R on Spark (With Support for R on Spark Streaming)

This forked repository provides an R package to do Spark Streaming in R. SparkR is an R package that provides a light-weight frontend to use Spark from R.

NOTE: As of April 2015, SparkR has been merged into Apache Spark and is shipping in an upcoming release (1.4) due early summer 2015. This repo currently targets users using released versions of Spark. This repo no longer accepts new pull requests, and they should now be submitted to apache/spark; see here for some instructions.

Installing SparkR

Requirements

SparkR requires

• Scala 2.10, and
• Spark version >= 0.9.0 and <= 1.2.

Current build by default uses Apache Spark 1.1.0. You can also build SparkR against a different Spark version (>= 0.9.0) by modifying pkg/src/build.sbt.

DataFrame: DataFrame was introduced in Spark 1.3; the 1.3-compatible SparkR version can be found in the sparkr-sql branch, which includes a preliminary R API to work with DataFrames.

Package installation

To develop SparkR, you can build the scala package and the R package using

./install-dev.sh

If you wish to try out the package directly from github, you can use install_github from devtools. Note that you can specify which branch, tag etc to install from.

library(devtools)
install_github("amplab-extras/SparkR-pkg", subdir="pkg")

SparkR by default uses Apache Spark 1.1.0. You can switch to a different Spark version by setting the environment variable SPARK_VERSION. For example, to use Apache Spark 1.3.0, you can run

SPARK_VERSION=1.3.0 ./install-dev.sh

SparkR by default links to Hadoop 1.0.4. To use SparkR with other Hadoop versions, you will need to rebuild SparkR with the same version that Spark is linked to. For example to use SparkR with a CDH 4.2.0 MR1 cluster, you can run

SPARK_HADOOP_VERSION=2.0.0-mr1-cdh4.2.0 ./install-dev.sh

By default, SparkR uses sbt to build an assembly jar. If you wish to use maven instead, you can set the environment variable USE_MAVEN=1. For example

USE_MAVEN=1 ./install-dev.sh

If you are building SparkR from behind a proxy, you can setup maven to use the right proxy server.

Building from source from GitHub

Run the following within R to pull source code from GitHub and build locally. It is possible to specify build dependencies by starting R with environment values:

1. Start R

SPARK_VERSION=1.2.0 SPARK_HADOOP_VERSION=2.5.0 R

2. Run install_github

library(devtools)
install_github("repo/SparkR-pkg", ref="branchname", subdir="pkg")

note: replace repo and branchname

Running sparkR

If you have cloned and built SparkR, you can start using it by launching the SparkR shell with

./sparkR

The sparkR script automatically creates a SparkContext with Spark by default in local mode. To specify the Spark master of a cluster for the automatically created SparkContext, you can run

MASTER=<Spark master URL> ./sparkR

If you have installed it directly from github, you can include the SparkR package and then initialize a SparkContext. For example to run with a local Spark master you can launch R and then run

library(SparkR)
sc <- sparkR.init(master="local")

To increase the memory used by the driver you can export the SPARK_MEM environment variable. For example to use 1g, you can run

SPARK_MEM=1g ./sparkR

In a cluster setting to set the amount of memory used by the executors you can pass the variable spark.executor.memory to the SparkContext constructor.

library(SparkR)
sc <- sparkR.init(master="spark://<master>:7077",
                  sparkEnvir=list(spark.executor.memory="1g"))

Finally, to stop the cluster run

sparkR.stop()

sparkR.stop() can be invoked to terminate a SparkContext created previously via sparkR.init(). Then you can call sparkR.init() again to create a new SparkContext that may have different configurations.

Running SparkR Streaming

After starting the SparkR with sparkR script or RStudio or any other your favourite R frontends, you can initialize SparkR streaming with the Spark Context sc and the command

ssc <-sparkR.streaming.init(sc, batchDuration = 1L)

to start a Streaming Context ssc. Then you can create DStreams and apply transformations. You can start streaming by using command

startStreaming(ssc)

For window and state functions, you need to give SparkR streaming a checkpoint directory before starting streaming, by using command

checkpoint(ssc, "/Users/haolin/checkpoints")

Examples, Unit tests

SparkR comes with several sample programs in the examples directory. To run one of them, use ./sparkR <filename> <args>. For example:

./sparkR examples/pi.R local[2]

To run SparkR Streaming wordcount example:

./sparkR examples/streaming/hdfs_wordcount.R /home/haolin/rstreaming/

You can also run the unit-tests for SparkR by running (you need to install the testthat package first):

R -e 'install.packages("testthat", repos="http://cran.us.r-project.org")'
./run-tests.sh

Running on EC2

Instructions for running SparkR on EC2 can be found in the SparkR wiki.

Running on YARN

Currently, SparkR supports running on YARN with the yarn-client mode. These steps show how to build SparkR with YARN support and run SparkR programs on a YARN cluster:

# assumes Java, R, yarn, spark etc. are installed on the whole cluster.
cd SparkR-pkg/
USE_YARN=1 SPARK_YARN_VERSION=2.4.0 SPARK_HADOOP_VERSION=2.4.0 ./install-dev.sh

Alternatively, install_github can be use (on CDH in this case):

# assume devtools package is installed by install.packages("devtools")
USE_YARN=1 SPARK_VERSION=1.1.0 SPARK_YARN_VERSION=2.5.0-cdh5.3.0 SPARK_HADOOP_VERSION=2.5.0-cdh5.3.0 R

Then within R,

library(devtools)
install_github("amplab-extras/SparkR-pkg", ref="master", subdir="pkg")

Before launching an application, make sure each worker node has a local copy of lib/SparkR/sparkr-assembly-0.1.jar. With a cluster launched with the spark-ec2 script, do:

~/spark-ec2/copy-dir ~/SparkR-pkg

Or run the above installation steps on all worker node.

Finally, when launching an application, the environment variable YARN_CONF_DIR needs to be set to the directory which contains the client-side configuration files for the Hadoop cluster (with a cluster launched with spark-ec2, this defaults to /root/ephemeral-hdfs/conf/):

YARN_CONF_DIR=/root/ephemeral-hdfs/conf/ MASTER=yarn-client ./sparkR
YARN_CONF_DIR=/root/ephemeral-hdfs/conf/ ./sparkR examples/pi.R yarn-client

Running on a cluster using sparkR-submit

sparkR-submit is a script introduced to facilitate submission of SparkR jobs to a Spark supported cluster (e.g. Standalone, Mesos, YARN). It supports the same commandline parameters as spark-submit. SPARK_HOME and JAVA_HOME must be defined.

On YARN, YARN_CONF_DIR must be defined. sparkR-submit supports YARN deploy modes: yarn-client and yarn-cluster.

sparkR-submit is installed with the SparkR package. By default, it can be found under the default Library ('library' subdirectory of R_HOME)

For example, to run on YARN (CDH 5.3.0),

export SPARK_HOME=/opt/cloudera/parcels/CDH-5.3.0-1.cdh5.3.0.p0.30/lib/spark
export YARN_CONF_DIR=/etc/hadoop/conf
export JAVA_HOME=/usr/java/jdk1.7.0_67-cloudera
/usr/lib64/R/library/SparkR/sparkR-submit --master yarn-client examples/pi.R yarn-client 4

Report Issues/Feedback

For better tracking and collaboration, issues and TODO items are reported to the Apache Spark JIRA under the component tag "SparkR".

In your pull request, please cross reference the ticket item created and append "[SPARKR]" (e.g.: "[SPARK-1234] [SPARKR] Pull request").