Mapreduce by examples

MapReduce
by examples
The code is available on:
https://github.com/andreaiacono/MapReduce
Take a look at my blog:
https://andreaiacono.blogspot.com/

MapReduce by examples
MapReduce is a programming model for
processing large data sets with a parallel,
distributed algorithm on a cluster
[src: http://en.wikipedia.org/wiki/MapReduce]
What is MapReduce?
Originally published in 2004 from Google
engineers Jeffrey Dean and Sanjay Ghemawat

Hadoop is the open source implementation of
the model by Apache Software foundation
The main project is composed by:
- HDFS
- YARN
- MapReduce
Its ecosystem is composed by:
- Pig
- Hbase
- Hive
- Impala
- Mahout
- a lot of other tools

Hadoop 2.x
- YARN: the resource manager, now called YARN, is now
detached from mapreduce framework
- java packages are under org.apache.hadoop.mapreduce.*

MapReduce inspiration
The name MapReduce comes from functional programming:
- map is the name of a higher-order function that applies a given function
to each element of a list. Sample in Scala:
val numbers = List(1,2,3,4,5)
numbers.map(x => x * x) == List(1,4,9,16,25)
- reduce is the name of a higher-order function that analyze a recursive
data structure and recombine through use of a given combining
operation the results of recursively processing its constituent parts,
building up a return value. Sample in Scala:
val numbers = List(1,2,3,4,5)
numbers.reduce(_ + _) == 15
MapReduce takes an input, splits it into smaller parts, execute the code of
the mapper on every part, then gives all the results to one or more reducers
that merge all the results into one.
src: http://en.wikipedia.org/wiki/Map_(higher-order_function)
http://en.wikipedia.org/wiki/Fold_(higher-order_function)

Overall view

How does Hadoop work?
Init
- Hadoop divides the input file stored on HDFS into splits (tipically of the size
of an HDFS block) and assigns every split to a different mapper, trying to
assign every split to the mapper where the split physically resides
Mapper
- locally, Hadoop reads the split of the mapper line by line
- locally, Hadoop calls the method map() of the mapper for every line passing
it as the key/value parameters
- the mapper computes its application logic and emits other key/value pairs
Shuffle and sort
- locally, Hadoop's partitioner divides the emitted output of the mapper into
partitions, each of those is sent to a different reducer
- locally, Hadoop collects all the different partitions received from the
mappers and sort them by key
Reducer
- locally, Hadoop reads the aggregated partitions line by line
- locally, Hadoop calls the reduce() method on the reducer for every line of
the input
- the reducer computes its application logic and emits other key/value pairs
- locally, Hadoop writes the emitted pairs output (the emitted pairs) to HDFS

Simplied flow (for developers)

Serializable vs Writable
- Serializable stores the class name and the object representation to the
stream; other instances of the class are referred to by an handle to the
class name: this approach is not usable with random access
- For the same reason, the sorting needed for the shuffle and sort phase
can not be used with Serializable
- The deserialization process creates a new instance of the object, while
Hadoop needs to reuse objects to minimize computation
- Hadoop introduces the two interfaces Writable and WritableComparable
that solve these problem

Writable wrappers
Java
primitive
Writable
implementation
boolean BooleanWritable
byte ByteWritable
short ShortWritable
int IntWritable
VIntWritable
float FloatWritable
long LongWritable
VLongWritable
double DoubleWritable
Java class Writable
implementation
String Text
byte[] BytesWritable
Object ObjectWritable
null NullWritable
Java
collection
Writable
implementation
array ArrayWritable
ArrayPrimitiveWritable
TwoDArrayWritable
Map MapWritable
SortedMap SortedMapWritable
enum EnumSetWritable

Implementing Writable: the SumCount class
public class SumCount implements WritableComparable<SumCount> {
DoubleWritable sum;
IntWritable count;
public SumCount() {
set(new DoubleWritable(0), new IntWritable(0));
}
public SumCount(Double sum, Integer count) {
set(new DoubleWritable(sum), new IntWritable(count));
}
@Override
public void write(DataOutput dataOutput) throws IOException {
sum.write(dataOutput);
count.write(dataOutput);
}
@Override
public void readFields(DataInput dataInput) throws IOException {
sum.readFields(dataInput);
count.readFields(dataInput);
}
// getters, setters and Comparable overridden methods are omitted
}

