Re: increasing parallelism increases the end2end latency in flink sql

Posted by Yan Zhou [FDS Science] ­ on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/increasing-parallelism-increases-the-end2end-latency-in-flink-sql-tp20296p20298.html

The BoundedOutOfOrdernessTimestampExtractor  is assigned to datastream after kafka consumer. The graph is like:

KafkaSource-> map2Pojo -> BoundedOutOfOrdernessTimestampExtractor -> Table -> ......

---

From: Yan Zhou [FDS Science] <[hidden email]>
Sent: Wednesday, May 23, 2018 3:21:24 PM
To: [hidden email]
Subject: increasing parallelism increases the end2end latency in flink sql

 

Hi,

My application assigned timestamp to kafka event with BoundedOutOfOrdernessTimestampExtractor then converted them to a table. Finally flink SQL over-window aggregation is run against the table. 

When I double the parallelism of my flink application, the end2end latency is doubled.  What could be the cause? It seems to me that it's because of slower advance of watermark in operator of operators generated by sql.

In this email thread [1], it's said that flink sql remove the internal DataStream timestamp and move it into the record. Does the query ignore the internal DataStream watermarks and re-generate them from the record? Let say there are two operator instances for one task, do they have same watermark?

There is a similar issue that i can find in the email thread [2] .

Best

Yan

[1]: https://lists.apache.org/thread.html/c5182628272f018037ce832290f9b19976fe5c268aa72760635cf3cc@%3Cuser.flink.apache.org%3E

[2]: https://lists.apache.org/thread.html/bf789df06e979f80caf23f6b2c8676aaf07b007ae0d450ae887b6a82@%3Cuser.flink.apache.org%3E