[ 
https://issues.apache.org/jira/browse/FLINK-1523?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14382725#comment-14382725
 ] 

ASF GitHub Bot commented on FLINK-1523:
---------------------------------------

Github user vasia commented on the pull request:

    https://github.com/apache/flink/pull/537#issuecomment-86727594
  
    Hey @andralungu! Thanks for this PR. I know it took more effort than you 
might have thought initially :-)
    
    I have a few comments:
    - In my opinion, the degrees availability and the total number of vertices 
should be *optional*. Otherwise, we are introducing unnecessary overhead for 
the majority of cases, i.e. when these are not actually used by an algorithm.
    - The edge direction option and the degrees / numOfVertices should be 
*separate extensions* and not tight together, i.e. we should allow a user to 
choose the edge direction option independently of whether they also want to 
access the degrees. To be consistent with the rest of the 
VertexCentricIteration options, we could introduce methods 
`setMessagingDirection()`, `setInDegreesAvailable()` and so on.
    - The change should be transparent to user code, i.e. if a user has turned 
on the option for accessing inDegrees, they should be able to call 
`vertex.getInDegree()` inside the messaging/update functions, but 
`vertex.getValue()` should still return the vertex value. In other words, this 
Tuple3 you're using to store the degrees should be hidden from the public 
methods.
    - Finally, we should add documentation for these options, including 
description of how they work and some illustrative examples.
    
    Let me know what you think about the above and whether you have any 
questions!


> Vertex-centric iteration extensions
> -----------------------------------
>
>                 Key: FLINK-1523
>                 URL: https://issues.apache.org/jira/browse/FLINK-1523
>             Project: Flink
>          Issue Type: Improvement
>          Components: Gelly
>            Reporter: Vasia Kalavri
>            Assignee: Andra Lungu
>
> We would like to make the following extensions to the vertex-centric 
> iterations of Gelly:
> - allow vertices to access their in/out degrees and the total number of 
> vertices of the graph, inside the iteration.
> - allow choosing the neighborhood type (in/out/all) over which to run the 
> vertex-centric iteration. Now, the model uses the updates of the in-neighbors 
> to calculate state and send messages to out-neighbors. We could add a 
> parameter with value "in/out/all" to the {{VertexUpdateFunction}} and 
> {{MessagingFunction}}, that would indicate the type of neighborhood.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to