h2oai / sparkling-water

Sparkling Water provides H2O functionality inside Spark cluster

Github

Sparkling Water

Join the chat at https://gitter.im/h2oai/sparkling-water Powered by H2O.ai

Sparkling Water integrates H2O's fast scalable machine learning engine with Spark. It provides:

  • Utilities to publish Spark data structures (RDDs, DataFrames) as H2O's frames and vice versa.
  • DSL to use Spark data structures as input for H2O's algorithms
  • Basic building blocks to create ML applications utilizing Spark and H2O APIs
  • Python interface enabling use of Sparkling Water directly from pySpark

Getting Started

Select right version

The Sparkling Water is developed in multiple parallel branches. Each branch corresponds to a Spark major release (e.g., branch rel-1.5 provides implementation of Sparkling Water for Spark 1.5).

Please, switch to the right branch:

Note Older releases are available here:

Note The master branch includes the latest changes for the latest Spark version. They are back-ported into older Sparkling Water versions.

Requirements

  • Linux/OS X/Windows
  • Java 7+
  • Spark 1.6+
    • SPARK_HOME shell variable must point to your local Spark installation

Build

Download Spark installation and point environment variable SPARK_HOME to it.

Before building this project, you may want to build Spark in case you are using Spark source distribution: go to Spark folder and do sbt assembly.

Then use the provided gradlew to build project:

In order to build the whole project inlucding Python module, one of the following properties needs to be set: H2O_HOME, which should point to location of local h2o project directory, or H2O_PYTHON_WHEEL, which should point to H2O Python Wheel.

If you are not sure which property to set, just run

./gradlew build

and the commands which sets the H2O_PYTHON_WHEEL will be shown on your console and can be copy-pasted into your terminal. After setting the property, the build needs to be rerun.

Now, to have everything you need for Python, you may need to install Python future library, via pip install future. First, you may need to install pip. See http://stackoverflow.com/questions/17271319/installing-pip-on-mac-os-x Use brew package manager or

curl https://bootstrap.pypa.io/ez_setup.py -o - | sudo python
sudo easy_install pip
pip install future

To avoid running tests, use the -x test -x integTest or -x check option.

To build only a specific module, use, for example, ./gradlew :sparkling-water-examples:build. To build and test a specific module, use, for example, ./gradlew :sparkling-water-examples:check.


Download Binaries

For each Sparkling Water you can download binaries here:


Maven

Each Sparkling Water release is published into Maven central. Right now we publish artifacts only for Scala 2.10 and Scala 2.11.

The artifacts coordinates are:

  • ai.h2o:sparkling-water-core_{{scala_version}}:{{version}} - includes core of Sparkling Water.
  • ai.h2o:sparkling-water-examples_{{scala_version}}:{{version}} - includes example applications.

Note: The {{version}} reference to a release version of Sparkling Water, the {{scala_version}} references to Scala base version (2.10 or 2.11). For example: ai.h2o:sparkling-water-examples_2.11:2.0.0

The full list of published packages is available here.


Use Sparkling Water

Sparkling Water is distributed as a Spark application library which can be used by any Spark application. Furthermore, we provide also zip distribution which bundles the library and shell scripts.

There are several ways of using Sparkling Water:

  • Sparkling Shell
  • Sparkling Water driver
  • Spark Shell and include Sparkling Water library via --jars or --packages option
  • Spark Submit and include Sparkling Water library via --jars or --packages option
  • pySpark with pySparkling

Run Sparkling shell

The Sparkling shell encapsulates a regular Spark shell and append Sparkling Water library on the classpath via --jars option. The Sparkling Shell supports creation of an H2O cloud and execution of H2O algorithms.

  1. First, build a package containing Sparkling water:
./gradlew assemble
  1. Configure the location of Spark cluster:
export SPARK_HOME="/path/to/spark/installation"
export MASTER="local[*]"

In this case, local[*] points to an embedded single node cluster.

