How to sort comma separated keys in Reducer ouput?

Question

I am running an RFM Analysis program using MapReduce. The OutputKeyClass is Text.class and I am emitting comma separated R (Recency), F (Frequency), M (Monetory) as the key from Reducer where R=BigInteger, F=Binteger, M=BigDecimal and the value is also a Text representing Customer_ID. I know that Hadoop sorts output based on keys but my final result is a bit wierd. I want the output keys to be sorted by R first, then F and then M. But I am getting the following output sort order for unknown reasons:

545,1,7652    100000
545,23,390159.402343750    100001
452,13,132586    100002
452,4,32202    100004
452,1,9310    100007
452,1,4057    100018
452,3,18970    100021

But I want the following output:

545,23,390159.402343750    100001
545,1,7652    100000
452,13,132586    100002
452,4,32202    100004
452,3,18970    100021
452,1,9310    100007
452,1,4057    100018

NOTE: The customer_ID was the key in Map phase and all the RFM values belonging to a particular Customer_ID are brought together at the Reducer for aggregation.

Answer 1

So after a lot of searching I found some useful material the compilation of which I am posting now:

1. You have to start with your custom data type. Since I had three comma separated values which needed to be sorted descendingly, I had to create a TextQuadlet.java data type in Hadoop. The reason I am creating a quadlet is because the first part of the key will be the natural key and the rest of the three parts will be the R, F, M:

   import java.io.*;
   import org.apache.hadoop.io.*;
   public class TextQuadlet implements WritableComparable<TextQuadlet> {
   private String customer_id;
   private long R;
   private long F;
   private double M;
   public TextQuadlet() {
   }
   public TextQuadlet(String customer_id, long R, long F, double M) {
       set(customer_id, R, F, M);
   }
   public void set(String customer_id2, long R2, long F2, double M2) {
       this.customer_id = customer_id2;
       this.R = R2;
       this.F = F2;
       this.M=M2;
   }
   public String getCustomer_id() {
       return customer_id;
   }
   public long getR() {
       return R;
   }
   public long getF() {
       return F;
   }
   public double getM() {
       return M;
   }
   @Override
   public void write(DataOutput out) throws IOException {
       out.writeUTF(this.customer_id);
       out.writeLong(this.R);
       out.writeLong(this.F);
       out.writeDouble(this.M);
   }
   @Override
   public void readFields(DataInput in) throws IOException {
       this.customer_id = in.readUTF();
       this.R = in.readLong();
       this.F = in.readLong();
       this.M = in.readDouble();
   }
   // This hashcode function is important as it is used by the custom
   // partitioner for this class.
   @Override
   public int hashCode() {
       return (int) (customer_id.hashCode() * 163 + R + F + M);
   }
   @Override
   public boolean equals(Object o) {
       if (o instanceof TextQuadlet) {
           TextQuadlet tp = (TextQuadlet) o;
           return customer_id.equals(tp.customer_id) && R == (tp.R) && F==(tp.F) && M==(tp.M);
       }
       return false;
   }
   @Override
   public String toString() {
       return customer_id + "," + R + "," + F + "," + M;
   }
   // LHS in the conditional statement is the current key
   // RHS in the conditional statement is the previous key
   // When you return a negative value, it means that you are exchanging
   // the positions of current and previous key-value pair
   // Returning 0 or a positive value means that you are keeping the
   // order as it is
   @Override
   public int compareTo(TextQuadlet tp) {
   // Here my natural is is customer_id and I don't even take it into
   // consideration.
   
