Point-in-Time Feature Join

This is the join a feature store exists to get right, and doing it wrong is the most expensive mistake in applied ML: because the model looks better, not worse.

You have two frames:

• events: columns entity_id, timestamp: the moments you need features for
• features: columns entity_id, timestamp, value: a feature that changes over time

For each event, attach the value of the feature as it was known at that moment: the most recent feature row for that entity with a timestamp less than or equal to the event's. If no such row exists, the value is NaN.

Return entity_id, timestamp, value, sorted by entity_id then timestamp.

The trap: a plain merge on entity_id attaches every feature value, including ones recorded after the event. Training on that produces a model with excellent offline metrics that fails on day one in production, because at serving time the future does not exist.