PREDICTIVE MODELS

Ship a model without leaving the database.

Train on a table, read how trustworthy the result is, promote a version and write scores back to a table your dashboards already query.

STEP 1

Train on a table you own

One call trains a model on the rows you select. The model, its metrics and its training data reference are recorded together.

No export, no separate ML platform

psql

-- train a model on a table you own
SELECT status, metrics
FROM openpgray.train('churn_model',
  'SELECT tenure, plan, usage, label FROM core.customers',
  '{}', 'LogisticRegression');

osdbstore

Predictive models: the model health tab1400 x 1000 pxDelete this block and drop in an Image or Video widget

STEP 2

Read the health report

Metrics come with plain-language checks: ROC AUC, calibration, and warnings about features that leak the answer.

Leakage warnings before you promote
Compare versions side by side

STEP 3

Promote and score

Promote the version you trust, then score new rows in batch. Predictions land in a table with the probability next to each row.

Registry with promote and roll back

psql

-- score new rows
SELECT row_id, prediction, probability
FROM openpgray.predict('churn_model',
  'SELECT tenure, plan, usage FROM core.customers_new ORDER BY id');

osdbstore

Predictive models: drift check and retrain history1400 x 1000 pxDelete this block and drop in an Image or Video widget

STEP 4

Retrain when the data drifts

A drift check watches incoming data. When it moves enough, retraining runs on its own and the new version waits for your approval.

Drift-gated retraining
Every run is recorded

Train your first model.

Start with the sample data, then point it at your own table.