  1. Run Sparkling Shell:
bin/sparkling-shell

Sparkling Shell accepts common Spark Shell arguments. For example, to increase memory allocated by each executor, use the spark.executor.memory parameter: bin/sparkling-shell --conf "spark.executor.memory=4g"

  1. Initialize H2OContext
import org.apache.spark.h2o._
val hc = H2OContext.getOrCreate(sparkSession)

H2OContext start H2O services on top of Spark cluster and provides primitives for transformations between H2O and Spark datastructures.


Run examples

The Sparkling Water distribution includes also a set of examples. You can find there implementation in example folder. You can run them in the following way:

  1. Build a package that can be submitted to Spark cluster:
./gradlew assemble
  1. Set the configuration of the demo Spark cluster (for example, local[*] or local-cluster[3,2,1024])
export SPARK_HOME="/path/to/spark/installation"
export MASTER="local[*]"

In this example, the description local[*] causes creation of a single node local cluster.

  1. And run the example:
bin/run-example.sh

For more details about examples, please see the README.md file in the examples directory.

Additional Examples

You can find more examples in the examples folder.


Run PySparkling

Sparkling Water can be also used directly from PySpark.

See py/README.md to learn about PySparkling.


Use Sparkling Water via Spark Packages

Sparkling Water is also published as a Spark package. You can use it directly from your Spark distribution.

For example, if you have Spark version 2.0 and would like to use Sparkling Water version 2.0.0 and launch example CraigslistJobTitlesStreamingApp, then you can use the following command:

$SPARK_HOME/bin/spark-submit --packages ai.h2o:sparkling-water-core_2.11:2.0.0,ai.h2o:sparkling-water-examples_2.11:2.0.0 --class org.apache.spark.examples.h2o.CraigslistJobTitlesStreamingApp /dev/null

The Spark option --packages points to published Sparkling Water packages in Maven repository.

The similar command works for spark-shell:

$SPARK_HOME/bin/spark-shell --packages ai.h2o:sparkling-water-core_2.11:2.0.0,ai.h2o:sparkling-water-examples_2.11:2.0.0

The same command works for Python programs:

$SPARK_HOME/bin/spark-submit --packages ai.h2o:sparkling-water-core_2.11:2.0.0,ai.h2o:sparkling-water-examples_2.11:2.0.0 example.py

Note: When you are using Spark packages you do not need to download Sparkling Water distribution! Spark installation is sufficient!


Docker Support

See docker/README.md to learn about Docker support.


Use Sparkling Water in Windows environments

The Windows environments require several additional steps to make Spark and later Sparkling Water working. Great summary of configuration steps is here.

On Windows it is required:

  1. Download Spark distribution

  2. Setup variable SPARK_HOME:

SET SPARK_HOME=<location of your downloaded Spark distribution>
  1. From https://github.com/steveloughran/winutils, download winutils.exe for Hadoop version which is referenced by your Spark distribution (for example, for spark-2.1.0-bin-hadoop2.6.tgz you need wintutils.exe for hadoop2.6).

  2. Put winutils.exe into a new directory %SPARK_HOME%\hadoop\bin and set:

SET HADOOP_HOME=%SPARK_HOME%\hadoop
  1. Create a new file %SPARK_HOME%\hadoop\conf\hive-site.xml which setup default Hive scratch dir. The best location is a writable temporary directory, for example %TEMP%\hive:
<configuration>
  <property>
    <name>hive.exec.scratchdir</name>
    <value>PUT HERE LOCATION OF TEMP FOLDER</value>
    <description>Scratch space for Hive jobs</description>
  </property>
</configuration>

Note: you can also use Hive default scratch directory which is /tmp/hive. In this case, you need to create directory manually and call winutils.exe chmod 777 \tmp\hive to setup right permissions.

  1. Set HADOOP_CONF_DIR property
SET HADOOP_CONF_DIR=%SPARK_HOME%\hadoop\conf
  1. Run Sparkling Water as described above.

Sparkling Water cluster backends

Sparkling water supports two backend/deployment modes. We call them internal and external back-ends. Sparkling Water applications are independent on selected backend, the before H2OContext is created we need to tell it which backend used.

For more details regarding the internal or external backend, please see doc/backends.md.


