[
https://issues.apache.org/jira/browse/FLINK-16929?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17076137#comment-17076137
]
Aljoscha Krettek commented on FLINK-16929:
------------------------------------------
Could you check the {{numLateRecordsDropped}} metric to see if there are
actually records that are dropped as late? There could be something awry with
the watermark.
> Session Window produces sessions randomly
> -----------------------------------------
>
> Key: FLINK-16929
> URL: https://issues.apache.org/jira/browse/FLINK-16929
> Project: Flink
> Issue Type: Bug
> Affects Versions: 1.9.1
> Reporter: Ori Popowski
> Priority: Major
> Attachments: image-2020-04-01-19-56-00-239.png,
> image-2020-04-01-19-56-27-720.png
>
>
> We have a Flink job which keyBys session ID (sId), and uses a session window
> with 30 minutes gap:
> {code:java}
> inputStream
> .keyBy(keySelector)
> .window(EventTimeSessionWindows.withGap(Time.minutes(30)))
> .allowedLateness(Time.seconds(0L))
> {code}
> This Flink job reads from Kinesis stream.
> Lately (I suspect after upgrading from 1.5.4 to 1.9.1) we get too many
> sessions, with gaps of several seconds (instead of 30 minutes).
> We have no idea why it's happening and suspect a Flink bug or a state backend
> bug (we use RocksDB).
> I haven't found any indication in the logs except for some read throughput
> warnings which were resolved by a backoff.
> Attached is a table of derived sessions, and then the raw events
> *Sessions*
> !image-2020-04-01-19-56-00-239.png|width=753,height=406!
>
> *Events*
>
> !image-2020-04-01-19-56-27-720.png|width=312,height=383!
>
>
>
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
