spark-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Cody Koeninger <c...@koeninger.org>
Subject Re: replacement for SPARK_JAVA_OPTS
Date Thu, 07 Aug 2014 21:35:56 GMT
Just wanted to check in on this, see if I should file a bug report
regarding the mesos argument propagation.


On Thu, Jul 31, 2014 at 8:35 AM, Cody Koeninger <cody@koeninger.org> wrote:

> 1. I've tried with and without escaping equals sign, it doesn't affect the
> results.
>
> 2. Yeah, exporting SPARK_SUBMIT_OPTS from spark-env.sh works for getting
> system properties set in the local shell (although not for executors).
>
> 3. We're using the default fine-grained mesos mode, not setting
> spark.mesos.coarse, so it doesn't seem immediately related to that ticket.
> Should I file a bug report?
>
>
> On Thu, Jul 31, 2014 at 1:33 AM, Patrick Wendell <pwendell@gmail.com>
> wrote:
>
>> The third issue may be related to this:
>> https://issues.apache.org/jira/browse/SPARK-2022
>>
>> We can take a look at this during the bug fix period for the 1.1
>> release next week. If we come up with a fix we can backport it into
>> the 1.0 branch also.
>>
>> On Wed, Jul 30, 2014 at 11:31 PM, Patrick Wendell <pwendell@gmail.com>
>> wrote:
>> > Thanks for digging around here. I think there are a few distinct issues.
>> >
>> > 1. Properties containing the '=' character need to be escaped.
>> > I was able to load properties fine as long as I escape the '='
>> > character. But maybe we should document this:
>> >
>> > == spark-defaults.conf ==
>> > spark.foo a\=B
>> > == shell ==
>> > scala> sc.getConf.get("spark.foo")
>> > res2: String = a=B
>> >
>> > 2. spark.driver.extraJavaOptions, when set in the properties file,
>> > don't affect the driver when running in client mode (always the case
>> > for mesos). We should probably document this. In this case you need to
>> > either use --driver-java-options or set SPARK_SUBMIT_OPTS.
>> >
>> > 3. Arguments aren't propagated on Mesos (this might be because of the
>> > other issues, or a separate bug).
>> >
>> > - Patrick
>> >
>> > On Wed, Jul 30, 2014 at 3:10 PM, Cody Koeninger <cody@koeninger.org>
>> wrote:
>> >> In addition, spark.executor.extraJavaOptions does not seem to behave
>> as I
>> >> would expect; java arguments don't seem to be propagated to executors.
>> >>
>> >>
>> >> $ cat conf/spark-defaults.conf
>> >>
>> >> spark.master
>> >>
>> mesos://zk://etl-01.mxstg:2181,etl-02.mxstg:2181,etl-03.mxstg:2181/masters
>> >> spark.executor.extraJavaOptions -Dfoo.bar.baz=23
>> >> spark.driver.extraJavaOptions -Dfoo.bar.baz=23
>> >>
>> >>
>> >> $ ./bin/spark-shell
>> >>
>> >> scala> sc.getConf.get("spark.executor.extraJavaOptions")
>> >> res0: String = -Dfoo.bar.baz=23
>> >>
>> >> scala> sc.parallelize(1 to 100).map{ i => (
>> >>      |  java.net.InetAddress.getLocalHost.getHostName,
>> >>      |  System.getProperty("foo.bar.baz")
>> >>      | )}.collect
>> >>
>> >> res1: Array[(String, String)] = Array((dn-01.mxstg,null),
>> >> (dn-01.mxstg,null), (dn-01.mxstg,null), (dn-01.mxstg,null),
>> >> (dn-01.mxstg,null), (dn-01.mxstg,null), (dn-01.mxstg,null),
>> >> (dn-01.mxstg,null), (dn-01.mxstg,null), (dn-01.mxstg,null),
>> >> (dn-01.mxstg,null), (dn-01.mxstg,null), (dn-02.mxstg,null),
>> >> (dn-02.mxstg,null), ...
>> >>
>> >>
>> >>
>> >> Note that this is a mesos deployment, although I wouldn't expect that
>> to
>> >> affect the availability of spark.driver.extraJavaOptions in a local
>> spark
>> >> shell.