Glossary
Term Meaning
Job The whole process to execute: the input data,
the mapper and reducers execution and the
output data
Task Every job is divided among the several
mappers and reducers; a task is the job
portion that goes to every single mapper and
reducer
Split The input file is split into several splits (the
suggested size is the HDFS block size, 64Mb)
Record The split is read from mapper by default a line
at the time: each line is a record. Using a class
extending FileInputFormat, the record can
be composed by more than one line
Partition The set of all the key-value pairs that will be
sent to a single reducer. The default
partitioner uses an hash function on the key to
determine to which reducer send the data

Let's start coding!

WordCount
(the Hello World! for MapReduce, available in Hadoop sources)
Input Data:
The text of the book ”Flatland”
By Edwin Abbott.
Source:
http://www.gutenberg.org/cache/epub/201/pg201.txt
We want to count the occurrences of every word
of a text file

public static class TokenizerMapper extends Mapper<Object, Text, Text, IntWritable> {
private final static IntWritable one = new IntWritable(1);
private Text word = new Text();
@Override
public void map(Object key, Text value, Context context)
throws IOException, InterruptedException {
StringTokenizer itr = new StringTokenizer(value.toString());
while (itr.hasMoreTokens()) {
word.set(itr.nextToken().trim());
context.write(word, one);
}
}
}
WordCount mapper

public static class IntSumReducer extends Reducer<Text,IntWritable,Text,IntWritable>{
private IntWritable result = new IntWritable();
@Override
public void reduce(Text key, Iterable<IntWritable> values, Context context)
int sum = 0;
for (IntWritable val : values) {
sum += val.get();
}
result.set(sum);
context.write(key, result);
}
}
WordCount reducer

WordCount
a 936
ab 6
abbot 3
abbott 2
abbreviated 1
abide 1
ability 1
able 9
ablest 2
abolished 1
abolition 1
about 40
above 22
abroad 1
abrogation 1
abrupt 1
abruptly 1
absence 4
absent 1
absolute 2
...
Results:

MapReduce testing and debugging
- MRUnit is a testing framework based on Junit for unit
testing mappers, reducers, combiners (we'll see later what
they are) and the combination of the three
- Mocking frameworks can be used to mock Context or
other Hadoop objects
- LocalJobRunner is a class included in Hadoop that let us
run a complete Hadoop environment locally, in a single
JVM, that can be attached to a debugger. LocalJobRunner
can run at most one reducer
- Hadoop allows the creation of in-process mini clusters
programmatically thanks to MiniDFSCluster and
MiniMRCluster testing classes; debugging is more difficult
than LocalJobRunner because is multi-threaded and spread
over different VMs. Mini Clusters are used for testing
Hadoop sources.

MRUnit test for WordCount
@Test
public void testMapper() throws Exception {
new MapDriver<Object, Text, Text, IntWritable>()
.withMapper(new WordCount.TokenizerMapper())
.withInput(NullWritable.get(), new Text("foo bar foo"))
.withOutput(new Text("foo"), new IntWritable(1))
.withOutput(new Text("bar"), new IntWritable(1))
.runTest();
}
@Test
public void testReducer() throws Exception {
List<IntWritable> fooValues = new ArrayList<>();
fooValues.add(new IntWritable(1));
fooValues.add(new IntWritable(1));
List<IntWritable> barValue = new ArrayList<>();
barValue.add(new IntWritable(1));
new ReduceDriver<Text, IntWritable, Text, IntWritable>()
.withReducer(new WordCount.IntSumReducer())
.withInput(new Text("foo"), fooValues)
.withInput(new Text("bar"), barValue)
.runTest();
}

@Test
public void testMapReduce() throws Exception {
new MapReduceDriver<Object, Text, Text, IntWritable, Text, IntWritable>()
.withMapper(new WordCount.TokenizerMapper())
.withInput(NullWritable.get(), new Text("foo bar foo"))
.withReducer(new WordCount.IntSumReducer())
.runTest();
}
MRUnit test for WordCount

