How to sort comma separated keys in Reducer ouput? - sorting

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:
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;
}
}
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;
}
}
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());
}
}
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;
}
}
}
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.