Joins — regular, interval, temporal (point-in-time) & lookup

Prompt A. "Explain the difference between a regular join, an interval join and a temporal (versioned) join on streams. Take an orders × FX-rate enrichment. The business wants the rate that was in effect on the order date. Say which join is correct. Say why the other two give wrong or unbounded results."

Prompt B. "You implement the as-of enrichment by hand with a KeyedCoProcessFunction. Explain what you store in state and when you emit a result. Explain why you must wait for the watermark to pass the timestamp of the order before you pick a version. Connect this to how an SCD2 as-of join works in your warehouse."

Rubric — a strong explanation hits these (5–7):

• Regular join keeps both sides in state (unbounded) and emits retractions.
• Interval join bounds state to a time window around the match.
• Temporal/versioned join picks the dimension version effective at the event time of the fact (as-of).
• Names the as-of rule: greatest effective_from <= fact event time.
• Explains the watermark dependency: emit only after the watermark passes the fact time, so every eligible version has arrived.
• Maps a dimension version to an SCD2 row (valid_from/valid_to) and the join to an as-of join.
• Explains why a lookup join (reads "now") is wrong for reprocessing / back-dated corrections.