So, I am running into this again, net=host isn't an option for me, because then my hadoop client conflicts with the nodes hadoop client. So I am trying to work out net=bridge, however, even specifying the driver port and host in spark doesn't work because spark tries to listen on only the hostname specified in spark.driver.host, not on all interfaces, thus it doesn't work inside the container. Anyone else have luck with running Zeppelin in a Docker container, and using Yarn?
On Wed, Jun 17, 2015 at 12:20 PM, John Omernik <j...@omernik.com> wrote: > My challenge is I am running this in mesos, therefore I have a limited > amount of ports and I would need to know them ahead of time. Even with > that ,the IP provided to the Application Master by the is internal Docker > IP, so that proves challenging as well. Any other thoughts on this? > > On Wed, Jun 17, 2015 at 12:03 PM, Jong Wook Kim <jongw...@nyu.edu> wrote: > >> I just used the --net=host option to make the container just use the >> host’s network stack which is probably the easiest way. >> >> if you are not allowed to do that, change the container’s >> /proc/sys/net/ipv4/ip_local_port_range >> to have a narrower port range, like 59000-59999 and forward the entire >> range. >> >> >> On Jun 18, 2015, at 1:15 AM, John Omernik <j...@omernik.com> wrote: >> >> So I think I ran across an issue with Zeppelin/Docker but I wanted to >> describe here and see what people thought. >> >> Basically I have a YARN cluster (Yarn... well Myriad on Mesos, but for >> pyspark, it works like Yarn). >> >> I have setup a docker container with all I need including Spark etc. >> >> When I try to run pyspark, my Application Master in Yarn tries to connect >> back to the driver which is inconveniently located in a docker container, >> and the only ports exposed are those for Zeppelin. (My yarn spark >> applicaiton is accepted, but I see in the logs, and it never "runs" >> >> 15/06/17 11:04:29 ERROR yarn.ApplicationMaster: Failed to connect to driver >> at 172.17.0.16:59601, retrying ... >> 15/06/17 11:04:32 ERROR yarn.ApplicationMaster: Failed to connect to driver >> at 172.17.0.16:59601, retrying . >> >> >> Ok, so yarn-client mode won't work (easily), however, yarn-cluster mode >> doesn't work either because you can't run a shell in yarn-cluster mode. >> >> So this is all just trying to run $SPARK_HOME/bin/pyspark --master >> yarn-client from inside of my Zeppelin Docker container (not even running >> Zeppelin yet) >> >> I am guessing that the issue is part of my issues I am having getting the >> spark interpreter running in Zeppelin because Zeppelin is doing just that, >> running pyspark in yarn-client mode. >> >> Ok, so how have people running Zeppelin in Docker dealt with this? What >> other ideas should I look into here? >> >> >> >> >
