Using the Connector

Prerequisites

Before using the , ensure that you can identify:

The hostname of your Greenplum Database master node.
The port on which your Greenplum Database master server process is running, if it is not running on the default port (5432).
The name of the Greenplum database to which you want to connect.
The names of the Greenplum Database schema and table that you want to access.
The Greenplum Database user/role name and password that you have been assigned. Also ensure that this user/role name has the required privileges as described in Configuring Greenplum Database Role Privileges.

Downloading the Connector Package

The Connector is available as a separate download for Greenplum Database 5.x or 6.x from Broadcom Support Portal:

Download the Connector package by navigating to Broadcom Support Portal and locating Greenplum Spark Connector under the desired Greenplum release.

Note
For more information about download prerequisites, troubleshooting, and instructions, see Download Broadcom products and software.

The format of the Connector download file name is greenplum-connector-apache-spark-scala_<scala-version>-<gsc-version>.tar.gz. For example:

greenplum-connector-apache-spark-scala_2.12-2.0.0.tar.gz

The versions of Scala and Spark that you are developing for determine the package that you download:

Spark Version	Scala Version	Connector Package File
2.3.x , 2.4.x 2.4.x, 3.0.x	2.11 2.12	greenplum-connector-apache-spark-scala_2.11-2.0.0.tar.gz greenplum-connector-apache-spark-scala_2.12-2.0.0.tar.gz

The Connector download package includes the Connector JAR file and the product open source license. Extract the download package:
```
user@spark-node$ tar zxf greenplum-connector-apache-spark-scala_2.12-2.0.0.tar.gz
```
This command extracts the license text file and the JAR file named greenplum-connector-apache-spark-scala_2.12-2.0.0.jar into the current working directory.
Make note of the directory to which the Connector JAR file was extracted.

Using spark-shell

You can run Spark interactively through spark-shell, a modified version of the Scala shell. Refer to the spark-shell Spark documentation for detailed information on using this command.

To try out the Connector, run the spark-shell command providing a --jars option that identifies the file system path to the Connector JAR file. For example:

user@spark-node$ export GSC_JAR=/path/to/greenplum-connector-apache-spark-scala_2.12-2.0.0.jar
user@spark-node$ spark-shell --jars $GSC_JAR
< ... spark-shell startup output messages ... >
scala>

When you run spark-shell, you enter the scala> interactive subsystem. A SparkSession is instantiated for you and accessible via the spark local variable:

scala> println(spark)
org.apache.spark.sql.SparkSession@4113d9ab

Your SparkSession provides the entry points methods that you will use to transfer data between Spark and Greenplum Database.

Developing Applications with the Connector

If you are writing a stand-alone Spark application, you will bundle the Connector along with your other application dependencies into an "uber" JAR. The Spark Self-Contained Applications and Bundling Your Application's Dependencies documentation identifies additional considerations for stand-alone Spark application development.

You can use the spark-submit command to launch a Spark application assembled with the Connector. You can also run the spark-submit command providing a --jars option that identifies the file system path to the Connector JAR file. The spark-submit Spark documentation describes using this command.

Constructing the Greenplum Database JDBC URL

The Connector uses a JDBC connection to communicate with the Greenplum Database master node. The PostgreSQL JDBC driver version 42.2.14 is bundled with the Connector JAR file, so you do not need to manage this dependency. You may also use a custom JDBC driver with the Connector.

You must provide a JDBC connection string URL when you use the Connector to transfer data between Greenplum Database and Spark. This URL must include the Greenplum Database master hostname and port, as well as the name of the database to which you want to connect.

Parameter Name	Description
<master>	Hostname or IP address of the Greenplum Database master node.
<port>	The port on which the Greenplum Database server process is listening. Optional, default is 5432.
<database_name>	The Greenplum database to which you want to connect.

Note: The Connector requires that other connection options, including user name and password, be provided separately.

Using the Default PostgreSQL JDBC Driver

The JDBC connection string URL format for the default Connector JDBC driver is:

jdbc:postgresql://<master>[:<port>]/<database_name>

For example:

jdbc:postgresql://gpdb-master:5432/testdb

The syntax and semantics of the default JDBC connection string URL are governed by the PostgreSQL JDBC driver. For additional information about this syntax, refer to Connecting to the Database in the PostgreSQL JDBC documentation.