TopN
Input Data:
The text of the book ”Flatland”
By E. Abbott.
Source:
http://www.gutenberg.org/cache/epub/201/pg201.txt
We want to find the top-n used words of a text file

public static class TopNMapper extends Mapper<Object, Text, Text, IntWritable> {
private final static IntWritable one = new IntWritable(1);
private Text word = new Text();
private String tokens = "[_|$#<>^=[]*/,;,.-:()?!"']";
@Override
String cleanLine = value.toString().toLowerCase().replaceAll(tokens, " ");
StringTokenizer itr = new StringTokenizer(cleanLine);
word.set(itr.nextToken().trim());
context.write(word, one);
}
}
}
TopN mapper

public static class TopNReducer extends Reducer<Text, IntWritable, Text, IntWritable> {
private Map<Text, IntWritable> countMap = new HashMap<>();
@Override
int sum = 0;
sum += val.get();
}
countMap.put(new Text(key), new IntWritable(sum));
}
@Override
protected void cleanup(Context context) throws IOException, InterruptedException {
Map<Text, IntWritable> sortedMap = sortByValues(countMap);
int counter = 0;
for (Text key: sortedMap.keySet()) {
if (counter ++ == 20) {
break;
}
context.write(key, sortedMap.get(key));
}
}
}
TopN reducer

TopN
the 2286
of 1634
and 1098
to 1088
a 936
i 735
in 713
that 499
is 429
you 419
my 334
it 330
as 322
by 317
not 317
or 299
but 279
with 273
for 267
be 252
...
Results:

TopN
In the shuffle and sort phase, the partioner will send
every single word (the key) with the value ”1” to the
reducers.
All these network transmissions can be minimized if
we reduce locally the data that the mapper will emit.
This is obtained by a Combiner.

public static class Combiner extends Reducer<Text, IntWritable, Text, IntWritable> {
@Override
int sum = 0;
sum += val.get();
}
context.write(key, new IntWritable(sum));
}
}
TopN combiner

TopN
Without combiner
Map input records=4239
Map output records=37817
Map output bytes=359621
Input split bytes=118
Combine input records=0
Combine output records=0
Reduce input groups=4987
Reduce shuffle bytes=435261
Reduce input records=37817
Reduce output records=20
Hadoop output
With combiner

Combiners
If the function computed is
- commutative [a + b = b + a]
- associative [a + (b + c) = (a + b) + c]
we can reuse the reducer as a combiner!
Max function works:
max (max(a,b), max(c,d,e)) = max (a,b,c,d,e)
Mean function does not work:
mean(mean(a,b), mean(c,d,e)) != mean(a,b,c,d,e)

Combiners
Advantages of using combiners
- Network transmissions are minimized
Disadvantages of using combiners
- Hadoop does not guarantee the execution of a combiner:
it can be executed 0, 1 or multiple times on the same input
- Key-value pairs emitted from mapper are stored in local
filesystem, and the execution of the combiner could cause
expensive IO operations

private Map<String, Integer> countMap = new HashMap<>();
private String tokens = "[_|$#<>^=[]*/,;,.-:()?!"']";
@Override
String cleanLine = value.toString().toLowerCase().replaceAll(tokens, " ");
StringTokenizer itr = new StringTokenizer(cleanLine);
String word = itr.nextToken().trim();
if (countMap.containsKey(word)) {
countMap.put(word, countMap.get(word)+1);
}
else {
countMap.put(word, 1);
}
}
}
@Override
protected void cleanup(Context context) throws InterruptedException {
for (String key: countMap.keySet()) {
context.write(new Text(key), new IntWritable(countMap.get(key)));
}
}
TopN in-mapper combiner

private Map<Text, IntWritable> countMap = new HashMap<>();
@Override
int sum = 0;
sum += val.get();
}
countMap.put(new Text(key), new IntWritable(sum));
}
@Override
Map<Text, IntWritable> sortedMap = sortByValues(countMap);
int counter = 0;
for (Text key: sortedMap.keySet()) {
if (counter ++ == 20) {
break;
}
context.write(key, sortedMap.get(key));
}
}
TopN in-mapper reducer

