Spark Interpreter for Apache Zeppelin

Overview

Apache Spark is a fast and general-purpose cluster computing system. It provides high-level APIs in Java, Scala, Python and R, and an optimized engine that supports general execution graphs. Apache Spark is supported in Zeppelin with Spark interpreter group which consists of following interpreters.

Main Features

Play Spark in Zeppelin docker

For beginner, we would suggest you to play Spark in Zeppelin docker. In the Zeppelin docker image, we have already installed miniconda and lots of useful python and R libraries including IPython and IRkernel prerequisites, so %spark.pyspark would use IPython and %spark.ir is enabled. Without any extra configuration, you can run most of tutorial notes under folder Spark Tutorial directly.

First you need to download Spark, because there's no Spark binary distribution shipped with Zeppelin. e.g. Here we download Spark 3.1.2 to/mnt/disk1/spark-3.1.2, and we mount it to Zeppelin docker container and run the following command to start Zeppelin docker container.

docker run -u $(id -u) -p 8080:8080 -p 4040:4040 --rm -v /mnt/disk1/spark-3.1.2:/opt/spark -e SPARK_HOME=/opt/spark  --name zeppelin apache/zeppelin:0.10.0

After running the above command, you can open http://localhost:8080 to play Spark in Zeppelin. We only verify the spark local mode in Zeppelin docker, other modes may not work due to network issues. -p 4040:4040 is to expose Spark web ui, so that you can access Spark web ui via http://localhost:8081.

Configuration

The Spark interpreter can be configured with properties provided by Zeppelin. You can also set other Spark properties which are not listed in the table. For a list of additional properties, refer to Spark Available Properties.

Without any configuration, Spark interpreter works out of box in local mode. But if you want to connect to your Spark cluster, you'll need to follow below two simple steps.

• Set SPARK_HOME
• Set master

Set SPARK_HOME

There are several options for setting SPARK_HOME.

• Set SPARK_HOME in zeppelin-env.sh
• Set SPARK_HOME in interpreter setting page
• Set SPARK_HOME via inline generic configuration

Set SPARK_HOME in zeppelin-env.sh

If you work with only one version of Spark, then you can set SPARK_HOME in zeppelin-env.sh because any setting in zeppelin-env.sh is globally applied.

e.g.

export SPARK_HOME=/usr/lib/spark

You can optionally set more environment variables in zeppelin-env.sh

# set hadoop conf dir
export HADOOP_CONF_DIR=/usr/lib/hadoop

Set SPARK_HOME in interpreter setting page

If you want to use multiple versions of Spark, then you need to create multiple Spark interpreters and set SPARK_HOME separately. e.g. Create a new Spark interpreter spark24 for Spark 2.4 and set its SPARK_HOME in interpreter setting page as following,

Create a new Spark interpreter spark16 for Spark 1.6 and set its SPARK_HOME in interpreter setting page as following,

Set SPARK_HOME via inline generic configuration

Besides setting SPARK_HOME in interpreter setting page, you can also use inline generic configuration to put the configuration with code together for more flexibility. e.g.

Set master

After setting SPARK_HOME, you need to set spark.master property in either interpreter setting page or inline configuartion. The value may vary depending on your Spark cluster deployment type.

For example,

• local[*] in local mode
• spark://master:7077 in standalone cluster
• yarn-client in Yarn client mode (Not supported in Spark 3.x, refer below for how to configure yarn-client in Spark 3.x)
• yarn-cluster in Yarn cluster mode (Not supported in Spark 3.x, refer below for how to configure yarn-cluster in Spark 3.x)
• mesos://host:5050 in Mesos cluster

That's it. Zeppelin will work with any version of Spark and any deployment type without rebuilding Zeppelin in this way. For the further information about Spark & Zeppelin version compatibility, please refer to "Available Interpreters" section in Zeppelin download page.

