[
https://issues.apache.org/jira/browse/FLINK-1019?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14057348#comment-14057348
]
Asterios Katsifodimos commented on FLINK-1019:
----------------------------------------------
I have replaced the RPC service with Akka.io actors.
Instead of opening connections from the TaskManagers (TM) to the JobManager
(JM) and vice versa, there is an actor in each node (TM or JM) that receives
the serialized RPC calls and assigns them to a given internal Actor. The
internal actor executes the call and then returns the result to the calling
actor.
I've implemented this in two different ways:
1. Using Untyped actors
(http://doc.akka.io/docs/akka/2.3.4/java/untyped-actors.html)
2. Using TypedActors (http://doc.akka.io/docs/akka/2.3.4/java/typed-actors.html)
Both implementations are very small in size. 600 LOC for Untyped, around 300
LOC for the Typed one.
BUT: my implementation using Akka is slower than the old one. That's why I did
not pull-request it yet.
Today we will check it with [~StephanEwen] and see what can be wrong - I tried
lots of different configurations with no luck. However, I'm pretty confident
that this is not Akka's fault - there is something I'm missing and we will try
to find it today.
> Rework RPC service
> ------------------
>
> Key: FLINK-1019
> URL: https://issues.apache.org/jira/browse/FLINK-1019
> Project: Flink
> Issue Type: Improvement
> Components: Distributed Runtime
> Reporter: Ufuk Celebi
>
> There is work going on to improve the RPC service by using [Akka|akka.io]. I
> couldn't find a issue for it.
> Could one of the two people working on it ([~StephanEwen] and [~asteriosk])
> please give an overview of the changes and a status update?
--
This message was sent by Atlassian JIRA
(v6.2#6252)
