Load baseline and rolling evaluation feature distributions, confidence distributions, and
champion-vs-challenger pairs from pipeline-generated MJ: Feature Values (Source = 'Pipeline')
for a given RecordProcessID or MJ: ML ModelsModelID (MLModelID).
Important semantics:
Champion rows are upserted in place by ValueKeyHash (EntityID + RecordID + FeatureName), so each
window holds values computed (i.e. changed) in that window rather than a full historical population
snapshot.
Each challenger row computed in the evaluation window is paired with the latest champion row for the
same record and feature (the latest champion value is all that exists). A champion that has not changed keeps
its old ComputedAt, so champions older than the windows are read for pairing only (one batched read).
An ML Model Record Process with no challenger configured writes no Feature Values, so it has no drift
population and the sweep reports InsufficientData: true for it.
MJ: Feature Values rows carry no latency metrics, so the scheduled sweep reports no P95 latency
ratio (LatencyRatio is evaluated only when callers pass inline latencies).
Load baseline and rolling evaluation feature distributions, confidence distributions, and champion-vs-challenger pairs from pipeline-generated
MJ: Feature Values(Source = 'Pipeline') for a givenRecordProcessIDorMJ: ML ModelsModelID(MLModelID).Important semantics:
ValueKeyHash(EntityID + RecordID + FeatureName), so each window holds values computed (i.e. changed) in that window rather than a full historical population snapshot.ComputedAt, so champions older than the windows are read for pairing only (one batched read).InsufficientData: truefor it.MJ: Feature Valuesrows carry no latency metrics, so the scheduled sweep reports no P95 latency ratio (LatencyRatiois evaluated only when callers pass inline latencies).