This is an example project for HBase RDD. It currently runs on CDH 5.5, although it will run on other versions of CDH with minor modifications.

First, build the project with

sbt assembly

This will generate target/scala-2.12/hbase-rdd-examples-assembly-0.9.1.jar.

You can then copy this file, together with the files in the scripts directory, on a gateway machine of the cluster, and then run the scripts to launch the jobs. Of course, you may have to adapt some parameters in the scripts.

We assume the existence of a file test-input on the user directory on HDFS of the user running the job. Each line contains four fields, each one being a random printable string of length 10. Let us call k, col1, col2, col3 the four fields.

WriteSingleCf copies the contents of this file inside test-table, using k as rowkey, putting col1 and col2 under the column family cf1 and discarding col3.

WriteMultiCf copies the contents of this file inside test-table, using k as rowkey, putting col1 and col2 under the column family cf1 and col3 under cf2.

WriteBulk does the same as WriteSingleCf (on table test-table-bulk), by writing to HFiles on HDFS and then submitting these HFiles to the HBase servers.

All the write jobs create the table if it does not exist already - WriteBulk also takes care of computing splits appropriate for the file contents and a desired region size (128M in the example).

Read reads the contents of test-tables and reassambles a TSV output on HDFS under test-output, in the same format as the original. It does this by specifying both the column families and the columns to read, as a Map[String, Set[String]]. ReadTS is the same as above but also reads timestamps.

ReadCf does the same as Read but only specifies the column families, as a Set[String]. The whole column families are read - this is useful in the cases where the set of columns in a family is not known a priori, e.g. when the column families are used as a set (using a dummy marker value for all columns). ReadTSCf is the same as above but also reads timestamps.

DeleteSingleCf deletes contents from test-table, using k as rowkey, deleting col1 and col2 under the column family cf1.

DeleteMultiCf deletes contents from test-table, using k as rowkey, deleting col1 and col2 under the column family cf1 and col3 under cf2.

DeleteRows deletes rows from test-table, using k as rowkey.

In all jobs we are using String values for the cells, but HBaseRDD is not limited to this. Any other type A is supported, provided there is an implicit Reads[A] or Writes[A] in scope. These are traits defined by HBaseRDD that essentially wrap conversions from Array[Byte] to A and viceversa.

By default, HBaseRDD ships converters for Array[Byte] (duh!), String and JValue from Json4s, but you can write your own implicit conversions as necessary.

You can generate the test file as you prefer. A quick way would be to open a Scala console and write

import java.io._
import scala.util.Random
def printToFile(path: String)(op: PrintWriter => Unit) {
  val p = new PrintWriter(new File(path))
  try { op(p) } finally { p.close() }
}

def nextString = (1 to 10) map (_ => Random.nextPrintableChar) mkString
def nextLine = (1 to 4) map (_ => nextString) mkString "\t"

printToFile(args(0)) { p =>
  for (_ <- 1 to 100000000) {
    p.println(nextLine)
  }
}