Introduced on day 2
Point-in-time correctness
Building each training row only from values as they existed at that row's cutoff, not as they look today after updates.
In detail
Requires event and load timestamps, snapshot tables or versioned records. Feature stores implement it with point-in-time joins that look backward from each row's timestamp.