● What is Big Data?
● What is the solution of Big data?
● How Apache Spark can help us?
● Apache Spark advantages over Hadoop MapReduce
What is Big Data?
● Lots of Data (Terabytes or Petabytes).
● Large and complex.
● Difficult to deal using Relational Databases.
● Challenges faced in - searching, storing, transfer, analysis, visualisation.
● Require Parallel processing on 100s of machines.
● Allows distributed processing of large datasets across clusters.
● It is open source database management with scale out storage and distributed
● Map - Data is converted into tuples (key/value pair).
● Reduce - Takes input from map and combines input from map to form smaller set of
○ Scale data
○ Parallel Processing
○ Built in fault tolerant
Shortcomings of MapReduce
1. Slow for Iterative Jobs.
2. Slow for Interactive Ad-hoc queries.
3. Operations - Forces task be of type Map and Reduce.
4. Difficult to program - Even simple join operations also require extensive code.
Lacks data sharing. Data sharing done through stable storage (HDFS)→ slow.
Slow due to replication and Disk I/O but it is essential for fault tolerance.
Can we use memory? How will it be fault tolerant?
● Developed in 2009 by UC Berkeley.
● Processing engine.
● Used for speed, ease of use, and sophisticated analytics.
● It is based on Hadoop MapReduce but it extends MapReduce for performing
more types of computations.
● Spark participated in Daytona Gray category, Spark sorted 100 TB of data (1
trillion records) the same data three time faster using ten times fewer
machines as compared to Hadoop.
● Improves efficiency through
○ In-memory data sharing.
○ General computation graph.
● Improves usability through
○ Rich APIs in Java, Scala, Python.
○ Interactive Shell.
Upto 100x faster in memory
and 10x faster on disk
Upto 2-5x less code
Resilient Distributed Dataset (RDD)
● Fundamental Data Structure of Apache Spark.
● Read-only collection of objects partitioned across a set of machines.
● Perform In-memory Computation.
● Build on transformation operations like map, filter etc.
● Fault tolerant through lineage.
○ Lazy Evaluated
Resilient Distributed Dataset (RDD)
Two types of operation can be performed:
○ Create new RDD from existing RDD.
○ Creates DAG.
○ Lazily evaluated.
○ Increases efficiency by not returning large dataset.
○ Eg. GroupByKey, ReduceByKey, filter.
○ All queries are executed.
○ Performs computation.
○ Returns result to driver program.
○ Eg. collect, count, take.
# Creates a list of animal.
animals = ['cat', 'dog', 'elephant', 'cat', 'mouse', ’cat’]
# Parallelize method is used to create RDD from list. Here “animalRDD” is created.
#sc is Object of Spark Context.
animalRDD = sc.parallelize(animals)
# Since RDD is lazily evaluated, to print it we perform an action operation, i.e.
collect() which is used to print the RDD.
Output - ['cat', 'dog', 'elephant', 'cat', 'mouse', 'cat']
Creating RDD from file
#The file words.txt has names of animals through which animalsRDD is made.
animalsRDD = sc.textFile('/path/to/file/words.txt')
#collect() is the action operation.
Map operation on RDD
‘’’’’ To count the frequency of animals, we make (key/value) pair - (animal,1) for all
the animals and then perform reduce operation which counts all the values.
Lambda is used to write inline functions in python.
mapRDD = animalRDD.map(lambda x:(x,1))
Output - [('cat',1), ('dog',1), ('elephant',1), ('cat',1), ('mouse',1), ('cat',1)]
Reduce operation on RDD
‘’’’’ reduceByKey is used to perform reduce operation on same key. So in its
arguments, we have defined a function to add the values for same key. Hence, we
get the count of animals.
reduceRDD = mapRDD.reduceByKey(lambda x,y:x+y)
Output - [('cat',3), ('dog',1), ('elephant',1), ('mouse',1)]
Filter operation on RDD
‘’’’’ Filter all the animals obtained from reducedRDD with count greater than 2. x is
a tuple made of (animal, count), i.e. x=animal name and x=count of animal.
Therefore we filter the reduceRDD based on x>2.
filterRDD = reduceRDD.filter(lambda x:x>2)
Output - [('cat',3)]
Please refer http://spark.apache.org/docs/latest/programming-guide.html for
more about programming in Apache Spark.
Spark vs. Hadoop
○ Spark better as it does in-memory computation.
○ Hadoop is good for one pass ETL jobs and where data does not fit in memory.
● Ease of use
○ Spark is easier to program and provides API in Java, Scala, R, Python.
○ Spark has an interactive mode.
○ Hadoop MapReduce is more difficult to program but many tools are available to
make it easier.
○ Spark is cost effective according to benchmark, though staffing can be costly.
○ Compatibility to data types and data sources is the same for both.
Spark vs. Hadoop
● Data Processing
○ Spark can perform real time processing and batch processing.
○ Hadoop MapReduce is good for batch processing. Hadoop requires storm for real
time processing, Giraph for graph processing, Mahout for machine learning.
● Fault tolerant
○ Hadoop MapReduce is slightly more tolerant.
○ Spark can cache the input data.
Companies that uses Hadoop and Spark are:
● Hadoop - Hadoop is used good for static operation.
○ Dell, IBM, Cloudera, AWS and many more.
○ Real-time marketing campaign, online product recommendations etc.
○ eBay, Amazon, Yahoo, Nokia and many more.
○ Data mining 40x times faster than Hadoop (Conviva).
○ Traffic Prediction via EM (Mobile Millennium).
○ DNA Sequence Analysis (SNAP).
○ Twitter Spam Classification (Monarch).
Apache Spark helping companies grow in
● Spark Helps Pinterest Identify Trends - Using Spark, Pinterest is able to
identify—and react to—developing trends as they happen.
● Netflix Leans on Spark for Personalization Aid - Netflix uses Spark to support
real-time stream processing for online recommendations and data monitoring.
Libraries of Apache Spark
Spark provides libraries to provide generality. We can combine these libraries
seamlessly in the same application to provide more functionality.
Libraries provided by Apache Spark are:
1. Spark Streaming - It supports scalable and fault tolerant processing of
2. Spark SQL - It allows spark to work with structured data.
3. Spark MLlib - It provides scalable machine learning library and has machine
learning and statistical algorithms.
4. Spark GraphX - It is used to compute graphs over data.
Refer http://spark.apache.org/docs/latest/ for more information.