Note that without setting SPARK_HOME, it's running in local mode with included version of Spark. The included version may vary depending on the build profile. And this included version Spark has limited function, so it is always recommended to set SPARK_HOME.

Yarn client mode and local mode will run driver in the same machine with zeppelin server, this would be dangerous for production. Because it may run out of memory when there's many Spark interpreters running at the same time. So we suggest you only allow yarn-cluster mode via setting zeppelin.spark.only_yarn_cluster in zeppelin-site.xml.

Configure yarn mode for Spark 3.x

Specifying yarn-client & yarn-cluster in spark.master is not supported in Spark 3.x any more, instead you need to use spark.master and spark.submit.deployMode together.

Interpreter binding mode

The default interpreter binding mode is globally shared. That means all notes share the same Spark interpreter.

So we recommend you to use isolated per note which means each note has own Spark interpreter without affecting each other. But it may run out of your machine resource if too many Spark interpreters are created, so we recommend to always use yarn-cluster mode in production if you run Spark in hadoop cluster. And you can use inline configuration via %spark.conf in the first paragraph to customize your spark configuration.

You can also choose scoped mode. For scoped per note mode, Zeppelin creates separated scala compiler/python shell for each note but share a single SparkContext/SqlContext/SparkSession.

SparkContext, SQLContext, SparkSession, ZeppelinContext

SparkContext, SQLContext, SparkSession (for spark 2.x, 3.x) and ZeppelinContext are automatically created and exposed as variable names sc, sqlContext, spark and z respectively, in Scala, Kotlin, Python and R environments.

Note that Scala/Python/R environment shares the same SparkContext, SQLContext, SparkSession and ZeppelinContext instance.

Yarn Mode

Zeppelin support both yarn client and yarn cluster mode (yarn cluster mode is supported from 0.8.0). For yarn mode, you must specify SPARK_HOME & HADOOP_CONF_DIR. Usually you only have one hadoop cluster, so you can set HADOOP_CONF_DIR in zeppelin-env.sh which is applied to all Spark interpreters. If you want to use spark against multiple hadoop cluster, then you need to define HADOOP_CONF_DIR in interpreter setting or via inline generic configuration.

K8s Mode

Regarding how to run Spark on K8s in Zeppelin, please check this doc.

PySpark

There are 2 ways to use PySpark in Zeppelin:

• Vanilla PySpark
• IPySpark

Vanilla PySpark (Not Recommended)

Vanilla PySpark interpreter is almost the same as vanilla Python interpreter except Spark interpreter inject SparkContext, SQLContext, SparkSession via variables sc, sqlContext, spark.

By default, Zeppelin would use IPython in %spark.pyspark when IPython is available (Zeppelin would check whether ipython's prerequisites are met), Otherwise it would fall back to the vanilla PySpark implementation.

IPySpark (Recommended)

You can use IPySpark explicitly via %spark.ipyspark. IPySpark interpreter is almost the same as IPython interpreter except Spark interpreter inject SparkContext, SQLContext, SparkSession via variables sc, sqlContext, spark. For the IPython features, you can refer doc Python Interpreter

SparkR

Zeppelin support SparkR via %spark.r, %spark.ir and %spark.shiny. Here's configuration for SparkR Interpreter.

Refer R doc for how to use R in Zeppelin.

SparkSql

Spark sql interpreter share the same SparkContext/SparkSession with other Spark interpreters. That means any table registered in scala, python or r code can be accessed by Spark sql. For examples:

%spark

case class People(name: String, age: Int)
var df = spark.createDataFrame(List(People("jeff", 23), People("andy", 20)))
df.createOrReplaceTempView("people")

%spark.sql

select * from people

You can write multiple sql statements in one paragraph. Each sql statement is separated by semicolon. Sql statement in one paragraph would run sequentially. But sql statements in different paragraphs can run concurrently by the following configuration.