Combiners - output
Without combiner
With combiner
With in-mapper
With in-mapper and combiner

Mean
Input Data:
Temperature in Milan
(DDMMYYY, MIN, MAX)
01012000, -4.0, 5.0
02012000, -5.0, 5.1
03012000, -5.0, 7.7
…
29122013, 3.0, 9.0
30122013, 0.0, 9.8
31122013, 0.0, 9.0
We want to find the mean max temperature for every month
Data source:
http://archivio-meteo.distile.it/tabelle-dati-archivio-meteo/

Mean mapper
private Map<String, List<Double>> maxMap = new HashMap<>();
@Override
String[] values = value.toString().split((","));
if (values.length != 3) return;
String date = values[DATE];
Text month = new Text(date.substring(2));
Double max = Double.parseDouble(values[MAX]);
if (!maxMap.containsKey(month)) {
maxMap.put(month, new ArrayList<Double>());
}
maxMap.get(month).add(max);
}
@Override
protected void cleanup(Mapper.Context context) throws InterruptedException {
for (Text month: maxMap.keySet()) {
List<Double> temperatures = maxMap.get(month);
Double sum = 0d;
for (Double max: temperatures) {
sum += max;
}
context.write(month, new DoubleWritable(sum));
}
}

Is this correct?
Mean mapper
private Map<String, List<Double>> maxMap = new HashMap<>();
@Override
}
}
@Override
protected void cleanup(Mapper.Context context) throws InterruptedException {
Double sum = 0d;
for (Double max: temperatures) {
sum += max;
}
context.write(month, new DoubleWritable(sum));
}
}

Mean
Mapper #1: lines 1, 2
Mapper #2: lines 3, 4, 5
Mapper#1: mean = (10.0 + 20.0) / 2 = 15.0
Mapper#2: mean = (2.0 + 4.0 + 3.0) / 3 = 3.0
Reducer mean = (15.0 + 3.0) / 2 = 9.0
But the correct mean is:
(10.0 + 20.0 + 2.0 + 4.0 + 3.0) / 5 = 7.8
Sample input data:
01012000, 0.0, 10.0
02012000, 0.0, 20.0
03012000, 0.0, 2.0
04012000, 0.0, 4.0
05012000, 0.0, 3.0
Not correct!

private Map<Text, List<Double>> maxMap = new HashMap<>();
@Override
}
}
@Override
Double sum = 0d;
for (Double max: temperatures) sum += max;
context.write(month, new SumCount(sum, temperatures.size()));
}
}
Mean mapper
This is correct!

private Map<Text, SumCount> sumCountMap = new HashMap<>();
@Override
public void reduce(Text key, Iterable<SumCount> values, Context context)
SumCount totalSumCount = new SumCount();
for (SumCount sumCount : values) {
totalSumCount.addSumCount(sumCount);
}
sumCountMap.put(new Text(key), totalSumCount);
}
@Override
for (Text month: sumCountMap.keySet()) {
double sum = sumCountMap.get(month).getSum().get();
int count = sumCountMap.get(month).getCount().get();
context.write(month, new DoubleWritable(sum/count));
}
}
Mean reducer

Mean
022012 7.230769230769231
022013 7.2
022010 7.851851851851852
022011 9.785714285714286
032013 10.741935483870968
032010 13.133333333333333
032012 18.548387096774192
032011 13.741935483870968
022003 9.278571428571428
022004 10.41034482758621
022005 9.146428571428572
022006 8.903571428571428
022000 12.344444444444441
022001 12.164285714285715
022002 11.839285714285717
...
Results:

Mean
Result:
R code to plot data:
temp <- read.csv(file="results.txt", sep="t", header=0)
names(temp) <- c("date","temperature")
ym <- as.yearmon(temp$date, format = "%m-%Y");
year <- format(ym, "%Y")
month <- format(ym, "%m")
ggplot(temp, aes(x=month, y=temperature, group=year)) + geom_line(aes(colour = year))

