[jira] [Created] (SPARK-5171) Standalone cluster: masters scheduling independently

["Roberto Vaquerizo Rodriguez (JIRA)" <ji...@apache.org>]

Roberto Vaquerizo Rodriguez created SPARK-5171:
--------------------------------------------------

             Summary: Standalone cluster: masters scheduling independently
                 Key: SPARK-5171
                 URL: https://issues.apache.org/jira/browse/SPARK-5171
             Project: Spark
          Issue Type: Bug
          Components: Deploy
    Affects Versions: 1.2.0
            Reporter: Roberto Vaquerizo Rodriguez


Hi. 

I'm trying to configure a spark standalone cluster (release = spark 1.2), with three master nodes (bigdata1, bigdata2 and bigdata3) managed by Zookeeper. 

It seems there's a configuration problem, since every master node is saying it is the cluster leader: 

     ......... 
     14/12/30 13:54:59 INFO Master: I have been elected leader! New state: ALIVE 

The message above is dumped by every master I start up. 


Zookeeper is configured identically in all of them, as follows: 
........ 
dataDir=/spark 
........ 

The only difference is the "myid" file in the /spark directory, of course. 


The masters are started using the following configuration: 
......... 
export SPARK_DAEMON_JAVA_OPTS=" \ 
-Dspark.deploy.recoverymode=ZOOKEEPER \ 
-Dspark.deploy.zookeeper.url=bigdata1:2181,bigdata2:2181,bigdata3:2181" 

I have not set the spark.deploy.zookeeper.dir variable, since I'm using the default value, /spark, configured in zookeeper, as I mentioned before. 

I would like to know if there is any other thing I have to configure, in order to make the masters behave correctly (only one master node active at a time, while the others are in stand-by mode). 

With the current situation, I can connect workers and applications to the whole cluster. For instance, I can connect a worker to the cluster using: 

spark-class org.apache.spark.deploy.worker.Worker spark://bigdata1:2181,bigdata2:2181,bigdata3:2181 


But the worker gets registered to each of the masters independently. 
If I stop one of the masters, it tries to re-register to it. 
The notion of only one active master is completely lost. All masters are scheduling independently. 

If I run "zkServer.sh status" in the cluster, only one of the master nodes says "leader", and the other ones say "follower". So, from a Zookeeper point of view, it seems that the configuration is correct. 
But the standalone cluster is not  interacting with zookeeper properly. 

Do you have any idea? 



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

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org