Compare a challenger model's holdout metric vs the incumbent's and recommend
promote (challenger wins by ≥ margin) or hold. Pure / synchronous — reads
the frozen HoldoutMetrics JSON off both models. This is a RECOMMENDATION; no
promotion is performed here (that is a separate signed-off Action).
the current model
the freshly-retrained model
the resolved policy (margin + comparison metric)
Decide whether a binding's model is stale per policy. Pure with respect to the injected seams: reads counts/drift via deps and the clock, mutates nothing.
the MJ: ML Model Scoring Bindings row
the MJ: ML Models row the binding scores with
the staleness/retraining policy (partial → merged with defaults)
the injected dependency bundle
Re-score a Mode='Scheduled' binding and stamp its monitoring fields
(LastScoredAt = now, LastRowCount = scoredCount). Bindings that are not
Scheduled are skipped (returns null).
the binding to re-score
the model it scores with
the injected dependency bundle
the re-score result, or null when the binding is not Scheduled
Run the full maintenance pass over a set of bindings — the entry the scheduler
(or an Action) calls. For each binding: load its model, detect staleness,
optionally re-score (Scheduled bindings), and optionally retrain+compare when
stale. Each binding is isolated: a failure is captured on its entry and the
pass continues.
the bindings to maintain (resolve via the loader for "all scheduled")
the staleness/retraining policy (partial → merged with defaults)
the injected dependency bundle
pass toggles (re-score scheduled / retrain stale)
Retrain when stale: if staleness.stale, retrain the binding's model against
the SAME pipeline → a new immutable version, then compare the challenger's
holdout metric vs the incumbent's and return a promotion RECOMMENDATION
(promote / hold). Never auto-promotes. A retrain whose gate decision ships no model has no challenger: the
outcome carries the run without a comparison, the incumbent stays, and the stored reason is logged.
the binding whose model may be retrained
the current (stale) model
the staleness verdict driving the decision
the policy (for the promotion margin + comparison metric)
the injected dependency bundle (carries the training engine + deps)
OptionaltrainInputOverrides: Partial<Omit<TrainModelInput, "pipelineId">>optional extra train inputs (labelEventDates / maxRows / etc.)
Maintenance orchestrator. Stateless across calls; construct once and reuse. The dependency bundle (MaintenanceDeps) and RetrainingPolicy are supplied per call.