API Reference

This reference is generated directly from the source docstrings via mkdocstrings, so it always matches the code.

flowchart TD
    subgraph Entry["Entrypoints & SDK"]
        SDK["Forecaster & Registry (sdk)"]
        REV["monitor_run & review_run (review)"]
        PG["Single-Series Playground (playground)"]
        MAIN["run & CLI (main)"]
    end
    subgraph Plan["Planning, Sizing & Preflight"]
        CFG["RunConfig & Settings (config, settings)"]
        DAG["DAG & Router (dag, router, launch_plan)"]
        PROF["Profiling, Resources, Quota & Capacity"]
    end
    subgraph Core["Forecasting Core & Engines"]
        WRK["run_cell (worker)"]
        MOD["Models, Features, Seasonality & HPO"]
        BT["Backtest, Metrics & Calibration"]
        ENS["Ensembler & Ensemble Driver"]
        ENG["Spark, Ray, Vertex CustomJob, GCE, GKE, BigQuery & Vertex AutoML Engines"]
    end
    subgraph Ops["Registry, Probes & Repair"]
        REG["Registry Views & Ops (registry.views, registry.ops)"]
        PRB["Probes, Surgical Repair & Ray Reaper"]
    end
    Entry --> Plan --> Core --> Ops

scale_forecasting exposes three doors onto the same forecasting core, plus a full operational and registry management surface:

• The easy path — Forecaster, review, playground, agent_surfaces, and mcp: point at a config, call dry_run() / run() / review(), monitor live progress with monitor_run(), inspect single-series fits in a notebook, probe local environment readiness (probe_environment), or serve the built-in Model Context Protocol server over stdio.
• The orchestration & planning path — run, dag, router, launch_plan, vertex_submit, gce_submit, gke_submit, automl_submit, and staging: plan the family-to-job DAG, run quota preflight, stage artifacts to GCS, and launch across Local, Dataproc Serverless, Dataproc Cluster, Vertex AI Ray, Vertex AI CustomJob, Compute Engine (gce), Google Kubernetes Engine (gke), BigQuery, and Vertex AI AutoML / Tabular Workflows (vertex_automl).
• The direct path — the cell primitives (run_cell and the group/chunk runners in Spark core, Ray core, Vertex AI CustomJob, GCE & GKE engine, BigQuery-native engine, and Vertex AI AutoML & Tabular Workflows engine) for embedding the model machinery in your own pipelines.
• The manage & repair path — registry.ops, registry.views, probes, retry_run, and ray_reaper: inspect registry health, reconcile live platform state, surgically repair errored cells, reap orphaned clusters, or snapshot/export runs.

Supporting surfaces: configuration, settings, errors, the model factory and BaseModel contract, features, seasonality, HPO, ensembler and ensemble driver, hierarchical reconciliation, metrics and BaseMetric, backtesting, calibration, compute profiling, resource planning, quota preflight, capacity fallback, and the synthetic data generator.

The package front door lazy-loads the heavy names via __getattr__, so import scale_forecasting stays fast; each page below documents the concrete module a name resolves to.