Hello,
I build Spark on HDFS/YARN cluster on Docker Containers.
* Spark on YARN version
- Spark 1.6.0
- Hadoop 2.6.0 (CDH 5.6.0)
- Oracle Java 1.8.0_74
There are one HDFS/YARN master and one HDFS/YARN worker on each containers.
spark-yarn-master container has below hostname and IP addr.
hostname: spark-yarn-master-1-sxegt (pod name)
IP addr.: 172.17.0.11
hostname: spark-yarn-master (alias DNS name)
IP addr.: 172.30.242.57 (alias IP addr.)
bash-4.2$ cat /etc/hosts
# Kubernetes-managed hosts file.
127.0.0.1 localhost
::1 localhost ip6-localhost ip6-loopback
fe00::0 ip6-localnet
fe00::0 ip6-mcastprefix
fe00::1 ip6-allnodes
fe00::2 ip6-allrouters
172.17.0.11 spark-yarn-master-1-sxegt
bash-4.2$
bash-4.2$ ip -4 addr show dev eth0
50: eth0@if51: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state
UP link-netnsid 0
inet 172.17.0.11/16 scope global eth0
valid_lft forever preferred_lft forever
bash-4.2$
bash-4.2$ hostname -f
spark-yarn-master-1-sxegt
bash-4.2$
bash-4.2$ curl -v spark-yarn-master:8020
* About to connect() to spark-yarn-master port 8020 (#0)
* Trying 172.30.242.57...
* Connected to spark-yarn-master (172.30.242.57) port 8020 (#0)
> GET / HTTP/1.1
> User-Agent: curl/7.29.0
> Host: spark-yarn-master:8020
> Accept: */*
>
< HTTP/1.1 404 Not Found
< Content-type: text/plain
* no chunk, no close, no size. Assume close to signal end
<
It looks like you are making an HTTP request to a Hadoop IPC port. This is not
the correct port for the web interface on this daemon.
* Closing connection 0
bash-4.2$
spark-yarn-worker container has below hostname and IP addr.
hostname: spark-yarn-worker-1-pshqi (pod name)
IP addr.: 172.17.0.12
hostname: spark-yarn-worker (alias DNS name)
IP addr.: 172.30.1.53 (alias IP addr.)
bash-4.2$ cat /etc/hosts
# Kubernetes-managed hosts file.
127.0.0.1 localhost
::1 localhost ip6-localhost ip6-loopback
fe00::0 ip6-localnet
fe00::0 ip6-mcastprefix
fe00::1 ip6-allnodes
fe00::2 ip6-allrouters
172.17.0.12 spark-yarn-worker-1-pshqi
bash-4.2$
bash-4.2$ ip -4 addr show dev eth0
52: eth0@if53: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state
UP link-netnsid 0
inet 172.17.0.12/16 scope global eth0
valid_lft forever preferred_lft forever
bash-4.2$
bash-4.2$ hostname -f
spark-yarn-worker-1-pshqi
bash-4.2$
bash-4.2$ curl -v spark-yarn-worker:8040
* About to connect() to spark-yarn-worker port 8040 (#0)
* Trying 172.30.1.53...
* Connected to spark-yarn-worker (172.30.1.53) port 8040 (#0)
> GET / HTTP/1.1
> User-Agent: curl/7.29.0
> Host: spark-yarn-worker:8040
> Accept: */*
>
< HTTP/1.1 404 Not Found
< Content-type: text/plain
* no chunk, no close, no size. Assume close to signal end
<
It looks like you are making an HTTP request to a Hadoop IPC port. This is not
the correct port for the web interface on this daemon.
* Closing connection 0
bash-4.2$
Spark HDFS/YARN master/worker nodes can connect each other by alias DNS name.
On master, To worker (alias DNS name):
bash-4.2$ hostname -f ; curl spark-yarn-worker:8040
spark-yarn-master-1-sxegt
It looks like you are making an HTTP request to a Hadoop IPC port. This is not
the correct port for the web interface on this daemon.
bash-4.2$
On worker, To master (alias DNS name):
bash-4.2$ hostname -f ; curl spark-yarn-master:8020
spark-yarn-worker-1-pshqi
It looks like you are making an HTTP request to a Hadoop IPC port. This is not
the correct port for the web interface on this daemon.
bash-4.2$
They cannot connect each other by hostname.
On master, To worker (hostname):
bash-4.2$ hostname -f ; curl spark-yarn-master-1-sxegt:8020
spark-yarn-worker-1-pshqi
curl: (6) Could not resolve host: spark-yarn-master-1-sxegt; Name or service
not known
bash-4.2$
On worker, To master (hostname):
bash-4.2$ hostname -f ; curl spark-yarn-worker-1-pshqi:8040
spark-yarn-master-1-sxegt
curl: (6) Could not resolve host: spark-yarn-worker-1-pshqi; Name or service
not known
bash-4.2$
So, I want HDFS/YARN to use alias DNS name instead of hostname.
But YARN nodemanager always uses hostname
even configuredyarn.nodemanager.hostname to alias DNS name.
HDFS/YARN worker log:
16/03/07 10:04:38 INFO datanode.DataNode: Configured hostname is
spark-yarn-worker
:
16/03/07 10:04:42 INFO security.NMContainerTokenSecretManager: Updating node
address : spark-yarn-worker-1-pshqi:39352
:
16/03/07 10:04:42 INFO containermanager.ContainerManagerImpl: ContainerManager
started at spark-yarn-worker-1-pshqi/172.17.0.12:39352
:
16/03/07 10:04:44 INFO nodemanager.NodeStatusUpdaterImpl: Registered with
ResourceManager as spark-yarn-worker-1-pshqi:39352 with total resource of
<memory:8192, vCores:8>
:
16/03/07 10:04:44 INFO common.Storage: Lock on
/var/lib/hadoop-hdfs/cache/hdfs/dfs/data/in_use.lock acquired by nodename
16@spark-yarn-worker-1-pshqi
:
HDFS/YARN master log:
16/03/07 10:04:44 INFO resourcemanager.ResourceTrackerService: NodeManager from
node spark-yarn-worker-1-pshqi(cmPort: 39352 httpPort: 8042) registered with
capability: <memory:8192, vCores:8>, assigned nodeId
spark-yarn-worker-1-pshqi:39352
:
16/03/07 10:04:45 INFO rmnode.RMNodeImpl: spark-yarn-worker-1-pshqi:39352 Node
Transitioned from NEW to RUNNING
16/03/07 10:04:45 INFO fair.FairScheduler: Added node
spark-yarn-worker-1-pshqi:39352 cluster capacity: <memory:8192, vCores:8>
When submit application to spark master, spark master connects to worker,
but it uses "unresolved" worker's hostname not alias DNS name,
it throws java.net.UnknownHostException.
16/03/07 11:58:51 INFO scheduler.SchedulerNode: Assigned container
container_1457312681433_0001_01_000001 of capacity <memory:1024, vCores:1> on
host spark-yarn-worker-1-pshqi:39352, which has 1 containers, <memory:1024,
vCores:1> used and <memory:7168, vCores:7> available after allocation
16/03/07 11:58:51 ERROR scheduler.SchedulerApplicationAttempt: Error trying to
assign container token and NM token to an allocated container
container_1457312681433_0001_01_000001
java.lang.IllegalArgumentException: java.net.UnknownHostException:
spark-yarn-worker-1-pshqi
Can I configure YARN nodemanager to use arbitrary hostname?
Sorry to say, there is no chance to modify container's hostname, /etc/hosts and
DNS.
Regards,
dai
--
HIGUCHI Daisuke <[email protected]>
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
