LoginSign Up
← Back to all sparks
O

orbital

ANALYTICS
Velocity0.0

Turn Tidymodels Workflows Into Series of Equations

github.com/tidymodels/orbital ↗

Turning fitted tidymodels into SQL, one model family at a time — and the boosting engines just landed.

tidymodelsin-database-scoringsql-generationmodel-deploymentr-package
◆Current state
orbital converts a fitted tidymodels workflow into a database expression so prediction runs where the data lives, no R session in the loop. Its value is entirely a function of coverage, and 0.5.0 was the largest coverage release yet: catboost and lightgbm boosted trees, rpart decision trees, earth-backed MARS, glmnet multinomial regression, and both randomForest and ranger random forests, all for numeric, class, and probability predictions. The 0.5.1 follow-up is corrective, fixing SQL that Snowflake and other engines rejected because it cast booleans directly to numeric.
◆Where it's heading
The package has been working outward in rings: recipe preprocessing steps first, then model types, then post-processing via the tailor package in 0.4.0, with show_query() added so users can inspect what actually gets sent. Recent releases show the constraint shifting from R-side translation to SQL dialect compatibility — the bugs now are about what a specific database will accept, not whether a model can be expressed. estimate_orbital_size() in 0.5.1 acknowledges the other practical limit, since generated expressions can grow large enough to matter before you generate them.
◆Prediction
With the major boosting and ensemble engines covered, expect the next releases to keep chasing dialect-specific SQL correctness across warehouses rather than adding model families.

◆Recent moves

  1. 1mo ago

    Snowflake-compatible SQL for dummy and NA steps

    step_dummy() and step_indicate_na() now emit SQL that works on Snowflake and other databases that refuse a direct boolean-to-numeric cast. Adds estimate_orbital_size() to predict an expression's character count without generating it, which matters because generated SQL for tree ensembles can get unwieldy.

    View source ↗
  2. 5mo ago

    catboost, lightgbm, ranger and four more model families translate to SQL

    ⚡ SPARK

    The coverage release that brings the models teams actually deploy into scope: catboost and lightgbm boosted trees, rpart decision trees, earth MARS, glmnet multinomial regression, and randomForest and ranger forests, across numeric, class, and probability predictions. A separate_trees argument gives tree ensembles a way to manage the size of the generated expression.

    View source ↗
  3. 8mo ago

    Compatibility with new xgboost versions

    A single-line release restoring compatibility with newer xgboost releases. Routine upstream tracking of the kind a translation layer accumulates.

    View source ↗
  4. 8mo ago

    Post-processing adjustments from tailor become translatable

    Extends translation past the model itself to the tailor package's post-processing — equivocal zones, numeric range clamping, probability thresholds, and custom prediction adjustments — plus their workflow integration. A show_query() method lets users see the emitted SQL directly.

    View source ↗
  5. 11mo ago

    PCA step translation bugs cleared

    Four related fixes to PCA steps: failures above 99 trained predictors, zero-valued terms emitted by step_pca_sparse(), errors depending on predictor selection, and large PCA results failing against databases. Narrow corrections within an already-supported step family.

    View source ↗
  6. 1y ago

    Class and probability predictions arrive, with glm and xgboost

    Adds a type argument for prediction type and extends support to logistic_reg with glm, boost_tree with xgboost, and decision_tree with partykit for class and probability predictions. The first release where orbital handles classification output rather than numeric prediction alone, plus an augment() method and a prefix argument for renaming prediction columns.

    View source ↗