GPUdb Spark Connector

The documentation can be found at http://www.gpudb.com/docs/5.2/index.html. The connector specific documentation can be found at:

• www.gpudb.com/docs/5.2/connectors/spark_guide.html

For changes to the connector API, please refer to CHANGELOG.md. For changes to GPUdb functions, please refer to CHANGELOG-FUNCTIONS.md.

---

---

Spark Developer Manual

The following guide provides step by step instructions to get started integrating GPUdb with Spark.

This project is aimed to make GPUdb Spark accessible, meaning an RDD or DStream can be generated from a GPUdb table or can be saved to a GPUdb table.

Source code for the connector can be found at https://github.com/GPUdb/gpudb-connector-spark

Connector Classes

The three connector classes that integrate GPUdb with Spark are:

com.gpudb.spark.input

• GPUdbReader - Reads data from a table into an RDD
• GPUdbReceiver - A Spark streaming Receiver that receives data from a GPUdb table monitor stream

com.gpudb.spark.output

• GPUdbWriter - Writes data from an RDD or DStream into GPUdb

---

Spark Context

The GPUdb Spark connector uses the Spark configuration to pass GPUdb instance information to the Spark workers. Ensure that the following properties are set in the Spark configuration (SparkConf) of the Spark context (JavaSparkContext) when using each of the following connector interfaces:

GPUdbReader

• gpudb.host - The hostname or IP address of the GPUdb instance
• gpudb.port - The port number on which the GPUdb service is listening
• gpudb.threads - The number of threads GPUdb should use
• gpudb.table - The name of the GPUdb table being accessed

GPUdbWriter

• gpudb.host - The hostname or IP address of the GPUdb instance
• gpudb.port - The port number on which the GPUdb service is listening
• gpudb.threads - The number of threads GPUdb should use
• gpudb.table - The name of the GPUdb table being accessed
• gpudb.insert.size - The number of records to queue before inserting into GPUdb

---

Loading Data from GPUdb into a Spark RDD

To read from a GPUdb table, create a class that extends RecordObject and implements Serializable. For example::

  public class PersonRecord extends RecordObject implements Serializable
  {
     @RecordObject.Column(order = 0, properties = { ColumnProperty.DATA })
     public long id;

     @RecordObject.Column(order = 1) 
     public String name;        

     @RecordObject.Column(order = 2) 
     public long birthDate;   

     public PersonRecord(){}
  }

Note: The column order specified in the class must correspond to the table schema.

Next, instantiate a GPUdbReader for that class and call the readTable method with an optional filter expression::

  GPUdbReader<PersonRecord> reader = new GPUdbReader<PersonRecord>(sparkConf);
  JavaRDD<PersonRecord> rdd = reader.readTable(PersonRecord.class, "PeopleTable", expression, sparkContext);

The expression in the readTable call is equivalent to a SQL WHERE clause. For details, read the Expressions section of the Concepts page.

---

Saving Data from a Spark RDD to GPUdb

Creating a GPUdb table::

  GPUdbUtil.createTable(gpudbUrl, tableName, PersonRecord.class);

Writing to a GPUdb table::

  final GPUdbWriter<PersonRecord> writer = new GPUdbWriter<PersonRecord>(sparkConf);
  writer.write(rdd);

---

Receiving Data from GPUdb into a Spark DStream

The following creates a DStream from any new data inserted into the table tableName::

  GPUdbReceiver receiver = new GPUdbReceiver(gpudbUrl, gpudbStreamUrl, tableName);

  JavaReceiverInputDStream<AvroWrapper> dstream = javaStreamingContext.receiverStream(receiver);

Each record in the DStream is of type AvroWrapper, which is an Avro object along with its schema to decode it.

Note: At this time, only add and bulkadd functions will trigger the GPUdb to publish added records to ZMQ to be received by the Spark streaming interface. New records can also be added via the GPUdb administration page.

---

Saving Data from a Spark DStream to GPUdb

Creating a GPUdb table::

  GPUdbUtil.createTable(gpudbUrl, tableName, PersonRecord.class);

Writing to a GPUdb table::

  final GPUdbWriter<PersonRecord> writer = new GPUdbWriter<PersonRecord>(sparkConf);
  writer.write(dstream);

---

Examples

Examples can be found in the com.gpudb.spark package:

• BatchExample - Reading & writing GPUdb data via Spark using an RDD
• StreamExample - Reading & writing GPUdb data via Spark using a DStream

---

Installation & Configuration

The example code provided in this project assumes launching will be done on a Spark server using /bin/spark-submit. The example.sh script can run each example with minimal configuration via the example.properites file.

To install the example, the Spark connector RPM needs to be deployed onto the Spark driver host. The RPM generated by this project should be installed, where <X.Y.Z> is the GPUdb version and <YYYYMMDDhhmmss> is the build date::

    [root@local]# yum -y install gpudb-connector-spark-<X.Y.Z>-<YYYYMMDDhhmmss>.noarch.rpm

Once this RPM is installed, the following files should exist::

    /opt/gpudb/connectors/spark/example.properties
    /opt/gpudb/connectors/spark/example.sh
    /opt/gpudb/connectors/spark/gpudb-spark-5.2.0.jar
    /opt/gpudb/connectors/spark/gpudb-spark-5.2.0-jar-with-dependencies.jar
    /opt/gpudb/connectors/spark/gpudb-spark-5.2.0-node-assembly.jar
    /opt/gpudb/connectors/spark/gpudb-spark-5.2.0-shaded.jar
    /opt/gpudb/connectors/spark/README.md

The gpudb.host property in example.properties should be modified to be the name of the GPUdb host being accessed.

To run the example, issue this Unix command with no parameters to display usage information::

    [gpudb@local]$ ./example.sh