[
https://issues.apache.org/jira/browse/FLINK-2279?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14603173#comment-14603173
]
ASF GitHub Bot commented on FLINK-2279:
---------------------------------------
Github user gyfora commented on the pull request:
https://github.com/apache/flink/pull/870#issuecomment-115756972
There is nothing tricky there:
In java every record in the stream passed to closewith() will be fed back.
(And closeWith returns this stream).
In scala its even clearer as the stepfunction returns 2 streams in a tuple.
The first one is the feedback the second is the output.
The methods are not the same but the functionality is equivalent.
> Allow treating iteration head as ConnectedDataStream
> -----------------------------------------------------
>
> Key: FLINK-2279
> URL: https://issues.apache.org/jira/browse/FLINK-2279
> Project: Flink
> Issue Type: New Feature
> Components: Streaming
> Affects Versions: 0.10
> Reporter: Gyula Fora
> Assignee: Gyula Fora
> Priority: Minor
>
> Currently the streaming iterations are restricted to use the same input and
> feedback types which are routed through the same operator.
> This means that if the user want to distinguish between normal input and
> feedback record he/she needs to mark it somehow and also a wrapper type is
> necessary for handling separate input and feedback types.
> This makes implementing iterative algorithms (such as ML) quite ugly at some
> points.
> I propose to let the user treat the normal input if the iteration head
> operator and the feedback input as a ConnectedDataStream which can be used to
> apply co-operators both distinguishing the inputs and allowing different
> feedback types for elegant implementations.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
