[
https://issues.apache.org/jira/browse/FLINK-2644?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14736828#comment-14736828
]
Aljoscha Krettek commented on FLINK-2644:
-----------------------------------------
I think we have to ensure that the same partitioner (key-selector) is used for
all inputs. Otherwise elements could go to different parallel operators and
then the state for one key would not reside on only one machine.
> State partitioning does not respect the different partitioning of multiple
> inputs
> ---------------------------------------------------------------------------------
>
> Key: FLINK-2644
> URL: https://issues.apache.org/jira/browse/FLINK-2644
> Project: Flink
> Issue Type: Bug
> Components: Streaming
> Reporter: Gyula Fora
>
> Currently state partitioning is only available for one-input stream operators
> and even for those, the partitioning will only depend on the first input.
> This happens because the keyselector used to extract the partitioning key
> from each element is used from the first input stream. We need to make sure
> that for each input stream the corresponding key selector is used.
> To do this, ctx.nextRecord(record), should be extended by adding a parameter
> denoting the index of the input channel which can be used by the partitioner
> to select the appropriate key selector.
> The same mechanics can be extended to allow state partitioning for TwoInput
> stream operators as well.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
