Turning the Dial

Your job bounds out-of-orderness at 5 seconds and drops about 3% of events as late. You raise the bound to 2 minutes and change nothing else. What is the expected effect?

Events are now sorted by timestamp before the aggregation runs, so both result latency and completeness improve at the same time

Far fewer events are dropped, but each window now emits roughly two minutes after its end instead of five seconds, raising result latency

The same events are dropped, because the watermark decides only when a window fires and never which events that window is allowed to include

Throughput falls by roughly the factor the bound rose, since the operator now buffers 24 times as many records before it can emit any result