My Flink job is doing aggregations on top of event-time based windowing
across Kafka partitions. As I have been developing and restarting it, the
state for the catch-up periods becomes unreliable -- lots of duplicate emits
for time windows already seen before, that I have to discard since my sink
can't handle it. There may be a bug in my job, but I wanted to clarify
whether this might be a flaw in Flink's handling of this.

I understand there is m:n mapping of partitions to sources depending on the
parallelism. Each source will have its own watermark. During catchup,
watermark progression can become pretty fragile, e.g. in my case where
there's n partitions and parallelism is 1.

I feel like some kind of event time alignment is needed across partitions. I
may be completely off here, so I look forward to your thoughts on this!



--
View this message in context: 
http://apache-flink-user-mailing-list-archive.2336050.n4.nabble.com/Kafka-partition-alignment-for-event-time-tp4782.html
Sent from the Apache Flink User Mailing List archive. mailing list archive at 
Nabble.com.

Reply via email to