Using a Custom JDBC Driver

The Connector also supports using a custom JDBC driver. To use a custom Greenplum Database JDBC driver, you must:

Construct a JDBC connection string URL for your custom driver that includes the Greenplum Database master hostname and port and the name of the database to which you want to connect.
Provide the JAR file for the custom JDBC driver via one of the following options:
- Include a --jars <custom-jdbc-driver>.jar option on your spark-shell or spark-submit command line, identifying the full path to the custom JDBC driver JAR file.
- Bundle the Connector and custom JDBC JAR files along with your other application dependencies into an "uber" JAR.
- Install the custom JDBC JAR file in a known, configured location on your Spark executor nodes.

You must also identify the fully qualified Java class name of the JDBC driver in a Connector option (described in About Connector Options).

Configuring the Connector Server Address

The Connector uses the Greenplum Database gpfdist parallel file server to transfer data between Spark and Greenplum Database. When you run a Spark application that uses the Connector, the Connector starts a gpfdist server process on each Spark worker node.

By default, the Connector starts the gpfdist server process using the IP address of the Spark worker node and allows the operating system to select a random port for the server.

The Connector's default gpfdist addressing behaviour may not meet your needs if the hosts in your Spark cluster are configured with multiple network interfaces. The Connector exposes the following options to set the server address on multi-homed systems:

server.port - Specify the gpfdist server port number. You can specify any combination of one or more: single port number, comma-separated list of port numbers, or range of port numbers.
server.useHostname - Use the Spark worker host name instead of the IP Address.
server.hostEnv - Specify the IP address on which to start the gpfdist server.
server.nic - Specify the network interface on which to start the gpfdist server.

These options provide Spark applications finer-grained control over gpfdist server process addressing.

Per-Spark-Worker Configuration Using Environment Variables

The Connector allows you to set each of the port, host, and interface via an environment variable name of your choosing. When you set an option via an environment variable, you can configure a different value for each Spark worker node. Setting the Spark worker gpfdist address options in your Spark application is described in detail in Specifying the Connector Server Options.

If you choose to specify an option via an environment variable, set the environment variable on each Spark worker node before you run the start-slave.sh command on that node. For example, if you set server.port to the environment variable named GSC_EXTERNAL_PORT, and server.nic to the environment variable named GSC_NIF, you would start the Spark worker as follows:

user@spark-worker$ GSC_EXTERNAL_PORT="12900" GSC_NIF="eth1" start-slave.sh

You may also choose to set an environment variable in your spark-env.sh file. For information about the Spark spark-env.sh file, refer to the Environment Variables section of the Spark Configuration documentation.

Configuring the Connector Server Timeout

Greenplum Database uses the gpfdist protocol to communicate with the gpfdist server on each Spark worker node. The Connector exposes the server.timeout option to specify the "no activity" timeout for gpfdist server connections from Greenplum. The Connector defines activity as one or more bytes read from, or written to, the connection. The gpfdist server on the Spark node throws a timeout exception when the server.timeout period elapses without activity from Greenplum.

The default server.timeout is 300,000 milliseconds (5 minutes).

Setting the server.timeout option in your Spark application is described in detail in Specifying the Connector Server Options.

JDBC Connection Pooling

The Connector pools JDBC connections for each Spark application. The Connector creates a new connection pool for each unique combination of JDBC connection string URL, username, and password.

You can use Connector options to configure the size of the JDBC connection pool (pool.maxSize), the amount of time after which an inactive JDBC connection is considered idle (pool.timeoutMs), and the minimum number of idle JDBC connections allowed in the pool (pool.minIdle). These connection pool options bound the number of open JDBC connections between a Spark application and the Greenplum Database server.

Setting connection pool options in your Spark application is described in Specifying Connection Pool Options.

Take into consideration both the performance required for your Spark application and the desired resource impact on the Greenplum Database cluster when you change connection pool configuration:

Decreasing the maximum size of the connection pool bounds the number of open connections to the Greenplum Database server. Setting this option too low may decrease the parallelism of your Spark application.
The default minimum number of idle connections in the pool is zero (0). If you increase this value, be aware that the Connector maintains at least that number of open connections, and that some or all of the connections may be idle.
If you decrease the idle timeout for connections in the pool, the Connector closes idle connections sooner. Very short timeout values may defeat the purpose of connection pooling.

The Greenplum Database max_connections server configuration parameter identifies the maximum number of concurrent open connections that are permitted to the database server. Each running Spark application that uses the Connector owns a number of open connections on the Greenplum Database server. Greenplum Database Connection Errors provides troubleshooting information should you encounter Greenplum Database connection errors in your Spark application.