Join
Input Data - Users file:
"user_ptr_id" "reputation" "gold" "silver" "bronze"
"100006402" "18" "0" "0" "0"
"100022094" "6354" "4" "12" "50"
"100018705" "76" "0" "3" "4"
…
Input Data - Posts file:
"id" "title" "tagnames" "author_id" "body" "node_type" "parent_id" "abs_parent_id" "added_at" "score" …
"5339" "Whether pdf of Unit and Homework is available?" "cs101 pdf" "100000458" "" "question" "N" "N"
"2012-02-25 08:09:06.787181+00" "1"
"2312" "Feedback on Audio Quality" "cs101 production audio" "100005361" "We are looking for feedback on
the audio in our videos. Tell us what you think and try to be as specific as possible." "question"
"N" "N" "2012-02-23 00:28:02.321344+00" "2"
"2741" "where is the sample page for homework?" "cs101 missing_info homework" "100001178" "I am sorry if I
am being a nob ... but I do not seem to find any information regarding the sample page reffered to on the 1
question of homework 1." "question" "N" "N" "2012-02-23 09:15:02.270861+00" "0"
...
We want to combine information from the users file with
Information from the posts file (a join)
Data source:
http://content.udacity-data.com/course/hadoop/forum_data.tar.gz

@Override
FileSplit fileSplit = (FileSplit) context.getInputSplit();
String filename = fileSplit.getPath().getName();
String[] fields = value.toString().split(("t"));
if (filename.equals("forum_nodes_no_lf.tsv")) {
if (fields.length > 5) {
String authorId = fields[3].substring(1, fields[3].length() - 1);
String type = fields[5].substring(1, fields[5].length() - 1);
if (type.equals("question")) {
context.write(new Text(authorId), one);
}
}
}
else {
String authorId = fields[0].substring(1, fields[0].length() - 1);
String reputation = fields[1].substring(1, fields[1].length() - 1);
try {
int reputationValue = Integer.parseInt(reputation) + 2;
context.write(new Text(authorId),new IntWritable(reputationValue));
}
catch (NumberFormatException nfe) {
// just skips this record
}
}
}
Join mapper

@Override
int postsNumber = 0;
int reputation = 0;
String authorId = key.toString();
for (IntWritable value : values) {
int intValue = value.get();
if (intValue == 1) {
postsNumber ++;
}
else {
reputation = intValue -2;
}
}
context.write(new Text(authorId), new Text(reputation + "t" + postsNumber));
}
Join reducer

Join
USER_ID REPUTATION SCORE
00081537 1019 3
100011949 12 1
100105405 36 1
100000628 60 2
100011948 231 1
100000629 2090 1
100000623 1 2
100011945 457 4
100000624 167 1
100011944 114 3
100000625 1 1
100000626 93 1
100011942 11 1
100000620 1 1
100011940 35 1
100000621 2 1
100080016 11 2
100080017 53 1
100081549 1 1
...
Results:

Join
Result:
users <- read.csv(file="part-r-00000",sep='t', header=0)
users$V2[which(users$V2 > 10000,)] <- 0
plot(users$V2, users$V3, xlab="Reputation", ylab="Number of posts", pch=19, cex=0.4)

K-means
Input Data:
A random set of points
2.2705 0.9178
1.8600 2.1002
2.0915 1.3679
-0.1612 0.8481
-1.2006 -1.0423
1.0622 0.3034
0.5138 2.5542
...
We want to aggregate 2D points in clusters using
K-means algorithm
R code to generate dataset:
N <- 100
x <- rnorm(N)+1; y <- rnorm(N)+1; dat <- data.frame(x, y)
x <- rnorm(N)+5; y <- rnorm(N)+1; dat <- rbind(dat, data.frame(x, y))
x <- rnorm(N)+1; y <- rnorm(N)+5; dat <- rbind(dat, data.frame(x, y))

K-means mapper
@Override
protected void setup(Context context) throws IOException, InterruptedException {
URI[] cacheFiles = context.getCacheFiles();
centroids = Utils.readCentroids(cacheFiles[0].toString());
}
@Override
String[] xy = value.toString().split(" ");
double x = Double.parseDouble(xy[0]);
double y = Double.parseDouble(xy[1]);
int index = 0;
double minDistance = Double.MAX_VALUE;
for (int j = 0; j < centroids.size(); j++) {
double cx = centroids.get(j)[0];
double cy = centroids.get(j)[1];
double distance = Utils.euclideanDistance(cx, cy, x, y);
if (distance < minDistance) {
index = j;
minDistance = distance;
}
}
context.write(new IntWritable(index), value);
}

