http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/EventTimeSessionWindow-firing-too-soon-tp35959p35990.html
Hi, thanks for answering.
> I guess you consume from Kafka from the earliest offset, so you consume historical data and Flink is catching-up.
Yes, it's what's happening. But Kafka is partitioned on sessionId, so skew between partitions cannot explain it.
I think the only way it can happen is when when suddenly there's one event with very late timestamp
> Just to verify, if you do keyBy sessionId, do you check the gaps of events from the same session?
Good point. sessionId is unique in this case, and even if it's not - every single session suffers from this problem of early triggering so it's very unlikely that all millions sessions within that hour had duplicates.
I'm suspecting that the fact I have two ProcessWindowFunctions one after the other somehow causes this.
I deployed a version with one window function which just prints the timestamps to S3 (to find out if I have event-time jumps) and suddenly it doesn't trigger early (I'm running for 10 minutes and not a single event has arrived to the sink)