1. Set zeppelin.spark.concurrentSQL to true to enable the sql concurrent feature, underneath zeppelin will change to use fairscheduler for Spark. And also set zeppelin.spark.concurrentSQL.max to control the max number of sql statements running concurrently.
2. Configure pools by creating fairscheduler.xml under your SPARK_CONF_DIR, check the official spark doc Configuring Pool Properties

3. Set pool property via setting paragraph local property. e.g.

   %spark(pool=pool1)
   
   sql statement
   

This pool feature is also available for all versions of scala Spark, PySpark. For SparkR, it is only available starting from 2.3.0.

Dependency Management

For Spark interpreter, it is not recommended to use Zeppelin's Dependency Management for managing third party dependencies (%spark.dep is removed from Zeppelin 0.9 as well). Instead, you should set the standard Spark properties as following:

As general Spark properties, you can set them in via inline configuration or interpreter setting page or in zeppelin-env.sh via environment variable SPARK_SUBMIT_OPTIONS. For examples:

export SPARK_SUBMIT_OPTIONS="--files <my_file> --jars <my_jar> --packages <my_package>"

To be noticed, SPARK_SUBMIT_OPTIONS is deprecated and will be removed in future release.

ZeppelinContext

Zeppelin automatically injects ZeppelinContext as variable z in your Scala/Python environment. ZeppelinContext provides some additional functions and utilities. See Zeppelin-Context for more details. For Spark interpreter, you can use z to display Spark Dataset/Dataframe.

Setting up Zeppelin with Kerberos

Logical setup with Zeppelin, Kerberos Key Distribution Center (KDC), and Spark on YARN:

There are several ways to make Spark work with kerberos enabled hadoop cluster in Zeppelin.

1. Share one single hadoop cluster. In this case you just need to specify zeppelin.server.kerberos.keytab and zeppelin.server.kerberos.principal in zeppelin-site.xml, Spark interpreter will use these setting by default.

2. Work with multiple hadoop clusters. In this case you can specify spark.yarn.keytab and spark.yarn.principal to override zeppelin.server.kerberos.keytab and zeppelin.server.kerberos.principal.

Configuration Setup

1. On the server that Zeppelin is installed, install Kerberos client modules and configuration, krb5.conf. This is to make the server communicate with KDC.

2. Add the two properties below to Spark configuration ([SPARK_HOME]/conf/spark-defaults.conf):

   spark.yarn.principal
   spark.yarn.keytab
   

NOTE: If you do not have permission to access for the above spark-defaults.conf file, optionally, you can add the above lines to the Spark Interpreter setting through the Interpreter tab in the Zeppelin UI.

1. That's it. Play with Zeppelin!

User Impersonation

In yarn mode, the user who launch the zeppelin server will be used to launch the Spark yarn application. This is not a good practise. Most of time, you will enable shiro in Zeppelin and would like to use the login user to submit the Spark yarn app. For this purpose, you need to enable user impersonation for more security control. In order the enable user impersonation, you need to do the following steps

Step 1 Enable user impersonation setting hadoop's core-site.xml. E.g. if you are using user zeppelin to launch Zeppelin, then add the following to core-site.xml, then restart both hdfs and yarn.

<property>
  <name>hadoop.proxyuser.zeppelin.groups</name>
  <value>*</value>
</property>
<property>
  <name>hadoop.proxyuser.zeppelin.hosts</name>
  <value>*</value>
</property>

Step 2 Enable interpreter user impersonation in Spark interpreter's interpreter setting. (Enable shiro first of course)

Step 3(Optional) If you are using kerberos cluster, then you need to set zeppelin.server.kerberos.keytab and zeppelin.server.kerberos.principal to the user(aka. user in Step 1) you want to impersonate in zeppelin-site.xml.

Deprecate Spark 2.2 and earlier versions

Starting from 0.9, Zeppelin deprecate Spark 2.2 and earlier versions. So you will see a warning message when you use Spark 2.2 and earlier. You can get rid of this message by setting zeppelin.spark.deprecatedMsg.show to false.

Community

Join our community to discuss with others.