Setup Tableau with Amazon EMR-Spark

Target audience: Intermediate
Estimated reading time: 3'

Posts history

This article describes the configuration of Amazon EMR service and set up of Tableau desktop to query and visualize Spark SQL datasets and content stored on S3. The installation process can daunting as documentation and useful tips are spread across various sites, chat rooms.

Table of contents

Overview

Setup

References

Overview

Tableau is a popular, powerful visualization platform that leverages a large variety of data sources from files, databases, applications to frameworks such as Apache Spark. Tableau is particularly suitable to visualize the results of queries to a Spark dataset.
Tableau desktop is a charts/query builder that relies on simple drag and drop to render results of query. It support a very large variety (50+) data source connectors from files, databases, CRMs, enterprise and cloud applications to scalable frameworks such as Apache Spark. Tableau's powerful statistical and computational capabilities help data scientists and product managers to spot critical data patterns.
There are few options to query and visualize Apache Spark datasets in Tableau:

1. Using an ODBC driver to access Spark data-frame using the Spark SQL connector
2. Through Hive2 thrift server using the Amazon EMR Hadoop Hive connector

For this post we select the second option and describe a common use case: installation and configuration of Thrift server, loading data from S3, transformation applied to a Spark dataset and leveraging parquet format.

Setup

Starting configuration

We assume that the data has been previously stored on Amazon S3. The same procedure would apply to any other storage such as HDFS, database or local file. The other assumptions is that Apache Spark has been deployed through an Amazon EMR.

Steps

1- Download Tableau Desktop Tableau Desktop

2- Download Simba ODBC driver for MacOS for Amazon EMR Hadoop Hive connector Driver Download 

3- Implement processing of the dataset loaded from S3 in CSV format, and Hive table using parquet. The procedure (Scala code snippet below) can be also easily implemented using Python/PySpark.

1. final private val dataDir = "/tmp/records_table"
   final private val s3Bucket = "...."
   final private val s3CSVFile = "...."
   
   final val createTable = "CREATE EXTERNAL TABLE records_table(" +
    "id varchar(64), " +
    "description varchar(512), " +
    "value float, " +
    "unit char(8) " +
    "STORED AS PARQUET LOCATION"
   
   
   def loadRecords(implicit sparkSession: SparkSession): Unit = {
     import sparkSession.implicits._
     try {
       // Load dataframe from CSV file
       val recordsDF  = s3CSVToDataFrame(s3CSVFile , s3Bucket, true, true)
       // Convert data frame into a typed data set
       val recordsDS = recordsDF.map(Record(_))
   
       // Generate and store data in Parquet columnar structure
       recordDS.write.mode(SaveMode.Overwrite).parquet(dataDir)
   
       // Create the HIVE table pre-populated from Parquet structure stored on HDFS
       sparkSession.sql("DROP TABLE IF EXISTS records_table")
       sparkSession.sql(s"${createTable} '$dataDir'")
       logger.info(s"New table for $dataDir was created")
   
       // Just a quick validation test
       sparkSession.sql("SELECT id, value FROM records_table")

                   .show
       sparkSession.close
     }
     catch {
       case e: Exception => 
         logger.error(s"Failed to create HIVE table ${e.toString}")
     }
   }
   
   
   
   @throws(clazz = classOf[IllegalStateException])
   def s3CSVToDataFrame(
        s3CSVInputFile: String, 
        header: Boolean,
        s3Bucket: String,
        isMultiLine: Boolean
   )(implicit sparkSession: SparkSession): DataFrame = {
    import sparkSession.implicits._
    