   // So as you might have concluded, I am sorting R,F,M descendingly.
       if (this.R != tp.R) {
           if(this.R < tp.R) {
               return 1;
           }
           else{
               return -1;
           }
       }
       if (this.F != tp.F) {
           if(this.F < tp.F) {
               return 1;
           }
           else{
               return -1;
           }
       }
       if (this.M != tp.M){
           if(this.M < tp.M) {
               return 1;
           }
           else{
               return -1;
           }
       }
       return 0;
   }
   public static int compare(TextQuadlet tp1, TextQuadlet tp2) {
       int cmp = tp1.compareTo(tp2);
       return cmp;
   }
   public static int compare(Text customer_id1, Text customer_id2) {
       int cmp = customer_id1.compareTo(customer_id1);
       return cmp;
   }
   }
   

2. Next you'll need a custom partitioner so that all the values which have the same key end up at one reducer:

   import org.apache.hadoop.io.Text;
   import org.apache.hadoop.mapreduce.Partitioner;
   
   public class FirstPartitioner_RFM extends Partitioner<TextQuadlet, Text> {
   @Override
   public int getPartition(TextQuadlet key, Text value, int numPartitions) {
       return (int) key.hashCode() % numPartitions;
      }
   }
   

3. Thirdly, you'll need a custom group comparater so that all the values are grouped together by their natural key which is customer_id and not the composite key which is customer_id,R,F,M:

   import org.apache.hadoop.io.WritableComparable;
   import org.apache.hadoop.io.WritableComparator;
   
   public class GroupComparator_RFM_N extends WritableComparator {
   protected GroupComparator_RFM_N() {
       super(TextQuadlet.class, true);
   }
   @SuppressWarnings("rawtypes")
   @Override
   public int compare(WritableComparable w1, WritableComparable w2) {
       TextQuadlet ip1 = (TextQuadlet) w1;
       TextQuadlet ip2 = (TextQuadlet) w2;
       // Here we tell hadoop to group the keys by their natural key.
       return ip1.getCustomer_id().compareTo(ip2.getCustomer_id());
       }
   }
   

4. Fourthly, you'll need a key comparater which will again sort the keys based on R,F,M descendingly and implement the same sort technique which is used in TextQuadlet.java. Since I got lost while coding, I slightly changed the way I compared data types in this function but the underlying logic is the same as in TextQuadlet.java:

   import org.apache.hadoop.io.WritableComparable;
   import org.apache.hadoop.io.WritableComparator;
   
   public class KeyComparator_RFM extends WritableComparator {
   protected KeyComparator_RFM() {
       super(TextQuadlet.class, true);
   }
   @SuppressWarnings("rawtypes")
   @Override
   public int compare(WritableComparable w1, WritableComparable w2) {
       TextQuadlet ip1 = (TextQuadlet) w1;
       TextQuadlet ip2 = (TextQuadlet) w2;
       // LHS in the conditional statement is the current key-value pair
       // RHS in the conditional statement is the previous key-value pair
       // When you return a negative value, it means that you are exchanging
       // the positions of current and previous key-value pair
       // If you are comparing strings, the string which ends up as the argument
       // for the `compareTo` method turns out to be the previous key and the
       // string which is invoking the `compareTo` method turns out to be the
       // current key.
       if(ip1.getR() == ip2.getR()){
           if(ip1.getF() == ip2.getF()){
               if(ip1.getM() == ip2.getM()){
                   return 0;
               }
               else{
                   if(ip1.getM() < ip2.getM())
                       return 1;
                   else
                       return -1;
               }
           }
           else{
               if(ip1.getF() < ip2.getF())
                   return 1;
               else
                   return -1;
           }
       }
       else{
           if(ip1.getR() < ip2.getR())
               return 1;
           else
               return -1;
           }
       }
   }
   

5. And finally, in your driver class, you'll have to include our custom classes. Here I have used TextQuadlet,Text as k-v pair. But you can choose any other class depending on your needs.:

   job.setPartitionerClass(FirstPartitioner_RFM.class);
   job.setSortComparatorClass(KeyComparator_RFM.class);
   job.setGroupingComparatorClass(GroupComparator_RFM_N.class);
   job.setMapOutputKeyClass(TextQuadlet.class);
   job.setMapOutputValueClass(Text.class);
   job.setOutputKeyClass(TextQuadlet.class);
   job.setOutputValueClass(Text.class);
   

Do correct me if I am technically going wrong somewhere in the code or in the explanation as I have based this answer purely on my personal understanding from what I read on the internet and it works for me perfectly.