[
https://issues.apache.org/jira/browse/SPARK-10987?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14947893#comment-14947893
]
Marcelo Vanzin commented on SPARK-10987:
----------------------------------------
Anyway, here's what I found so far.
Driver launches AM; AM connects back to driver and sends stuff. But driver
never sends any messages to AM. That means that in
{{NettyRpcHandler::connectionTerminated}}, the {{Disassociated}} message is not
sent, because since no message was sent *to* to the AM, the code in
{{NettyRpcHandler::receive}} was never run, so the driver connection was never
recorded.
So there must be a way for {{NettyRpcHandler}} to know when outgoing
connections are killed, not just incoming ones.
In a way this is caused by the code trying to mimic what akka does, but failing
at it; since the AM is purely a client, it shouldn't need to listen for
connections and rely on incoming connections for anything - it should be able
to register itself and do everything using the client socket it opened. That's
probably going to be tricky to fix, though.
> yarn-client mode misbehaving with netty-based RPC backend
> ---------------------------------------------------------
>
> Key: SPARK-10987
> URL: https://issues.apache.org/jira/browse/SPARK-10987
> Project: Spark
> Issue Type: Bug
> Components: Spark Core, YARN
> Affects Versions: 1.6.0
> Reporter: Marcelo Vanzin
> Priority: Blocker
>
> YARN running in cluster deploy mode seems to be having issues with the new
> RPC backend; if you look at unit test runs, tests that run in cluster mode
> are taking several minutes to run, instead of the more usual 20-30 seconds.
> For example,
> https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/43349/consoleFull:
> {noformat}
> [info] YarnClusterSuite:
> [info] - run Spark in yarn-client mode (13 seconds, 953 milliseconds)
> [info] - run Spark in yarn-cluster mode (6 minutes, 50 seconds)
> [info] - run Spark in yarn-cluster mode unsuccessfully (1 minute, 53 seconds)
> [info] - run Python application in yarn-client mode (21 seconds, 842
> milliseconds)
> [info] - run Python application in yarn-cluster mode (7 minutes, 0 seconds)
> [info] - user class path first in client mode (1 minute, 58 seconds)
> [info] - user class path first in cluster mode (4 minutes, 49 seconds)
> {noformat}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
