Notesstream🌱 seedling
Windowing
Also: Window · Tumbling Window · Hopping Window · Sliding Window · Session Window · Cumulating Window · Global Window
Windows slice an unbounded stream into finite, per-key groups by event time, so aggregates and joins can finish. The window shape (tumbling, hopping, session) decides how many results you get, how much state you hold, and when a result can be emitted.
Why it exists
An aggregate over an infinite stream never completes. Windows bound both the computation and the state, and give each result a time it belongs to.
How it works
- Tumbling: fixed size, no overlap. Each event is in exactly one window (per-minute metrics, billing).
- Hopping (sliding): fixed size, advancing by a smaller slide. Each event is in size/slide windows (moving averages).
- Session: per key, closed after an inactivity gap. A late event can bridge two sessions and merge them.
- Cumulating: grows within a period ("so far today, every minute"). Global: one window per key, which needs a custom trigger.
- A window's state lives until the Watermark passes its end, plus any allowed lateness.
Trade-offs
| Window | Gains | Costs |
|---|---|---|
| tumbling | simple, one result per period | bursts split across boundaries |
| hopping | smooth trends | size/slide × state and output |
| session | matches user behaviour | data-dependent boundaries, merges, unbounded if keys never go quiet |
In practice
- Flink: DataStream window assigners, plus the SQL window TVFs TUMBLE, HOP, CUMULATE and SESSION (SESSION since 1.19).
- Kafka Streams: tumbling, hopping, sliding and session windows. Since 3.0 you choose the grace period explicitly.
Pitfalls
- Tiny slides on hopping windows multiply state and output.
- Session windows on keys that never go quiet (bots, health checks) never close.
- Windowing in processing time by accident.
Connections
- Part of
- Stream Processing
- Builds on this
- Stream Join
- In real systems
- Apache Flink
Recall
These cards are also in your review deck, scheduled with spaced repetition.