Point-in-Time Join

A point-in-time join selects related data for an entity at a specified time. For a type 2 dimension, this commonly means matching the entity and the interval containing the fact’s timestamp. For feature observations, an as-of lookup may instead choose the latest eligible observation no later than prediction time. State which interpretation the query implements.

For timestamp ranges, start ≤ t < end avoids subtracting an arbitrary fraction of a second. Define timezone, tie handling, and missing-history behavior. Check matches per fact key: equal total row counts can hide one dropped row and one duplicated row. Aggregating before checking can conceal that error.

Historical validity does not prove feature availability. A move effective March 14 but received March 20 cannot inform a March 16 prediction. Preserve the version actually available in the prediction system, respect freshness limits, and construct labels from a separate future outcome window. A correct temporal join is one part of leakage prevention, not a complete guarantee.

Further reading

• Feast point-in-time joins

See the worked examples in Data Modeling for Operational and Analytical Systems.