    // Initialize the access configuration for Hadoop
    val loadDS = Seq[String]().toDS
    val accessConfig = loadDS.sparkSession.sparkContext.hadoopConfiguration
   
    try {
      accessConfig.set("fs.s3a.access.key", "xxxxxx")
      accessConfig.set("fs.s3a.secret.key", "xxxxxxx")
      
      val headerStr = if (header) "true" else "false"
      // Read the content of the CSV file from S3 to generate a data frame
      sparkSession
            .read
            .format("csv")
            .option("header", headerStr)
            .option("delimiter", ",")
            .option("multiLine", isMultiLine)
            .load(path = s"s3a://${s3Bucket}/${s3CSVInputFile}")
     } catch {
        case e: FileNotFoundException =>  

            throw new IllegalStateException(e.getMessage)
        case e: SparkException =>  

            throw new IllegalStateException(e.getMessage)
        case e: Exception => 

            throw new IllegalStateException(e.getMessage)
     }
   }
   

4- Log into the target EMR instance and upload the jar file for execution

5- Add or edit few spark configuration parameters 

• spark.sql.warehouse.dir=/hive/spark-warehouse 
• spark.hadoop.hive.metastore.warehouse.dir=/hive/spark-warehouse 
• spark.sql.hive.server2.thrift.port=10000 
• spark.sql.hive.thriftServer.singleSession=true

6- Execute the code to generate the Hive table from Spark dataset  

7- Set up/edit HIVE configuration file /usr/lib/spark/conf/hive-site.xml as a super user (sudo). 

 

Note: The default Derby embedded driver is used for convenience but can be easily replaced by a MySql or PostgreSQL driver by updating the javax.jdo.option.ConnectionURL and javax.jdo.option.ConnectionDriverName values. 

The default port for the thrift server is 10000 (hive.server2.thrift.port) may have to be changed to avoid conflict with other services.

<configuration>
    <property>
        <name>hive.metastore.connect.retries</name>
        <value>10</value>
    </property>
    <property>
        <name>javax.jdo.option.ConnectionURL</name>
        <value>jdbc:derby:;databaseName=metastore_db;create=true</value>
    </property>
    <property>
        <name>javax.jdo.option.ConnectionDriverName</name>
        <value>org.apache.derby.jdbc.EmbeddedDriver</value>
    </property>
    <property>
        <name>hive.metastore.warehouse.dir</name>
        <value>~/hive/warehouse</value>
    </property>
    <property>
        <name>hive.server2.authentication</name>
        <value>NONE</value>
    </property>
    <property>
        <name>hive.server2.thrift.client.user</name>
        <value>root</value>
    </property>
    <property>
        <name>hive.server2.thrift.client.password</name>
        <value>xxxxxx</value>
    </property>
    <property>
        <name>hive.server2.thrift.port</name>   
        <value>10000</value>
    </property>
    <property>
        <name>hive.security.authorization.enabled</name>
        <value>true</value>
     </property>
    <property>
        <name>javax.jdo.option.ConnectionUserName</name>
        <value>hive</value>
    </property>
    <property>
        <name>javax.jdo.option.ConnectionPassword</name>
        <value>xxxxx</value>
    </property>
    <property>
        <name>hive.exec.local.scratchdir</name>
        <value>~/tmp/hive</value>
    </property>
</configuration>

Note: The configuration variables defined in the spark configuration file overrides some of these entries. 

• Remove potential locks in the metastore: rm -r metastore/*.lck (Locked access to the store will generate an error accessing and reading the table)
• Stop the Hive2 thrift server sudo /usr/lib/spark/sbin/stop-thriftserver.sh 
• Optionally kill the 2 processes related to thrift server ps -ef | grep RunJar sudo kill -9 {processId} 
• Restart the thrift server sudo /usr/lib/spark/sbin/start-thriftserver.sh --master local 
• Verify the parquet data is correctly stored on HDFS hdfs dfs -ls /tmp/metrics/ 
• Verify table is created and populated in the EMR instance hive => show tables
• Launch Tableau desktop
• Select Amazon EMR-Hadoop connector
• Configure the connection through UI (see attached snapshot): 1) Enter the public DNS URL for the EMR master instance, 2) Select authentication = user name 3) Enter user name = hadoop 4) Select SSL required.

Thank you for reading this article. For more information ...

