MapReduce job execution flow in Hadoop

Introduction

MapReduce helps process big data by breaking tasks into small parts and running them on many computers at once.

When you have a huge amount of data to analyze, like logs from a website.

When you want to count words in many documents quickly.

When you need to sort or filter large datasets stored across many machines.

When you want to run data processing jobs that can be split into independent tasks.

When you want to use a simple programming model to handle complex data processing.

Syntax

Hadoop

MapReduce job runs in three main steps:
1. Map phase: Processes input data and creates key-value pairs.
2. Shuffle and Sort phase: Groups data by keys.
3. Reduce phase: Processes grouped data to produce final results.

The Map phase runs many tasks in parallel on chunks of data.

The Shuffle phase moves data between machines to group keys together.

Examples

This example counts how many times each word appears in a text.

Hadoop

Map phase: Read lines from a file and output (word, 1) for each word.
Shuffle phase: Group all (word, 1) pairs by word.
Reduce phase: Sum all counts for each word to get total occurrences.

This example summarizes user activity from log data.

Hadoop

Map phase: Extract user IDs and actions from logs.
Shuffle phase: Group actions by user ID.
Reduce phase: Summarize actions per user.

Sample Program

This Python program uses the MRJob library to count words in input data. The mapper splits lines into words and outputs each word with count 1. The reducer sums counts for each word.

Hadoop

from mrjob.job import MRJob

class MRWordCount(MRJob):
    def mapper(self, _, line):
        for word in line.split():
            yield word.lower(), 1

    def reducer(self, word, counts):
        yield word, sum(counts)

if __name__ == '__main__':
    MRWordCount.run()

Important Notes

The MapReduce framework handles splitting data and running tasks in parallel.

Shuffle and sort happen automatically between map and reduce phases.

Reducers receive all values for a key together to process them.

Summary

MapReduce breaks big data jobs into map and reduce tasks.

Map tasks process input and create key-value pairs.

Reduce tasks aggregate results by key to produce final output.

Practice

1. What is the main role of the Map task in a MapReduce job?

easy

A. Process input data and create key-value pairs

B. Aggregate results from all nodes

C. Store the final output in HDFS

D. Schedule tasks on cluster nodes

Solution

1. Step 1: Understand the Map task function

   The Map task reads input data and transforms it into intermediate key-value pairs.

2. Step 2: Differentiate from Reduce task

   The Reduce task aggregates these key-value pairs, so it is not responsible for initial processing.

3. Final Answer:

   Process input data and create key-value pairs -> Option A

4. Quick Check:

   Map task = create key-value pairs [OK]

Hint: Map creates key-value pairs from input data [OK]

Common Mistakes:

• Confusing Map with Reduce task role
• Thinking Map stores final output
• Assuming Map schedules tasks

2. Which of the following is the correct sequence of steps in a MapReduce job execution?

easy

A. Map -> Shuffle -> Reduce

B. Shuffle -> Map -> Reduce

C. Reduce -> Map -> Shuffle

D. Reduce -> Shuffle -> Map

Solution

1. Step 1: Recall MapReduce execution order

   The job starts with Map tasks processing input, then Shuffle sorts and transfers data, followed by Reduce tasks aggregating results.

2. Step 2: Match sequence to options

   Only Map -> Shuffle -> Reduce correctly shows Map first, then Shuffle, then Reduce.

3. Final Answer:

   Map -> Shuffle -> Reduce -> Option A

4. Quick Check:

   Execution order = Map, Shuffle, Reduce [OK]

Hint: Remember: Map first, then Shuffle, then Reduce [OK]

Common Mistakes:

• Mixing up Shuffle and Map order
• Placing Reduce before Map
• Ignoring Shuffle step

3. Consider this simplified MapReduce code snippet:

map(String key, String value) {
  emit(value, 1);
}

reduce(String key, Iterator values) {
  int sum = 0;
  while (values.hasNext()) {
    sum += values.next();
  }
  emit(key, sum);
}

What will be the output for input: ["a", "apple"], ["b", "apple"], ["c", "banana"]?

medium

A. [('apple', 1), ('banana', 1)]

B. [('a', 1), ('b', 1), ('c', 1)]

C. [('apple', 2), ('banana', 1)]

D. [('apple', 3), ('banana', 1)]

Solution

1. Step 1: Analyze Map output

   Map emits (value, 1) pairs: ('apple',1), ('apple',1), ('banana',1).

2. Step 2: Analyze Reduce aggregation

   Reduce sums counts for each key: 'apple' = 1+1=2, 'banana' = 1.

3. Final Answer:

   [('apple', 2), ('banana', 1)] -> Option C

4. Quick Check:

   Count of 'apple' = 2, 'banana' = 1 [OK]

Hint: Map emits value with 1, Reduce sums counts by key [OK]

Common Mistakes:

• Counting keys instead of values
• Confusing input keys with values
• Adding counts incorrectly

4. Identify the error in this MapReduce job execution flow:

Map task reads input and emits key-value pairs.
Reduce task tries to read input directly from HDFS instead of shuffle output.
Reduce task aggregates values by key.

medium

A. Reduce task should emit key-value pairs before Map

B. Map task should aggregate values, not Reduce task

C. Map task should read from shuffle output

D. Reduce task should read from shuffle output, not directly from HDFS

Solution

1. Step 1: Understand Reduce input source

   Reduce tasks receive data from the shuffle phase, not directly from HDFS input files.

2. Step 2: Identify incorrect flow

   Reading directly from HDFS in Reduce breaks the flow and causes errors.

3. Final Answer:

   Reduce task should read from shuffle output, not directly from HDFS -> Option D

4. Quick Check:

   Reduce reads shuffle output [OK]

Hint: Reduce reads shuffle output, never direct HDFS input [OK]

Common Mistakes:

• Thinking Reduce reads input files
• Swapping Map and Reduce roles
• Ignoring shuffle phase

5. You want to optimize a MapReduce job that processes large logs by filtering out entries with empty user IDs before counting occurrences. Where should you apply this filter for best performance?

hard

A. In the Reduce task after aggregation

B. In the Map task before emitting key-value pairs

C. During the shuffle phase

D. After the job completes on the output files

Solution

1. Step 1: Understand filtering impact

   Filtering early reduces data volume, saving network and processing time.

2. Step 2: Apply filter in Map task

   Filtering in Map avoids sending unnecessary data to shuffle and Reduce phases.

3. Final Answer:

   In the Map task before emitting key-value pairs -> Option B

4. Quick Check:

   Filter early in Map for best performance [OK]

Hint: Filter data early in Map to reduce workload [OK]

Common Mistakes:

• Filtering only in Reduce wastes resources
• Trying to filter during shuffle (not possible)
• Filtering after job completion is too late