>> >>
>> >>
>> >> On Wed, Jul 30, 2014 at 4:18 PM, Cody Koeninger <cody@koeninger.org>
>> wrote:
>> >>
>> >>> Either whitespace or equals sign are valid properties file formats.
>> >>> Here's an example:
>> >>>
>> >>> $ cat conf/spark-defaults.conf
>> >>> spark.driver.extraJavaOptions -Dfoo.bar.baz=23
>> >>>
>> >>> $ ./bin/spark-shell -v
>> >>> Using properties file: /opt/spark/conf/spark-defaults.conf
>> >>> Adding default property:
>> spark.driver.extraJavaOptions=-Dfoo.bar.baz=23
>> >>>
>> >>>
>> >>> scala>  System.getProperty("foo.bar.baz")
>> >>> res0: String = null
>> >>>
>> >>>
>> >>> If you add double quotes, the resulting string value will have double
>> >>> quotes.
>> >>>
>> >>>
>> >>> $ cat conf/spark-defaults.conf
>> >>> spark.driver.extraJavaOptions "-Dfoo.bar.baz=23"
>> >>>
>> >>> $ ./bin/spark-shell -v
>> >>> Using properties file: /opt/spark/conf/spark-defaults.conf
>> >>> Adding default property:
>> spark.driver.extraJavaOptions="-Dfoo.bar.baz=23"
>> >>>
>> >>> scala>  System.getProperty("foo.bar.baz")
>> >>> res0: String = null
>> >>>
>> >>>
>> >>> Neither one of those affects the issue; the underlying problem in my
>> case
>> >>> seems to be that bin/spark-class uses the SPARK_SUBMIT_OPTS and
>> >>> SPARK_JAVA_OPTS environment variables, but nothing parses
>> >>> spark-defaults.conf before the java process is started.
>> >>>
>> >>> Here's an example of the process running when only
>> spark-defaults.conf is
>> >>> being used:
>> >>>
>> >>> $ ps -ef | grep spark
>> >>>
>> >>> 514       5182  2058  0 21:05 pts/2    00:00:00 bash
>> ./bin/spark-shell -v
>> >>>
>> >>> 514       5189  5182  4 21:05 pts/2    00:00:22
>> /usr/local/java/bin/java
>> >>> -cp
>> >>>
>> ::/opt/spark/conf:/opt/spark/lib/spark-assembly-1.0.1-hadoop2.3.0-mr1-cdh5.0.2.jar:/etc/hadoop/conf-mx
>> >>> -XX:MaxPermSize=128m -Djava.library.path= -Xms512m -Xmx512m
>> >>> org.apache.spark.deploy.SparkSubmit spark-shell -v --class
>> >>> org.apache.spark.repl.Main
>> >>>
>> >>>
>> >>> Here's an example of it when the command line --driver-java-options
is
>> >>> used (and thus things work):
>> >>>
>> >>>
>> >>> $ ps -ef | grep spark
>> >>> 514       5392  2058  0 21:15 pts/2    00:00:00 bash
>> ./bin/spark-shell -v
>> >>> --driver-java-options -Dfoo.bar.baz=23
>> >>>
>> >>> 514       5399  5392 80 21:15 pts/2    00:00:06
>> /usr/local/java/bin/java
>> >>> -cp
>> >>>
>> ::/opt/spark/conf:/opt/spark/lib/spark-assembly-1.0.1-hadoop2.3.0-mr1-cdh5.0.2.jar:/etc/hadoop/conf-mx
>> >>> -XX:MaxPermSize=128m -Dfoo.bar.baz=23 -Djava.library.path= -Xms512m
>> >>> -Xmx512m org.apache.spark.deploy.SparkSubmit spark-shell -v
>> >>> --driver-java-options -Dfoo.bar.baz=23 --class
>> org.apache.spark.repl.Main
>> >>>
>> >>>
>> >>>
>> >>>
>> >>> On Wed, Jul 30, 2014 at 3:43 PM, Patrick Wendell <pwendell@gmail.com>
>> >>> wrote:
>> >>>
>> >>>> Cody - in your example you are using the '=' character, but in our
>> >>>> documentation and tests we use a whitespace to separate the key
and
>> >>>> value in the defaults file.
