Hi Ted, Thanks for your reply. I think zookeeper is an optional dependency of Spark. To enable it, I essentially use this flags on all my spark-env.sh:
SPARK_DAEMON_JAVA_OPTS="-Dspark.deploy.recoveryMode=ZOOKEEPER -Dspark.deploy.zookeeper.url=my-zoo-ip:2181" and of course, I have my zookeeper runing on my-zoo-ip:2181 (just 1 zookeeper node at this stage) spark master election seems to work as only 1 master is active, and the remaining one are in standby. I can't tell you how the election / discovery of master works, I assume the zookeeper module of spark is somehow using zookeeper primitives to do the election. The thing that is blocking me is that the worker don't seem to discover the first-ever elected-master. Cheers, Christophe On Fri, Jul 31, 2015 at 2:18 PM, Ted Yu <yuzhih...@gmail.com> wrote: > zookeeper is not a direct dependency of Spark. > > Can you give a bit more detail on how the election / discovery of master > works ? > > Cheers > > On Thu, Jul 30, 2015 at 7:41 PM, Christophe Schmitz <cofcof...@gmail.com> > wrote: > >> Hi there, >> >> I am trying to run a 3 node spark cluster where each nodes contains a >> spark worker and a spark maser. Election of the master happens via >> zookeeper. >> >> The way I am configuring it is by (on each node) giving the IP:PORT of >> the local master to the local worker, and I wish the worker could >> autodiscover the elected master automatically. >> >> But unfortunatly, only the local worker of the elected master registered >> to the elected master. Why aren't the other worker getting to connect to >> the elected master? >> >> The interessing thing is that if I kill the elected master and wait a >> bit, then the new elected master sees all the workers! >> >> I am wondering if I am missing something to make this happens without >> having to kill the elected master. >> >> Thanks! >> >> >> PS: I am on spark 1.2.2 >> >> >
