spark-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jonathan Kelly <jonathaka...@gmail.com>
Subject Re: Spark 2.0 on YARN - Files in config archive not ending up on executor classpath
Date Mon, 20 Jun 2016 20:20:08 GMT
Thanks for the confirmation! Shall I cut a JIRA issue?

On Mon, Jun 20, 2016 at 10:42 AM Marcelo Vanzin <vanzin@cloudera.com> wrote:

> I just tried this locally and can see the wrong behavior you mention.
> I'm running a somewhat old build of 2.0, but I'll take a look.
>
> On Mon, Jun 20, 2016 at 7:04 AM, Jonathan Kelly <jonathakamzn@gmail.com>
> wrote:
> > Does anybody have any thoughts on this?
> >
> > On Fri, Jun 17, 2016 at 6:36 PM Jonathan Kelly <jonathakamzn@gmail.com>
> > wrote:
> >>
> >> I'm trying to debug a problem in Spark 2.0.0-SNAPSHOT (commit
> >> bdf5fe4143e5a1a393d97d0030e76d35791ee248) where Spark's
> log4j.properties is
> >> not getting picked up in the executor classpath (and driver classpath
> for
> >> yarn-cluster mode), so Hadoop's log4j.properties file is taking
> precedence
> >> in the YARN containers.
> >>
> >> Spark's log4j.properties file is correctly being bundled into the
> >> __spark_conf__.zip file and getting added to the DistributedCache, but
> it is
> >> not in the classpath of the executor, as evidenced by the following
> command,
> >> which I ran in spark-shell:
> >>
> >> scala> sc.parallelize(Seq(1)).map(_ =>
> >> getClass().getResource("/log4j.properties")).first
> >> res3: java.net.URL = file:/etc/hadoop/conf.empty/log4j.properties
> >>
> >> I then ran the following in spark-shell to verify the classpath of the
> >> executors:
> >>
> >> scala> sc.parallelize(Seq(1)).map(_ =>
> >> System.getProperty("java.class.path")).flatMap(_.split(':')).filter(e =>
> >> !e.endsWith(".jar") && !e.endsWith("*")).collect.foreach(println)
> >> ...
> >>
> >>
> /mnt/yarn/usercache/hadoop/appcache/application_1466208403287_0003/container_1466208403287_0003_01_000003
> >>
> >>
> /mnt/yarn/usercache/hadoop/appcache/application_1466208403287_0003/container_1466208403287_0003_01_000003/__spark_conf__
> >> /etc/hadoop/conf
> >> ...
> >>
> >> So the JVM has this nonexistent __spark_conf__ directory in the
> classpath
> >> when it should really be __spark_conf__.zip (which is actually a
> symlink to
> >> a directory, despite the .zip filename).
> >>
> >> % sudo ls -l
> >>
> /mnt/yarn/usercache/hadoop/appcache/application_1466208403287_0003/container_1466208403287_0003_01_000003
> >> total 20
> >> -rw-r--r-- 1 yarn yarn   88 Jun 18 01:26 container_tokens
> >> -rwx------ 1 yarn yarn  594 Jun 18 01:26
> >> default_container_executor_session.sh
> >> -rwx------ 1 yarn yarn  648 Jun 18 01:26 default_container_executor.sh
> >> -rwx------ 1 yarn yarn 4419 Jun 18 01:26 launch_container.sh
> >> lrwxrwxrwx 1 yarn yarn   59 Jun 18 01:26 __spark_conf__.zip ->
> >> /mnt1/yarn/usercache/hadoop/filecache/17/__spark_conf__.zip
> >> lrwxrwxrwx 1 yarn yarn   77 Jun 18 01:26 __spark_libs__ ->
> >>
> /mnt/yarn/usercache/hadoop/filecache/16/__spark_libs__4490748779530764463.zip
> >> drwx--x--- 2 yarn yarn   46 Jun 18 01:26 tmp
> >>
> >> Does anybody know why this is happening? Is this a bug in Spark, or is
> it
> >> the JVM doing this (possibly because the extension is .zip)?
> >>
> >> Thanks,
> >> Jonathan
>
>
>
> --
> Marcelo
>

Mime
View raw message