References

• Tableau
• Tableau Desktop
• Augmented Analytics
• Amazon AWS
• Amazon EMR-Spark
• github.com/patnicolas

---------------------------

Patrick Nicolas has over 25 years of experience in software and data engineering, architecture design and end-to-end deployment and support with extensive knowledge in machine learning. 
He has been director of data engineering at Aideo Technologies since 2017 and he is the author of "Scala for Machine Learning" Packt Publishing ISBN 978-1-78712-238-3

Handling Scala 2.x Option Elegantly

Target audience: Beginner

Estimated reading time: 3'

Posts history

This post reviews the different alternative mechanisms in Scala to handle errors. It also illustrates the applicability of the Option monad.

Table of contents

Use case

Handling Option values

for comprehension for options

References

The Option monad is a great tool for handling error, in Scala: developers do not have worry about NullPointerException or handling a typed exception as in Java and C++.

Notes:

• For the sake of readability of the implementation of algorithms, all non-essential code such as error checking, comments, exception, validation of class and method arguments, scoping qualifiers or import is omitted 
• The code associated with this article is written using Scala 2.12.4

Use case

Let's consider the simple square root computation, which throws an exception if the input value is strictly negative (line 2). The most common "java-like" approach is to wrap the computation with a try - catch paradigm. In Scala catching exception can be implemented through the Try monad (lines 7-9).

1 2 3 4 5 6 7 8 9 10

def sqrt(x: Double): Double = if(x < 0.0) throw MathException(s"sqrt: Incorrect argument $x") else Math.sqrt(x) Try ( sqrt(a)) match { case Success(x) => {} case Failure(e) => Console.println(e.toString) }

This type of implementation put the burden on the client code to handle the exception. The Option monad provides developer an elegant to control the computation flow.

Handling Option values

Clearly, there is no need to compute y if x is negative.

The most common to handle a Scala option is to unwrap it. Let's consider the function      y = sin(sqrt(x))

def sqrt(x: Double): Option[Double] = {
    if(x < 0.0) None
    else Math.sqrt(x)
}

 
 
def y(x: Double): Option[Double] = sqrt(x) match {
    case Some(y) => Math.sin(x)
    case None => None
}

The computation of the square root is implemented by the method sqrt while the final computation of sin(sqrt(x)) is defined by the method y.

This implementation is quite cumbersome because the client code has to process an extra Option. An alternative is to provide a default value (i.e 0.0) if the first computational step fails.

def y(x: Double): Double = Math.sin(sqrt(x)).getOrElse(0.0)

A more functional and elegant approach uses the map higher order function to propagate the value of the Option.

def y(x: Double): Double =  sqrt(x).map(Math.sin(_)).getOrElse(0.0)

What about a sequence of nested options? Let's consider the function y = 1/sqrt(x). There are two types of errors:

• x < 0.0 for sqrt
• x == 0.0 for 1/x

A third solution consist of applying the test for x > 0.0 to meet the two conditions at once.

def y(xdef y(x: Double): Double = 
    if(x < 1e-30) None  else Some(1.0/(Math.sqrt(x)))

for comprehension for options

However anticipating the multiple complex conditions on the argument is not always possible. The for comprehensive for loop is an elegant approach to handle sequence of options.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

def inv(x: Double): Option[Double] = { if(Math.abs(x) < 1e-30) None else 1.0/x } def log(x: Double): Option[Double] = { if(x < 1e-30) None else Math.log(x) } def compose(x: Double): Double = (for { y <- sqrt(x) z <- inv(y) t <- log(z) } yield t).getOrElse(0.0)

The objective is to compose the computation of a square root with the inverse function inv (line 1) and natural logarithm log (line 6). The for comprehension construct (lines 11-15) propagates the result of each function to the next in the pipeline through the automatic conversion of option to its value. In case of error (None), the for method exists before completion.
For-comprehension is a monad that compose (cascading) multiple flatMap with a final map method.

Thank you for reading this article. For more information ...

References

• Programming in Scala M. Odersky, L. Spoon, B. Venners - Artima 2010
• github.com/patnicolas

---------------------------

Patrick Nicolas has over 25 years of experience in software and data engineering, architecture design and end-to-end deployment and support with extensive knowledge in machine learning. 
He has been director of data engineering at Aideo Technologies since 2017 and he is the author of "Scala for Machine Learning" Packt Publishing ISBN 978-1-78712-238-3