[
https://issues.apache.org/jira/browse/FLINK-7666?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16220118#comment-16220118
]
Kostas Kloudas commented on FLINK-7666:
---------------------------------------
If I understand correctly, then this is how is used to be, right?
The problem with closing the timer service before the operators is that the
continuous file reader defers its "official" closing time for when all splits
are read. In the meantime, if we close the timer service before that, then no
more watermarks are going to be emitted, which can be problematic.
> ContinuousFileReaderOperator swallows chained watermarks
> --------------------------------------------------------
>
> Key: FLINK-7666
> URL: https://issues.apache.org/jira/browse/FLINK-7666
> Project: Flink
> Issue Type: Improvement
> Components: Streaming Connectors
> Affects Versions: 1.3.2
> Reporter: Ufuk Celebi
> Assignee: Kostas Kloudas
> Priority: Blocker
> Fix For: 1.4.0
>
>
> I use event time and read from a (finite) file. I assign watermarks right
> after the {{ContinuousFileReaderOperator}} with parallelism 1.
> {code}
> env
> .readFile(new TextInputFormat(...), ...)
> .setParallelism(1)
> .assignTimestampsAndWatermarks(...)
> .setParallelism(1)
> .map()...
> {code}
> The watermarks I assign never progress through the pipeline.
> I can work around this by inserting a {{shuffle()}} after the file reader or
> starting a new chain at the assigner:
> {code}
> env
> .readFile(new TextInputFormat(...), ...)
> .setParallelism(1)
> .shuffle()
> .assignTimestampsAndWatermarks(...)
> .setParallelism(1)
> .map()...
> {code}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
