Skip to content

Window Computation - Middle

How do window assigners and watermarks produce finite event-time results?

Window Shape Typical use
Tumbling fixed, non-overlapping revenue per minute
Sliding fixed size, overlapping slides rolling 15-minute rate
Session gap-delimited per key user activity sessions
Global one logical window plus triggers custom incremental logic 
events
    .assignTimestampsAndWatermarks(
        WatermarkStrategy.<Event>forBoundedOutOfOrderness(Duration.ofMinutes(2))
            .withTimestampAssigner((event, previous) -> event.eventTime()))
    .keyBy(Event::accountId)
    .window(TumblingEventTimeWindows.of(Duration.ofMinutes(5)))
    .aggregate(new RevenueAggregate());

A watermark of 10:10 means the engine believes future events at or before 10:10 are unlikely. It is a progress estimate, not a guarantee. A five-minute window ending at 10:10 can now fire according to its trigger.

flowchart LR E1[Event 10:07] --> W[Window 10:05-10:10] E2[Event 10:06 arrives later] --> W WM[Watermark passes 10:10] --> F[Fire result]

Session windows may merge when a late event bridges two sessions, so their state and output behavior are more complex than fixed windows.

Test yourself

  1. When would a sliding window contain one event multiple times?
  2. What claim does a watermark make?
  3. Why can a late event merge two session windows?

Continue to senior.md.