Combine by key to find max

Use combine by key and use map transformation to find Max value for all the keys  in Spark

Problem Statement : Given Employee(Avro) data saved in parquet file , one needs to find the maximum salary received by the each employee








import com.databricks.spark.avro.SchemaConverters;
import org.apache.avro.AvroRuntimeException;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.conf.Configured;
import org.apache.hadoop.util.Tool;
import org.apache.hadoop.util.ToolRunner;
import org.apache.spark.SparkConf;
import org.apache.spark.sql.DataFrame;
import org.apache.spark.sql.Row;
import org.apache.spark.sql.SQLContext;
import org.apache.spark.sql.types.StructField;
import org.apache.spark.sql.types.StructType;
import scala.Tuple2;

import java.util.ArrayList;
import java.util.List;
import java.util.TreeSet;

public class EmployeeMaxSalary extends Configured implements Tool, Closeable, Serializable {

    public static final String INPUT_PATH = "spark.input.path";
    public static final String OUTPUT_PATH = "spark.output.path";
    public static final String IS_RUN_LOCALLY = "";
    public static final String DEFAULT_FS = "spark.default.fs";
    public static final String NUM_PARTITIONS = "spark.num.partitions";

    // Just check because of a function use , the outer class is forced to be serialized
    // Example which throws light of serialization of Lambda function .
    private transient SQLContext sqlContext;
    private transient JavaSparkContext javaSparkContext;

    protected <T> JavaSparkContext getJavaSparkContext(final boolean isRunLocal,
                                                       final String defaultFs,
                                                       final Class<T> tClass) {
        final SparkConf sparkConf = new SparkConf()
                //Set spark conf here , 
                //after one gets spark context you can set hadoop configuration for InputFormats
                .set("spark.serializer", "org.apache.spark.serializer.KryoSerializer");

        if (isRunLocal) {

        final JavaSparkContext sparkContext = new JavaSparkContext(sparkConf);

        if (defaultFs != null) {
            sparkContext.hadoopConfiguration().set("fs.defaultFS", defaultFs);

        return sparkContext;

    // Convert Row to Avro POJO (Employee)
    public Employee convert(Row row) {
        try {
            // Employee Schema => ParquetRow Schema =>Row Schema

            Employee avroInstance = new Employee();

            for (StructField field : row.schema().fields()) {

                //row.fieldIndex => pos of the field Name in the schema
                avroInstance.put(, row.get(row.fieldIndex(;


            return avroInstance;

        } catch (Exception e) {
            throw new AvroRuntimeException("Avro POJO  building failed ", e);

    public int run(String[] args) throws Exception {

        //The arguments passed has been split into Key value by ToolRunner
        Configuration conf = getConf();
        String inputPath = conf.get(INPUT_PATH);

        String outputPath = conf.get(OUTPUT_PATH);

        //Get spark context, This is the central context , which can be wrapped in Any Other context
        javaSparkContext = getJavaSparkContext(conf.getBoolean(IS_RUN_LOCALLY, Boolean.FALSE), 
                                                    conf.get(DEFAULT_FS), EmployeeMaxSalary.class);
        sqlContext = new SQLContext(javaSparkContext);

        // No input path has been read, no job has not been started yet .
        //To set any configuration use javaSparkContext.hadoopConfiguration().set(Key,value);
        // To set any custom inputformat use javaSparkContext.newAPIHadoopFile() and get a RDD

        // Avro schema to StructType conversion
        final StructType outPutSchemaStructType = (StructType) SchemaConverters

        // read data from parquetfile, the schema of the data is taken from the avro schema
        DataFrame inputDf =

        // convert DataFrame into JavaRDD
        // the rows read from the parquetfile is converted into a Row object .
        // Row has same schema as that of the parquet file row
        JavaRDD<Row> rowJavaRDD = inputDf.javaRDD();

        //Row has same schema as that of Parquet row , 
        //Parquet Row has same schema as that of Avro Object

                // convert each Row to Employee Object

                // if i use a method call e -> convert(e) instead of static class,
                // i will need to serialize the Outer class
                // Lambda Functions internall needs to be serialized and is causing this issue

                .map(e -> convert(e))

                // Key by empid so that we can collect all the object on Reducer

                .combineByKey(new CreateCombiner(), new MergeValue(), new MergeCombiner())

                .map(new MapSpark());

        DataFrame outputDf = sqlContext.createDataFrame(rowJavaRDD, outPutSchemaStructType);

        // Convert JavaRDD to dataframe and save into parquet file

        return 0;

    public static class MapSpark implements Function<Tuple2<Integer, Object>, Object> {

        // LambdaFuncation used inside the Transformation are instantiated on Driver .
        // The Serialized object is sent to the executor
        // Making a filed transient helps in not serializing it
        private transient TreeSet<Long> employeeBonusSet;

