Re: Jira issue Flink-11127

Posted by Konstantin Knauf-2 on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/Jira-issue-Flink-11127-tp26180p26254.html

Hi Boris,

the exact command depends on the docker-entrypoint.sh script and the image you are using. For the example contained in the Flink repository it is "task-manager", I think. The important thing is to pass "taskmanager.host" to the Taskmanager process. You can verify by checking the Taskmanager logs. These should contain lines like below:

2019-02-21 08:03:00,004 INFO  org.apache.flink.runtime.taskexecutor.TaskManagerRunner      [] -  Program Arguments:
2019-02-21 08:03:00,008 INFO  org.apache.flink.runtime.taskexecutor.TaskManagerRunner      [] -     -Dtaskmanager.host=10.12.10.173

In the Jobmanager logs you should see that the Taskmanager is registered under the IP above in a line similar to:

2019-02-21 08:03:26,874 INFO  org.apache.flink.runtime.resourcemanager.StandaloneResourceManager [] - Registering TaskManager with ResourceID a0513ba2c472d2d1efc07626da9c1bda (akka.tcp://flink@10.12.10.173:46531/user/taskmanager_0) at ResourceManager

A service per Taskmanager is not required. The purpose of the config parameter is that the Jobmanager addresses the taskmanagers by IP instead of hostname.

Hope this helps!

Cheers,

Konstantin

On Wed, Feb 20, 2019 at 4:37 PM Boris Lublinsky <[hidden email]> wrote:

Also, The suggested workaround does not quite work.

2019-02-20 15:27:43,928 WARN akka.remote.ReliableDeliverySupervisor - Association with remote system [akka.tcp://flink-metrics@flink-taskmanager-1:6170] has failed, address is now gated for [50] ms. Reason: [Association failed with [akka.tcp://flink-metrics@flink-taskmanager-1:6170]] Caused by: [flink-taskmanager-1: No address associated with hostname]
	2019-02-20 15:27:48,750 ERROR org.apache.flink.runtime.rest.handler.legacy.files.StaticFileServerHandler - Caught exception

I think the problem is that its trying to connect to flink-task-manager-1

Using busybody to experiment with nslookup, I can see

/ # nslookup flink-taskmanager-1.flink-taskmanager

Server:    10.0.11.151

Address 1: 10.0.11.151 ip-10-0-11-151.us-west-2.compute.internal

Name:      flink-taskmanager-1.flink-taskmanager

Address 1: 10.131.2.136 flink-taskmanager-1.flink-taskmanager.flink.svc.cluster.local

/ # nslookup flink-taskmanager-1

Server:    10.0.11.151

Address 1: 10.0.11.151 ip-10-0-11-151.us-west-2.compute.internal

nslookup: can't resolve 'flink-taskmanager-1'

/ # nslookup flink-taskmanager-0.flink-taskmanager

Server:    10.0.11.151

Address 1: 10.0.11.151 ip-10-0-11-151.us-west-2.compute.internal

Name:      flink-taskmanager-0.flink-taskmanager

Address 1: 10.131.0.111 flink-taskmanager-0.flink-taskmanager.flink.svc.cluster.local

/ # nslookup flink-taskmanager-0

Server:    10.0.11.151

Address 1: 10.0.11.151 ip-10-0-11-151.us-west-2.compute.internal

nslookup: can't resolve 'flink-taskmanager-0'

/ # 

So the name should be postfixed with the service name. How do I force it? I suspect I am missing config parameter

 

Boris Lublinsky
FDP Architect
[hidden email]
https://www.lightbend.com/

On Feb 19, 2019, at 4:33 AM, Konstantin Knauf <[hidden email]> wrote:

Hi Boris,

the solution is actually simpler than it sounds from the ticket. The only thing you need to do is to set the "taskmanager.host" to the Pod's IP address in the Flink configuration. The easiest way to do this is to pass this config dynamically via a command-line parameter. 

The Deployment spec could looks something like this:

containers:
- name: taskmanager
  [...]
  args:
  - "taskmanager.sh"
  - "start-foreground"
  - "-Dtaskmanager.host=$(K8S_POD_IP)"
  [...]
  env:
  - name: K8S_POD_IP
    valueFrom:
      fieldRef:
        fieldPath: status.podIP

Hope this helps and let me know if this works. 

Best, 

Konstantin

On Sun, Feb 17, 2019 at 9:51 PM Boris Lublinsky <[hidden email]> wrote:

I was looking at this issue https://issues.apache.org/jira/browse/FLINK-11127

Apparently there is a workaround for it.

Is it possible provide the complete helm chart for it.

Bits and pieces are in the ticket, but it would be nice to see the full chart

Boris Lublinsky
FDP Architect
[hidden email]
https://www.lightbend.com/

--

--