[
https://issues.apache.org/jira/browse/FLINK-2967?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14991682#comment-14991682
]
Robert Metzger commented on FLINK-2967:
---------------------------------------
The problem is caused by a too short initial timeout.
The TaskManager tried to connect to the JobManager 5 seconds after the JM
started, so there was enough time (in the case the user reported). The YARN AM
starts requesting containers after the actor system on the JM has been started,
so we can assume the JM to run.
I've created a fix for the affected user with the timeout set to 5 seconds
(because I know that InetAddress.getLocalHost() returns the right address).
Once the user confirmed the fix, we can increase the timeout in our code.
> TM address detection might not always detect the right interface on slow
> networks / overloaded JMs
> --------------------------------------------------------------------------------------------------
>
> Key: FLINK-2967
> URL: https://issues.apache.org/jira/browse/FLINK-2967
> Project: Flink
> Issue Type: Bug
> Affects Versions: 0.9, 0.10, 1.0
> Reporter: Robert Metzger
> Assignee: Robert Metzger
>
> I'm talking to a user which is facing the following issue:
> Some of the TaskManagers select the wrong IP address out of the available
> network interfaces.
> The first address we try to connect to is the one returned by
> {{InetAddress.getLocalHost()}}. This address is the right IP address to use,
> but the JobManager is not able to respond within the timeout (50ms) to that
> connection request.
> So the TM tries the next address, which is not publicly reachable. However,
> the TM can connect to the JM from there. Netty will later fail to connect to
> the TM from the other TMs.
> There are two solutions for this issue:
> - Allow users to configure a higher timeout for the first address detection
> strategy. In most cases, the address returned by
> {{InetAddress.getLocalHost()}} is correct. By setting a high timeout, users
> with slow networks / overloaded JMs can make sure the TM picks this address
> - add an Akka message which we send from the TM to the JM, and the JM tries
> to connect to the TM. If that succeeds, we know that the TM is reachable from
> the outside.
> The problem is that we have to start a separate actor system on the
> TaskManager first. We have to do this because might use a wrong ip address
> for the TM (so we might end up starting actor systems until we found an
> externally reachable ip)
> I'm first going to implement the first approach. If that solution works well
> for my user, I'll contribute this to 0.10 / 1.0.
> If not, I'll implement the second approach.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
