[
https://issues.apache.org/jira/browse/SPARK-2691?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14194679#comment-14194679
]
Chris Heller commented on SPARK-2691:
-------------------------------------
Ok here is the patch as a PR: https://github.com/apache/spark/pull/3074
[~tarnfeld] feel free to expand on this patch. I was looking at the code today
and realized the coarse mode support should be trivial (just setting a
ContainerInfo inside the TaskInfo created) -- it just cannot reuse the
fine-grained code path in its current form since that assumes passing of an
ExecutorInfo, but it could easily be generalized over a ContainerInfo instead.
We are not shipping the spark image as an executor URI, instead spark is
bundled in the image. It is just a stock spark needed in the image, a simple
docker file would look like (assuming you have a spark tar ball and libmesos in
your directory with the Dockerfile):
{noformat}
FROM ubuntu
RUN apt-get -y update
RUN apt-get -y install default-jre-headless
RUN apt-get -y install python2.7
ADD spark-1.1.0-bin-hadoop1.tgz /
RUN mv /spark-1.1.0-bin-hadoop1 /spark
COPY libmesos-0.20.1.so /usr/lib/libmesos.so
ENV SPARK_HOME /spark
ENV MESOS_JAVA_NATIVE_LIBRARY /usr/lib/libmesos.so
CMD ps -ef
{noformat}
[~yoeduardoj] one awesome thing, which is actually beyond the scope of docker
support, but still related to mesos would be the ability to support
configuration of what role and attributes in a mesos offer are filtered by
spark -- but this is not relevant just wanted to bring it up while folks are
digging into the mesos backend code.
> Allow Spark on Mesos to be launched with Docker
> -----------------------------------------------
>
> Key: SPARK-2691
> URL: https://issues.apache.org/jira/browse/SPARK-2691
> Project: Spark
> Issue Type: Improvement
> Components: Mesos
> Reporter: Timothy Chen
> Assignee: Timothy Chen
> Labels: mesos
> Attachments: spark-docker.patch
>
>
> Currently to launch Spark with Mesos one must upload a tarball and specifiy
> the executor URI to be passed in that is to be downloaded on each slave or
> even each execution depending coarse mode or not.
> We want to make Spark able to support launching Executors via a Docker image
> that utilizes the recent Docker and Mesos integration work.
> With the recent integration Spark can simply specify a Docker image and
> options that is needed and it should continue to work as-is.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