Develop with Sparkling Water

Setup Sparkling Water in IntelliJ IDEA

  • In IDEA, install the Scala plugin for IDEA
  • In a Terminal:
git clone https://github.com/h2oai/sparkling-water.git
cd sparkling-water
./gradlew idea
open sparkling-water.ipr
  • In IDEA, open the file sparkling-water/core/src/main/scala/water/SparklingWaterDriver.scala
  • [ Wait for IDEA indexing to complete so the Run and Debug choices are available ]
  • In IDEA, Run or Debug SparklingWaterDriver (via right-click)

Develop applications with Sparkling Water

An application using Sparkling Water is regular Spark application which bundling Sparkling Water library. See Sparkling Water Droplet providing an example application here.


Contributing

Look at our list of JIRA tasks for new contributors or send your idea to support@h2o.ai.


Issues

To report issues, please use our JIRA page at http://jira.h2o.ai/.


Mailing list

Follow our H2O Stream.


FAQ

  • Where do I find the Spark logs?

Standalone mode: Spark executor logs are located in the directory $SPARK_HOME/work/app-<AppName> (where <AppName> is the name of your application). The location contains also stdout/stderr from H2O.

YARN mode: The executors logs are available via yarn logs -applicationId <appId> command. Driver logs are by default printed to console, however, H2O also writes logs into current_dir/h2ologs.

The location of H2O driver logs can be controlled via Spark property spark.ext.h2o.client.log.dir (pass via --conf) option.

  • Spark is very slow during initialization or H2O does not form a cluster. What should I do?

Configure the Spark variable SPARK_LOCAL_IP. For example:

export SPARK_LOCAL_IP='127.0.0.1'
  • How do I increase the amount of memory assigned to the Spark executors in Sparkling Shell?

Sparkling Shell accepts common Spark Shell arguments. For example, to increase the amount of memory allocated by each executor, use the spark.executor.memory parameter: bin/sparkling-shell --conf "spark.executor.memory=4g"

  • How do I change the base port H2O uses to find available ports?

    The H2O accepts spark.ext.h2o.port.base parameter via Spark configuration properties: bin/sparkling-shell --conf "spark.ext.h2o.port.base=13431". For a complete list of configuration options, refer to Devel Documentation.

  • How do I use Sparkling Shell to launch a Scala test.script that I created?

Sparkling Shell accepts common Spark Shell arguments. To pass your script, please use -i option of Spark Shell: bin/sparkling-shell -i test.script

  • How do I increase PermGen size for Spark driver?

Specify --conf spark.driver.extraJavaOptions="-XX:MaxPermSize=384m"

  • How do I add Apache Spark classes to Python path?

Configure the Python path variable PYTHONPATH:

export PYTHONPATH=$SPARK_HOME/python:$SPARK_HOME/python/build:$PYTHONPATH
export PYTHONPATH=$SPARK_HOME/python/lib/py4j-0.9-src.zip:$PYTHONPATH
  • Trying to import a class from the hex package in Sparkling Shell but getting weird error:

    error: missing arguments for method hex in object functions;
    follow this method with '_' if you want to treat it as a partially applied
    

    In this case you are probably using Spark 1.5 which is importing SQL functions into Spark Shell environment. Please use the following syntax to import a class from the hex package:

    import _root_.hex.tree.gbm.GBM
    
  • Trying to run Sparkling Water on HDP Yarn cluster, but getting error:

    java.lang.NoClassDefFoundError: com/sun/jersey/api/client/config/ClientConfig
    

    The Yarn time service is not compatible with libraries provided by Spark. Please disable time service via setting spark.hadoop.yarn.timeline-service.enabled=false. For more details, please visit https://issues.apache.org/jira/browse/SPARK-15343

  • Getting non-deterministic H2O Frames after the Spark Data Frame to H2O Frame conversion.

    This is caused by what we think is a bug in Apache Spark. On specific kinds of data combined with higher number of partitions we can see non-determinism in BroadCastHashJoins. This leads to to jumbled rows and columns in the output H2O frame. We recommend to disable broadcast based joins which seem to be non-deterministic as:

    sqlContext.sql(\"SET spark.sql.autoBroadcastJoinThreshold=-1\")
    

    The issue can be tracked as PUBDEV-3808. On the Spark side, the following issues are related to the problem: Spark-17806