What About Late Data?

Multi-Watermark Strategies and Result Retraction At the intermediate level, you learn that watermarks handle late data. The next level up, the interviewer tests whether you know that a single watermark is often insufficient. Real systems have multiple input streams with different lateness characteristics, and a single global watermark is bottlenecked by the slowest source. If you mention per-source watermarks unprompted, you've signaled serious depth. Result retraction is the advanced mechanism the interviewer may probe. Instead of dropping late events or holding windows open forever, the system emits a retraction of the previous result and then emits a corrected result. Downstream consumers see: (-1, old_value), (+1, new_value). Your answer should name retraction as the pattern and immedi