Feature engineering and preprocessing · insight-1109

Replay logged arrival times to mask training features the way serving will see them

Situation
A feature comes from a feed that lands hours after the event; the warehouse history is complete, but a live scoring request often lacks the value.
Action
Using ingestion timestamps, give each training row only the latest value that had arrived by prediction time (missing if none); evaluate on this masked set next to the unmasked one.
Reason
What the model could see at prediction time, not when the event occurred, defines its inputs; the missingness pattern is part of the signal it will meet and must be learned.
Boundary
Late-arriving partner feeds, daily batch aggregates and corrected or backfilled records.

Symptom or risk

The model leans on a feature that is routinely absent or stale at prediction time, so live accuracy falls below any offline estimate.

Features guaranteed to arrive before scoring, or truly optional features.

Needs reliable arrival timestamps and reduces the effective signal in training.

Using the event time as the arrival time, or masking with a uniform random rate that matches nothing real.

Feast point-in-time joins can also filter on the created timestamp so backfilled or corrected values do not leak into training data.
Point-in-time joins (Feast documentation)