Sunday, July 19, 2020

Setting up Tableau with Amazon EMR-Spark

Target audience: Intermediate
Estimated reading time: 20'

This post 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.


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.


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.


  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

    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) " +
    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 =
        // Generate and store data in Parquet columnar structure
        // 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'")"New table for $dataDir was created")
        // Just a quick validation test
        sparkSession.sql("SELECT id, value FROM records_table limit 4").show
      catch {
        case e: Exception => 
          logger.error(s"Failed to create HIVE table ${e.toString}")
    @throws(clazz = classOf[IllegalStateException])
    def s3CSVToDataFrame(
         s3CSVInputFile: String, 
         header: Boolean,
         s3Bucket: String,
         isMultiLine: Boolean
    )(implicit sparkSession: SparkSession): DataFrame = {
     import sparkSession.implicits._
     // Initialize the access configuration for Hadoop
     val loadDS = Seq[String]().toDS
     val accessConfig = loadDS.sparkSession.sparkContext.hadoopConfiguration
     try {
       accessConfig.set("fs.s3a.access.key", "xxxxxx")
       accessConfig.set("fs.s3a.secret.key", "xxxxxxx")
       val headerStr = if (header) "true" else "false"
       // Read the content of the CSV file from S3 to generate a data frame
             .option("header", headerStr)
             .option("delimiter", ",")
             .option("multiLine", isMultiLine)
             .load(path = s"s3a://${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


  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). 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.


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

  8. Remove potential locks in the metastore: rm -r metastore/*.lck (Locked access to the store will generate an error accessing and reading the table)

  9. Stop the Hive2 thrift server sudo /usr/lib/spark/sbin/

  10. Optionally kill the 2 processes related to thrift server ps -ef | grep RunJar sudo kill -9 {processId}

  11. Restart the thrift server sudo /usr/lib/spark/sbin/ --master local

  12. Verify the parquet data is correctly stored on HDFS hdfs dfs -ls /tmp/metrics/

  13. Verify table is created and populated in the EMR instance hive => show tables

  14. Launch Tableau desktop

  15. Select Amazon EMR-Hadoop connector

  16. 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.



No comments:

Post a Comment