[
https://issues.apache.org/jira/browse/SPARK-19900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15999741#comment-15999741
]
lyc edited comment on SPARK-19900 at 5/7/17 9:19 AM:
-----------------------------------------------------
I can successfully reproduce in spark master ( commit 63d90e7d, and spark
version is 2.3.0), though the symptom is a little different.
To make the application run slowly enough, I modify `examples/LogQuery.scala`
to this:
{code}
dataSet.map(line => (extractKey(line), extractStats(line)))
.reduceByKey((a, b) => {
Thread.sleep(20*60*1000)
a.merge(b)
})
{code}
run two worker each with 1G memory and 2 cores. submit the application with:
{code}
bin/spark-submit --class org.apache.spark.examples.LogQuery --master
spark://master:6066 --deploy-mode cluster --supervise --total-executor-cores 1
--num-executors 1 --executor-memory 512M --driver-memory 512M
/vagrant_data/spark/examples/target/scala-2.11/jars/spark-examples_2.11-2.3.0-SNAPSHOT.jar
{code}
and follow the steps as stated above.
The outcome is that at step 5, there are not two drivers, but there are two
applications, the old one is still running, the new one is waiting for
resources. After step 6, the old application is finished, the new one running,
and the only one driver is killed. This is problematic because the driver
corresponding to the running application is killed instead of running.
The problem is due to the fact that the relaunched driver uses the same driver
id with the original one and when old application is killed, the driver is
killed as a side effect. I have create a pull request to fix this.
was (Author: lyc):
I can successfully reproduce in spark master ( commit 63d90e7d, and spark
version is 2.3.0), though the symptom is a little different.
To make the application run slowly enough, I modify `examples/LogQuery.scala`
to this:
```
dataSet.map(line => (extractKey(line), extractStats(line)))
.reduceByKey((a, b) => {
Thread.sleep(20*60*1000)
a.merge(b)
})
```
run two worker each with 1G memory and 2 cores. submit the application with:
```
bin/spark-submit --class org.apache.spark.examples.LogQuery --master
spark://master:6066 --deploy-mode cluster --supervise --total-executor-cores 1
--num-executors 1 --executor-memory 512M --driver-memory 512M
/vagrant_data/spark/examples/target/scala-2.11/jars/spark-examples_2.11-2.3.0-SNAPSHOT.jar
```
and follow the steps as stated above.
The outcome is that at step 5, there are not two drivers, but there are two
applications, the old one is still running, the new one is waiting for
resources. After step 6, the old application is finished, the new one running,
and the only one driver is killed. This is problematic because the driver
corresponding to the running application is killed instead of running.
The problem is due to the fact that the relaunched driver uses the same driver
id with the original one and when old application is killed, the driver is
killed as a side effect. I have create a pull request to fix this.
> [Standalone] Master registers application again when driver relaunched
> ----------------------------------------------------------------------
>
> Key: SPARK-19900
> URL: https://issues.apache.org/jira/browse/SPARK-19900
> Project: Spark
> Issue Type: Bug
> Components: Deploy, Spark Core
> Affects Versions: 1.6.2
> Environment: Centos 6.5, spark standalone
> Reporter: Sergey
> Priority: Critical
> Labels: Spark, network, standalone, supervise
>
> I've found some problems when node, where driver is running, has unstable
> network. A situation is possible when two identical applications are running
> on a cluster.
> *Steps to Reproduce:*
> # prepare 3 node. One for the spark master and two for the spark workers.
> # submit an application with parameter spark.driver.supervise = true
> # go to the node where driver is running (for example spark-worker-1) and
> close 7077 port
> {code}
> # iptables -A OUTPUT -p tcp --dport 7077 -j DROP
> {code}
> # wait more 60 seconds
> # look at the spark master UI
> There are two spark applications and one driver. The new application has
> WAITING state and the second application has RUNNING state. Driver has
> RUNNING or RELAUNCHING state (It depends on the resources available, as I
> understand it) and it launched on other node (for example spark-worker-2)
> # open the port
> {code}
> # iptables -D OUTPUT -p tcp --dport 7077 -j DROP
> {code}
> # look an the spark UI again
> There are no changes
> In addition, if you look at the processes on the node spark-worker-1
> {code}
> # ps ax | grep spark
> {code}
> you will see that the old driver is still working!
> *Spark master logs:*
> {code}
> 17/03/10 05:26:27 WARN Master: Removing
> worker-20170310052240-spark-worker-1-35039 because we got no heartbeat in 60
> seconds
> 17/03/10 05:26:27 INFO Master: Removing worker
> worker-20170310052240-spark-worker-1-35039 on spark-worker-1:35039
> 17/03/10 05:26:27 INFO Master: Telling app of lost executor: 1
> 17/03/10 05:26:27 INFO Master: Telling app of lost executor: 0
> 17/03/10 05:26:27 INFO Master: Re-launching driver-20170310052347-0000
> 17/03/10 05:26:27 INFO Master: Launching driver driver-20170310052347-0000 on
> worker worker-20170310052411-spark-worker-2-40473
> 17/03/10 05:26:35 INFO Master: Registering app TestApplication
> 17/03/10 05:26:35 INFO Master: Registered app TestApplication with ID
> app-20170310052635-0001
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got status update for unknown executor
> app-20170310052354-0000/1
> 17/03/10 05:31:07 WARN Master: Got status update for unknown executor
> app-20170310052354-0000/0
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 WARN Master: Got heartbeat from unregistered worker
> worker-20170310052240-spark-worker-1-35039. Asking it to re-register.
> 17/03/10 05:31:07 INFO Master: Registering worker spark-worker-1:35039 with 8
> cores, 10.8 GB RAM
> 17/03/10 05:31:07 INFO Master: Launching executor app-20170310052354-0000/4
> on worker worker-20170310052240-spark-worker-1-35039
> 17/03/10 05:31:07 INFO Master: Launching executor app-20170310052354-0000/5
> on worker worker-20170310052240-spark-worker-1-35039
> {code}
> I expect the following behaviour:
> # when the driver is relaunching it should not create a new application or
> the old application should be removed
> # the process with old driver should be killed
> Correct me please if I do not understand something or I missed some settings.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
