developing-with-bigquery

Apply column pruning, predicate pushdown, early aggregation before joins, and common subexpression reuse. Rewrite IN subqueries as EXISTS, and choose TABLE over VIEW for large or reused intermediate results. Conditionally consider UNION ALL and APPROX_COUNT_DISTINCT for further gains.

Call AI.FORECAST as a table-valued function in the FROM clause with your input table, data_col, and timestamp_col as required arguments. Use named arguments with the => operator for optional parameters like horizon, id_cols, and confidence_level. It uses TimesFM models without requiring training.

BigFrames should be the default choice because it is more efficient than the BigQuery Python client library. Avoid read_gbq() for SQL execution and use BigFrames DataFrame methods instead, keeping computation in the cloud rather than downloading data with to_pandas().

No, you should import ML tools from bigframes.ml instead of scikit-learn. BigFrames ML handles scaling automatically, its predict() returns a DataFrame with predictions and features, and models do not accept a random_state argument.

AI.GENERATE_TABLE requires the input subquery to contain a column aliased as prompt, and the output_schema string must be enclosed in quotes. The function must also be placed in the FROM clause as a table-valued function with a remote Gemini model.

Use CREATE MODEL with REMOTE WITH CONNECTION DEFAULT and set the ENDPOINT option to a model like gemini-2.5-flash or text-embedding-005. The DEFAULT connection is the recommended behavior; otherwise specify a fully qualified connection ID.