Time & windows

Prompt A. "Explain why hourly GMV must use event time, not processing time. Then explain what a watermark is and how forBoundedOutOfOrderness(d) computes it. Explain what happens to an event that arrives after the watermark passes its window. Cover both cases: with a late side output and without one."

Prompt B. "A global hourly rollup stalls one quiet night, but most regions are busy. Explain the stuck-watermark trap (the minimum rule). Explain how withIdleness fixes it."

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

• Event time buckets by when the event happened. Processing time puts delayed events in the wrong hour and is not reproducible on replay.
• A watermark of T says that Flink expects no more events with timestamp ≤ T. It fires windows whose end ≤ T.
• forBoundedOutOfOrderness(d) emits (max event timestamp seen − d). A larger d gives more tolerance and later firing.
• Without late handling, an event past the watermark for its window is SILENTLY dropped.
• allowedLateness(d) keeps the window so that it can re-fire. sideOutputLateData(tag) routes too-late events to a stream (a dead-letter for late facts).
• The watermark of the operator is the MINIMUM across input partitions. So a quiet partition freezes it and stalls all windows.
• withIdleness(d) marks a quiet partition idle. Flink excludes it from the minimum, and windows fire again on the active partitions.