SQL should not be the cause here. It is true that Flink removes the timestamp from a record when entering the SQL API but this timestamp is set again before time-based operations such as OVER windows. Watermarks are not touched. I think your issue is related to . One explanation that I could image is that a watermark has to be broadcasted to all workers, depending on the frequency of watermark generation, a higher parallelism also causes a higher amout of watermarks to be broadcasted. Which Flink version are you using? Can you run your tests on the latest Flink RC, because there where some improvements to the buffer management that might help. I will loop in Piotr for that.
Am 24.05.18 um 01:02 schrieb Yan Zhou [FDS Science]: