Processing-Time Windowing Artifacts (Figure 12-8)
a 1-minute deploy pauses the stream processor — event-time windows stay honest, processing-time windows invent a spike
① event time — the true timeline
deploy / restart
processor down ≈ 1 min
each event keeps its event-time timestamp
0
1
2
3
4
5
6
7
8
9
10
event time (minutes) →
backlog flushed into one window
① processing time — grouped by the processor's clock
event-time counts (steady)
processing-time counts
3
3
3
3
3
3
3
3
down
6
spurious spike — artifact of
processing-time windowing
0
1
2
3
4
5
6
7
8
9
processing-time windows (1 min each)
processing time →
Use event timestamps
+ watermarks, not the
processor's clock
The same applies to
batch jobs: a batch's
output must depend on
the input data, not on
when the job happened.