Revision history for fbsql

0.1.0 2026-07-09
      Initial release.

      Extension infrastructure:
      - PostgreSQL extension skeleton: control file (requires plr), PGXS
        Makefile, install script, fbsql schema, fbsql.version().
      - Pinned Docker development environment (PostgreSQL 16 + PL/R 8.4.8.6
        + R 4.2.2) with helper scripts for build, PL/R verification, and
        installcheck.
      - pg_regress suite (11 tests) run locally and in GitHub Actions CI;
        all numeric results verified against R's stats::glm() /
        predict.glm() to 4 decimals.

      fbsql.fit_glm(relation, formula, family):
      - Fits a GLM via PL/R (stats::glm) from a relation given as an SQL
        string and an R formula string; gaussian (identity) and binomial
        (logit) families.
      - Numeric and factor predictors (text columns follow glm()
        conventions: sorted levels, treatment contrasts).
      - Complete Case Analysis with explicit n_obs / n_used / n_dropped.
      - Returns a single relation: one row per term (estimate, std_error,
        statistic, p_value, Wald 95% CIs) plus model-level columns (family,
        link, formula, aic, deviance, null_deviance) and a jsonb metadata
        column (meta_version 1: response, term_labels, intercept,
        data_classes, xlevels, contrasts, coef_terms).
      - Clear fit_glm-prefixed errors for unsupported families, invalid
        formulas, missing columns, and empty relations.

      fbsql.predict_glm(relation, model, on_new_levels):
      - Scores a relation from a fit_glm() model relation alone -- computed
        in PL/pgSQL without R, from the coefficients and metadata.
      - gaussian/identity (linear predictor) and binomial/logit
        (probabilities, as R's predict(type = "response")).
      - Numeric and factor predictors (treatment-contrast dummies rebuilt
        from stored factor levels); NULL predictors yield NULL predictions.
      - Factor levels unseen at fit time: on_new_levels => 'error'
        (default) or 'na' (NULL prediction for those rows only).
      - Returns SETOF record: the input relation's rows plus a
        <response>_predicted column.

      Documentation and examples:
      - Running example (customer churn: fit on 2025 data, predict 2026)
        covered end to end by the regression tests.
      - Companion repository FbSQL-experiments with reproducible
        comparisons against MADlib, PostgresML, and Spark MLlib.