQL. Use bigframes.ml instead of scikit-learn, and prefer built-in accessors over UDFs or lambdas.

Yes, the input table or subquery must contain a column named or aliased as prompt, and the output_schema argument must be a quoted string of SQL column definitions. The function must be placed in the FROM clause as a table-valued function.

Remote models require a connection to Vertex AI, created with REMOTE WITH CONNECTION DEFAULT or a fully qualified connection ID. Verify the endpoint name is valid, such as gemini-2.5-flash or text-embedding-005, and that the dataset exists via CREATE SCHEMA IF NOT EXISTS.

Use APPROX_COUNT_DISTINCT when approximate results are acceptable, since it is faster and uses less memory. This change is conditional and should be proposed with user confirmation because it trades exactness for performance.