[
https://issues.apache.org/jira/browse/SPARK-17428?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15473643#comment-15473643
]
Yanbo Liang commented on SPARK-17428:
-------------------------------------
[~sunrui] [~shivaram] [~felixcheung] Thanks for your reply.
Yes, we can compile packages at driver and send them to executors. But it
involves some issues:
* Usually the Spark job is not run as root, but we need root privilege to
install R packages on executors which is not permitted.
* After we run a SparkR job, the executors' R libraries will be polluted. And
when another job was running on that executor, it may failed due to some
conflict.
* The architecture of driver and executor may different, so the packages
compiled on driver may not work well when it was sending to executors if it
dependent on some architecture-related code.
These issues can not solved by SparkR currently. I investigated and found
packrat can help us on this direction, but may be need more experiments. If
this proposal make sense, I can work on this feature. Please feel free to let
me know what you concern about. Thanks!
> SparkR executors/workers support virtualenv
> -------------------------------------------
>
> Key: SPARK-17428
> URL: https://issues.apache.org/jira/browse/SPARK-17428
> Project: Spark
> Issue Type: New Feature
> Components: SparkR
> Reporter: Yanbo Liang
>
> Many users have requirements to use third party R packages in
> executors/workers, but SparkR can not satisfy this requirements elegantly.
> For example, you should to mess with the IT/administrators of the cluster to
> deploy these R packages on each executors/workers node which is very
> inflexible.
> I think we should support third party R packages for SparkR users as what we
> do for jar packages in the following two scenarios:
> 1, Users can install R packages from CRAN or custom CRAN-like repository for
> each executors.
> 2, Users can load their local R packages and install them on each executors.
> To achieve this goal, the first thing is to make SparkR executors support
> virtualenv like Python conda. I have investigated and found
> packrat(http://rstudio.github.io/packrat/) is one of the candidates to
> support virtualenv for R. Packrat is a dependency management system for R and
> can isolate the dependent R packages in its own private package space. Then
> SparkR users can install third party packages in the application
> scope(destroy after the application exit) and don’t need to bother
> IT/administrators to install these packages manually.
> I would like to know whether it make sense.
--
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
