Windowed operations in Kafka - Step-by-Step Execution

Process Flow - Windowed operations

Stream of events

↓

Assign event timestamps

↓

Define window size & type

↓

Group events into windows

↓

Aggregate or process events per window

↓

Emit windowed results

↓

Slide or advance window

↩Back to Group events into windows

Events flow in with timestamps, get grouped into time windows, processed, and results are emitted per window.

Execution Sample

Kafka

KStream<String, Integer> stream = builder.stream("input-topic");
KTable<Windowed<String>, Long> windowedCounts = stream
  .groupByKey()
  .windowedBy(TimeWindows.ofSizeWithNoGrace(Duration.ofSeconds(10)))
  .count();

This code counts events per key in 10-second tumbling windows.

Process Table

Step	Event Timestamp	Window Assigned	Key	Action	Windowed Count
1	00:00:03	[00:00:00 - 00:00:10)	A	Add event to window	A:1
2	00:00:07	[00:00:00 - 00:00:10)	A	Add event to window	A:2
3	00:00:12	[00:00:10 - 00:00:20)	A	Add event to new window	A:1
4	00:00:15	[00:00:10 - 00:00:20)	B	Add event to window	B:1
5	00:00:21	[00:00:20 - 00:00:30)	A	Add event to new window	A:1
6	00:00:25	[00:00:20 - 00:00:30)	A	Add event to window	A:2
7	Window [00:00:00 - 00:00:10) closes	-	-	Emit count for window	A:2
8	Window [00:00:10 - 00:00:20) closes	-	-	Emit count for window	A:1, B:1
9	Window [00:00:20 - 00:00:30) closes	-	-	Emit count for window	A:2

💡 All windows processed and counts emitted after their time range ends.

Status Tracker

Variable	Start	After 1	After 2	After 3	After 4	After 5	After 6	Final
Window [00:00:00 - 00:00:10) count for A	0	1	2	2	2	2	2	2
Window [00:00:10 - 00:00:20) count for A	0	0	0	1	1	1	1	1
Window [00:00:10 - 00:00:20) count for B	0	0	0	0	1	1	1	1
Window [00:00:20 - 00:00:30) count for A	0	0	0	0	0	1	2	2

Key Moments - 3 Insights

Why does the event at 00:00:12 start a new window instead of joining the first?

Because the first window covers [00:00:00 - 00:00:10), and 00:00:12 is outside that range, so it belongs to the next window [00:00:10 - 00:00:20). See execution_table rows 3 and 1.

Why are counts emitted only after the window closes?

Windowed operations emit results after the window time range ends to ensure all events in that window are counted. This is shown in execution_table rows 7-9.

What happens if events arrive late for a closed window?

In this example, no grace period is set, so late events are ignored or dropped. This is why windows close and emit counts strictly after their time range.

Visual Quiz - 3 Questions

Look at the execution_table at step 4, what is the window assigned to the event with key 'B'?

A [00:00:00 - 00:00:10)

B [00:00:20 - 00:00:30)

C [00:00:10 - 00:00:20)

D No window assigned yet

💡 Hint

Check the 'Window Assigned' column at step 4 in the execution_table.

At which step does the count for key 'A' in the first window reach 2?

A Step 1

B Step 2

C Step 3

D Step 6

💡 Hint

Look at the 'Windowed Count' column for key 'A' in execution_table rows 1 and 2.

If the window size was changed to 5 seconds, how would the window assigned at step 3 change?

A It would be [00:00:10 - 00:00:15)

B It would remain [00:00:10 - 00:00:20)

C It would be [00:00:05 - 00:00:10)

D No window would be assigned

💡 Hint

Window size defines the time range; smaller window means shorter intervals as seen in concept_flow.

Concept Snapshot

Windowed operations group streaming events by time intervals.
Define window size (e.g., 10 seconds) and type (tumbling, hopping).
Events are assigned to windows based on timestamps.
Aggregations (like count) happen per window.
Results emit after window closes.
Late events may be dropped or handled with grace period.

Full Transcript

Windowed operations in Kafka Streams process events grouped by time windows. Events flow in with timestamps and are assigned to windows based on these timestamps. For example, a 10-second tumbling window groups events from 0 to 10 seconds, then 10 to 20 seconds, and so on. Each event is added to its window's aggregation, such as counting occurrences per key. Counts are emitted only after the window time ends to ensure completeness. If events arrive late, they may be ignored if no grace period is set. This process repeats as new events arrive and windows slide forward.