ADR 0012: BigQuery ML explicitly out of scope for v1

• Status: Accepted (partially superseded by ADR 0047)
• Superseded by: 0047 (partial). The decision below is the v1.0.0 historical record. ADR 0047 redefines the BigQuery ML scope: model-metadata registration via CREATE MODEL, the Models REST resource (list/get/patch/delete), and the ML.PREDICT output shape move into scope, while training, evaluation, forecasting, generation, and prediction accuracy remain out of scope. See ADR 0047 for the current Models-resource scope, including the correction that BigQuery has no Models insert method.

Context

BigQuery ML (CREATE MODEL, ML.PREDICT, ML.EVALUATE, ML.FORECAST, clustering, ARIMA, linear/logistic regression, DNN, boosted trees, matrix factorization, AutoML) is a large surface area inside BigQuery.

Decision

BQML is explicitly out of scope for v1.0.0. CREATE MODEL and ML.* function invocations return a clear UnsupportedFeatureError with pointer to this ADR and docs/reference/out-of-scope.md.

Only Models resource CRUD (list/get/insert/patch/update/delete of model metadata) is supported — users can register externally-trained models as metadata and test their workflows around that metadata.

Rationale

• Emulating correct BQML training/inference would require shipping ML runtimes (scikit-learn, TensorFlow, statsmodels, etc.) and reproducing BigQuery's model semantics to a degree that approximates production-correct predictions. This is comparable in effort to the rest of the emulator combined.
• The most common testing needs for BQML users — workflow orchestration (DAGs, scheduled queries, service-account behavior) — are covered by testing against real BigQuery in a dedicated test project with the sandbox tier.

Consequences

• Positive: keeps v1 scope tractable and quality bar achievable.
• Positive: clear error messages prevent silent wrong behavior.
• Negative: users whose primary workload is BQML cannot test locally. Documented; reconsiderable for v2 as a separate product decision.