        // Please do not declare any field with static
        // as Multiple task can spawn inside same JVM(Executor) is Spark

        public Object call(Tuple2<Integer, Object> v1) throws Exception {

            if (employeeBonusSet == null) {
                employeeBonusSet = new TreeSet<>();

            // Object is being reused and not created on every call

            EmployeeAgregator aggregatedEvents = (EmployeeAgregator) v1._2();

                            .forEach(o -> employeeBonusSet.add(((Employee) o)

            // select one object from the Employee List
            Object output = aggregatedEvents.getEmployeeList().get(0);

            ((Employee) output).setBonus(employeeBonusSet.last());

            return output;

    public static class EmployeeAgregator {

        private List<Object> employeeList;

        public EmployeeAgregator() {
            employeeList = new ArrayList<>();

        public List<Object> getEmployeeList() {
            return employeeList;

        public void addEmployee(Employee emp) {

        public void addEmployeeAgregator(EmployeeAgregator aggregator) {


    // This class would be  instantiated on MapTask 
    // for Every Employee Group(for a group and not individual input Row).
    // Only for first Row in the Group it would be instantiated
    public static class CreateCombiner implements Function<Employee, Object> {

        public Object call(Employee v1) throws Exception {
            EmployeeAgregator aggregator = new EmployeeAgregator();

            return aggregator;

    // Any Subsequent input With same employeeId will be added into the EmployeeAgregator here
    // This is like combiner from MapReduce
    public static class MergeValue implements Function2<Object, Employee, Object> {

        public Object call(Object v1, Employee v2) throws Exception {

            // Type of v1 is EmployeeAgregator  , Whereas Type of v2 is Employee
            ((EmployeeAgregator) v1).addEmployee(v2);
            return v1;

    // This will Be executed on Reduce Task . 
    // All EmployeeAgregator will be coming from Mapper and just merged on Reducer
    public static class MergeCombiner implements Function2<Object, Object, Object> {

        public Object call(Object v1, Object v2) throws Exception {

            //Type of both v1 and v2 are  EmployeeAgregator
            ((EmployeeAgregator) v1).addEmployeeAgregator((EmployeeAgregator) v2);

            return v1;

    public void close() throws IOException {

    public static void main(String[] args) throws Exception { EmployeeMaxSalary(), args);


Integration Test:


import org.apache.avro.generic.GenericData;
import org.apache.avro.generic.GenericRecord;
import org.apache.hadoop.fs.Path;
import org.junit.After;
import org.junit.Assert;
import org.junit.Before;
import org.junit.Test;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import parquet.avro.AvroParquetReader;
import parquet.avro.AvroParquetWriter;
import parquet.hadoop.ParquetReader;
import parquet.hadoop.ParquetWriter;
import parquet.hadoop.metadata.CompressionCodecName;


public class EmployeeMaxSalaryTest {

    private static final Logger LOG = LoggerFactory.getLogger(EmployeeMaxSalaryTest.class);
    private static final String BASEDIR = 
    "/tmp/EmployeeMaxSalaryTest/avroparquetInputFile/" + System.currentTimeMillis() + "/";
    private String input;
    private String output;

    private Employee employee;

    public void setUp() throws IOException {

        input = BASEDIR + "input/";
        output = BASEDIR + "output/";

        employee = new Employee();

        Employee employee2 = new Employee();

        //Write parquet file with GZIP compression
        ParquetWriter<Object> writer = AvroParquetWriter
                                            .builder(new Path(input + "1.gz.parquet"))

    public void testSuccess() throws Exception {

        String[] args = new String[]{"-D" + EmployeeMaxSalary.INPUT_PATH + "=" + input,
                "-D" + EmployeeMaxSalary.OUTPUT_PATH + "=" + output,
                "-D" + EmployeeMaxSalary.IS_RUN_LOCALLY + "=true",
                "-D" + EmployeeMaxSalary.DEFAULT_FS + "=file:///",
                "-D" + EmployeeMaxSalary.NUM_PARTITIONS + "=1"};


        ParquetReader<GenericRecord> reader = AvroParquetReader.builder(new Path(output)).build();
        //Use .withConf(FS.getConf()) 
        //for reading from a diferent HDFS and not local , by default the fs is local

        GenericData.Record event = (GenericData.Record);
        Employee outputEvent = AvroUtils.convertByteArraytoAvroPojo(
                                         Employee.getClassSchema()), Employee.getClassSchema());
        reader.close();"Data read from Sparkoutput is {}", outputEvent.toString());
        Assert.assertEquals(employee.getEmpId(), outputEvent.getEmpId());
        Assert.assertEquals(100L, outputEvent.getBonus().longValue());

    public void cleanup() throws IOException {
        FileUtils.deleteDirectory(new File(BASEDIR));