Integrating Posit Workbench and Jupyter with PySpark#
This documentation describes the steps to use Posit Workbench, formerly RStudio Workbench, to connect to a Spark cluster using Jupyter Notebooks and PySpark.
In this example, Apache Hadoop YARN is used as a resource manager on the Spark cluster, and you'll create interactive Python sessions that use PySpark.
- Workbench configured with Jupyter Notebooks on a Single Server
- Hadoop cluster configured with Spark and YARN
- Access from Workbench to the Spark cluster
The Workbench server must have access to the Spark cluster and the underlying configuration files for YARN and the Hadoop Distributed File System (HDFS). This typically requires you to install Workbench on an edge node (i.e., gateway node) of the Spark cluster using a Hadoop administration tool such as Cloudera Manager or Apache Ambari. You can also achieve this by copying configuration files from the Spark cluster to the Workbench server.
Add Workbench as an edge/gateway node#
This section describes the process to add a single node as a Spark client to a Hadoop cluster. This step is typically performed by a Hadoop administrator. In this example, we use Cloudera Manager, but these steps can also be adapted to other Spark clusters such as Amazon EMR.
This process may vary depending on different versions of Cloudera Distribution Hadoop (CDH), authentication, and other variables. Refer to the Cloudera Manager documentation for more information.
Step 1: Add a new host to the Hadoop cluster#
Since you already have a server with Workbench installed, the first step is to add the Workbench node to an existing Cloudera CDH cluster.
From the Cloudera Manager dashboard, select the drop-down menu and click Add Hosts.
Next, click the Classic Wizard on the first page.
Step 2: Specify the hostname for the Workbench node#
Continue with the installation until it asks you to specify the hostname of the node to add. Add the hostname of the Workbench node.
This hostname should be accessible from the Cloudera CDH cluster, Cloudera Manager will verify this when you click Search.
When everything is verified click Continue.
Step 3: Specify the credentials for the Workbench node#
- Continue with the installation wizard until it asks for the login credentials for the Workbench node.
In this example, we are using an Amazon EC2 instance with the username
and authentication via a private key. You might be using a different
authentication/credential mechanism depending on how you access the Workbench node.
Step 4: Wait for the Cloudera Manager agent and parcels to be installed#
If the hostname and credentials are correct, Cloudera Manager installs the Cloudera Manager Agent on the Workbench node.
The Cloudera CDH parcels will then be installed on the node (this might take a couple of minutes).
Step 5: Verify that the new host with Workbench has been added#
- Continue with the installation, and Cloudera Manager will inspect the hosts. If everything installed correctly, then the Workbench node will join the Hadoop cluster.
- Verify that the Workbench node appears in the list of hosts. Initially, this node will not have any roles, but you will add the necessary roles in the following step.
Step 6: Add roles to the Workbench node#
You can now add roles to the Workbench node. The roles that you will need to add are listed as follows:
- HDFS Gateway
- YARN Gateway
- Hive Gateway
- Spark Gateway
- Spark2 Gateway (if your Cloudera CDH cluster has Spark2 installed)
The following steps show an example of how to add the
Spark Gateway role to
the Workbench node in Cloudera Manager. You can then repeat this
process for all of the necessary roles.
Navigate to the Cloudera Manager Home page, select the tab of the service that you want to add and click Add Role Instances.
Under the Gateway option, click Select hosts.
Select the Workbench node from the list of nodes.
You should now see the Workbench node selected under the Gateway option.
Be sure to verify the hostname of the Workbench node.
Follow the steps in the wizard and then re-deploy the client configuration by clicking the Deploy button.
You can then repeat this process to add all of the necessary roles that are listed above.
After you've added all of the necessary roles, the cluster roles for the Workbench node should look similar to the following figure.
Step 7: Verify that users exist on the Hadoop cluster and HDFS#
It's important that the same users that log into Workbench also exist within the Hadoop cluster. This is because the RStudio/Jupyter sessions will run as that user, and any Spark contexts will inherit the YARN and HDFS permissions of that user.
Synchronizing users across your Workbench instance and your Hadoop cluster can be accomplished using multiple approaches. For example, both systems might be configured to the same identity provider via LDAP/AD. For more information, you can discuss this more with your Hadoop administrator.
To manually create a user in HDFS, you can run the following command (replace
<rstudio>with the actual username):Terminal
$ hdfs dfs -mkdir /user/<rstudio> $ hdfs dfs -chown rstudio:rstudio /user/rstudio/
Step 8: Verify network connectivity between Workbench and the Hadoop cluster#
- Ensure that the Workbench node has network access to the Cloudera CDH cluster. In Amazon AWS, we recommend allowing all communication between the Cloudera CDH security group and the Workbench security group.
Using Workbench with Jupyter and PySpark#
Now that Workbench is a member of the Hadoop/Spark cluster, you can install and configure PySpark to work on Workbench Jupyter sessions.
This section describes the process for a user to work with Workbench and Jupyter Notebooks to connect to the Spark cluster via PySpark.
Step 1: Install PySpark in the Python environment#
Install PySpark in the environments that are configured as Python kernels, for example:Terminal
sudo /opt/python/2.7.16/bin/pip install pyspark
PySpark cannot run with different minor versions of Python installed, be sure to use the same version of Python in Workbench and the Spark cluster.
Step 2: Configure environment variables for Spark#
To configure the Spark environment variables for all Jupyter sessions, create a file under
/etc/profile.d/that exports the required configuration variables, for example:File: /etc/profile.d/spark.sh
export JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-22.214.171.124.b10-1.el7_7.x86_64/jre export HADOOP_CONF_DIR=/etc/hadoop/conf
Java versions and the JAVA_HOME variable
Ensure that you export a
JAVA_HOMEvariable that matches the Java version that PySpark was compiled with. In this example, we are using Java Version 8.
Step 3: Create a Spark session via PySpark#
Now you are ready to create a Spark session and connect to Spark.
From the Workbench home page, create a new Jupyter Notebook or JupyterLab session.
pysparkand create a new Spark session that uses YARN by running the following Python code in the notebook:Python Code
from pyspark import SparkConf from pyspark import SparkContext conf = SparkConf() conf.setMaster('yarn-client') conf.setAppName('rstudio-pyspark') sc = SparkContext(conf=conf)
Step 4: Verify that the Spark application is running in YARN#
At this point, you should be able to see that the Spark application is running in the YARN resource manager.
Step 5: Run a sample computation#
You can run the following sample code in the notebook to verify that the Spark connectivity is working as expected:Python Code
data = [1, 2, 3, 4, 5] distData = sc.parallelize(data) distData.mean()
Step 6: Verify read/write operations to HDFS#
You can run the following sample code in the notebook to verify that writes to HDFS are working as expected:Python Code
# Save a file to HDFS rdd = sc.parallelize(range(1, 4)).map(lambda x: (x, "a" * x)) rdd.saveAsSequenceFile("saved_file")
You can run the following sample code in the notebook to verify that reads from HDFS are working as expected:Python Code
# Read the same file from HDFS sorted(sc.sequenceFile("saved_file").collect())