K-means reducer
public class KMeansReducer extends Reducer<IntWritable, Text, Text, IntWritable> {
@Override
protected void reduce(IntWritable key, Iterable<Text> values, Context context)
Double mx = 0d;
Double my = 0d;
int counter = 0;
for (Text value: values) {
String[] temp = value.toString().split(" ");
mx += Double.parseDouble(temp[0]);
my += Double.parseDouble(temp[1]);
counter ++;
}
mx = mx / counter;
my = my / counter;
String centroid = mx + " " + my;
context.write(new Text(centroid), key);
}
}

K-means driver - 1public static void main(String[] args) throws Exception {
Configuration configuration = new Configuration();
String[] otherArgs = new GenericOptionsParser(configuration, args).getRemainingArgs();
if (otherArgs.length != 3) {
System.err.println("Usage: KMeans <in> <out> <clusters_number>");
System.exit(2);
}
int centroidsNumber = Integer.parseInt(otherArgs[2]);
configuration.setInt(Constants.CENTROID_NUMBER_ARG, centroidsNumber);
configuration.set(Constants.INPUT_FILE, otherArgs[0]);
List<Double[]> centroids = Utils.createRandomCentroids(centroidsNumber);
String centroidsFile = Utils.getFormattedCentroids(centroids);
Utils.writeCentroids(configuration, centroidsFile);
boolean hasConverged = false;
int iteration = 0;
do {
configuration.set(Constants.OUTPUT_FILE, otherArgs[1] + "-" + iteration);
if (!launchJob(configuration)) {
System.exit(1);
}
String newCentroids = Utils.readReducerOutput(configuration);
if (centroidsFile.equals(newCentroids)) {
hasConverged = true;
}
else {
Utils.writeCentroids(configuration, newCentroids);
}
centroidsFile = newCentroids;
iteration++;
} while (!hasConverged);
writeFinalData(configuration, Utils.getCentroids(centroidsFile));
}

private static boolean launchJob(Configuration config) {
Job job = Job.getInstance(config);
job.setJobName("KMeans");
job.setJarByClass(KMeans.class);
job.setMapperClass(KMeansMapper.class);
job.setReducerClass(KMeansReducer.class);
job.setMapOutputKeyClass(IntWritable.class);
job.setMapOutputValueClass(Text.class);
job.setNumReduceTasks(1);
job.addCacheFile(new Path(Constants.CENTROIDS_FILE).toUri());
FileInputFormat.addInputPath(job, new Path(config.get(Constants.INPUT_FILE)));
FileOutputFormat.setOutputPath(job, new Path(config.get(Constants.OUTPUT_FILE)));
return job.waitForCompletion(true);
}
K-means driver - 2

K-means
Results:
4.5700 0.5510 2
4.5179 0.6120 2
4.1978 1.5706 2
5.2358 1.7982 2
1.747 3.9052 0
1.0445 5.0108 0
-0.6105 4.7576 0
0.7108 2.8032 1
1.3450 3.9558 0
1.2272 4.9238 0
...
points <- read.csv(file="final-data", sep="t", header=0)
colnames(points)[1] <- "x"
colnames(points)[2] <- "y"
plot(points$x, points$y, col= points$V3+2)

Hints
- Use MapReduce only if you have really big data: SQL or scripting
are less expensive in terms of time needed to obtain the same
results
- Use a lot of defensive checks: when we have a lot of data, we don't
want the computation to be stopped by a trivial NPE :-)
- Testing can save a lot of time!

Thanks!
The code is available on:
https://github.com/andreaiacono/MapReduce
Take a look at my blog:
https://andreaiacono.blogspot.com/

Mapreduce by examples

Related slideshows

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

More Related Content

What's hot

What's hot (20)

Similar to Mapreduce by examples

Similar to Mapreduce by examples (20)

More from Andrea Iacono

More from Andrea Iacono (6)

Recently uploaded

Recently uploaded (20)

Mapreduce by examples