The user is asking about a sum over a rolling window, so a plain GROUP BY will double-count late-arriving rows. I should use the event timestamp, dedupe on the idempotency key first, then window the result.

Weighing two query plans against the size of the events table…