Handling late data in streaming windows

WHAT IT TESTS: whether you understand event time, watermarks, and the accuracy-latency tradeoff in streaming. ANSWER OUTLINE: aggregate on event time, not processing time, so windows reflect when events actually occurred; use watermarks to estimate progress and tolerate a configured amount of lateness, keeping window state open during a grace period so late events update results; finalize and emit when the watermark passes, and route events later than the allowed lateness to a side output or dead-letter sink rather than silently dropping.