>> >>>>
>> >>>> docs: http://spark.apache.org/docs/latest/configuration.html
>> >>>>
>> >>>> spark.driver.extraJavaOptions -Dfoo.bar.baz=23
>> >>>>
>> >>>> I'm not sure if the java properties file parser will try to interpret
>> >>>> the equals sign. If so you might need to do this.
>> >>>>
>> >>>> spark.driver.extraJavaOptions "-Dfoo.bar.baz=23"
>> >>>>
>> >>>> Do those work for you?
>> >>>>
>> >>>> On Wed, Jul 30, 2014 at 1:32 PM, Marcelo Vanzin <vanzin@cloudera.com
>> >
>> >>>> wrote:
>> >>>> > Hi Cody,
>> >>>> >
>> >>>> > Could you file a bug for this if there isn't one already?
>> >>>> >
>> >>>> > For system properties SparkSubmit should be able to read those
>> >>>> > settings and do the right thing, but that obviously won't work
for
>> >>>> > other JVM options... the current code should work fine in cluster
>> mode
>> >>>> > though, since the driver is a different process. :-)
>> >>>> >
>> >>>> >
>> >>>> > On Wed, Jul 30, 2014 at 1:12 PM, Cody Koeninger <
>> cody@koeninger.org>
>> >>>> wrote:
>> >>>> >> We were previously using SPARK_JAVA_OPTS to set java system
>> properties
>> >>>> via
>> >>>> >> -D.
>> >>>> >>
>> >>>> >> This was used for properties that varied on a
>> >>>> per-deployment-environment
>> >>>> >> basis, but needed to be available in the spark shell and
workers.
>> >>>> >>
>> >>>> >> On upgrading to 1.0, we saw that SPARK_JAVA_OPTS had been
>> deprecated,
>> >>>> and
>> >>>> >> replaced by spark-defaults.conf and command line arguments
to
>> >>>> spark-submit
>> >>>> >> or spark-shell.
>> >>>> >>
>> >>>> >> However, setting spark.driver.extraJavaOptions and
>> >>>> >> spark.executor.extraJavaOptions in spark-defaults.conf
is not a
>> >>>> replacement
>> >>>> >> for SPARK_JAVA_OPTS:
>> >>>> >>
>> >>>> >>
>> >>>> >> $ cat conf/spark-defaults.conf
>> >>>> >> spark.driver.extraJavaOptions=-Dfoo.bar.baz=23
>> >>>> >>
>> >>>> >> $ ./bin/spark-shell
>> >>>> >>
>> >>>> >> scala> System.getProperty("foo.bar.baz")
>> >>>> >> res0: String = null
>> >>>> >>
>> >>>> >>
>> >>>> >> $ ./bin/spark-shell --driver-java-options "-Dfoo.bar.baz=23"
>> >>>> >>
>> >>>> >> scala> System.getProperty("foo.bar.baz")
>> >>>> >> res0: String = 23
>> >>>> >>
>> >>>> >>
>> >>>> >> Looking through the shell scripts for spark-submit and
>> spark-class, I
>> >>>> can
>> >>>> >> see why this is; parsing spark-defaults.conf from bash
could be
>> >>>> brittle.
>> >>>> >>
>> >>>> >> But from an ergonomic point of view, it's a step back to
go from a
>> >>>> >> set-it-and-forget-it configuration in spark-env.sh, to
requiring
>> >>>> command
>> >>>> >> line arguments.
>> >>>> >>
>> >>>> >> I can solve this with an ad-hoc script to wrap spark-shell
with
>> the
>> >>>> >> appropriate arguments, but I wanted to bring the issue
up to see
>> if
>> >>>> anyone
>> >>>> >> else had run into it,
>> >>>> >> or had any direction for a general solution (beyond parsing
java
>> >>>> properties
>> >>>> >> files from bash).
>> >>>> >
>> >>>> >
>> >>>> >
>> >>>> > --
>> >>>> > Marcelo
>> >>>>
>> >>>
>> >>>
>>
>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message