Wednesday, 6 January 2016

Introduction to Apache Spark with Examples and Use Cases

Today, Spark is being adopted by major players like Amazon, eBay, and Yahoo! Many organizations run Spark on clusters with thousands of nodes. According to the Spark FAQ, the largest known cluster has over 8000 nodes. Indeed, Spark is a technology well worth taking note of and learning about.
apache spark tutorial
This article provides an introduction to Spark including use cases and examples. It contains information from the Apache Spark website as well as the book Learning Spark - Lightning-Fast Big Data Analysis.

What is Apache Spark? An Introduction

Spark is an Apache project advertised as “lightning fast cluster computing”. It has a thriving open-source community and is the most active Apache project at the moment.
Spark provides a faster and more general data processing platform. Spark lets you run programs up to 100x faster in memory, or 10x faster on disk, than Hadoop. Last year, Spark took over Hadoop by completing the 100 TB Daytona GraySort contest 3x faster on one tenth the number of machines and it also became thefastest open source engine for sorting a petabyte.
Spark also makes it possible to write code more quickly as you have over 80 high-level operators at your disposal. To demonstrate this, let’s have a look at the “Hello World!” of BigData: the Word Count example. Written in Java for MapReduce it has around 50 lines of code, whereas in Spark (and Scala) you can do it as simply as this:
            .flatMap(line => line.split(" "))
            .map(word => (word, 1)).reduceByKey(_ + _)
Another important aspect when learning how to use Apache Spark is the interactive shell (REPL) which it provides out-of-the box. Using REPL, one can test the outcome of each line of code without first needing to code and execute the entire job. The path to working code is thus much shorter and ad-hoc data analysis is made possible.
Additional key features of Spark include:
  • Currently provides APIs in Scala, Java, and Python, with support for other languages (such as R) on the way
  • Integrates well with the Hadoop ecosystem and data sources (HDFS, Amazon S3, Hive, HBase, Cassandra, etc.)
  • Can run on clusters managed by Hadoop YARN or Apache Mesos, and can also run standalone
The Spark core is complemented by a set of powerful, higher-level libraries which can be seamlessly used in the same application. These libraries currently include SparkSQL, Spark Streaming, MLlib (for machine learning), and GraphX, each of which is further detailed in this article. Additional Spark libraries and extensions are currently under development as well.
spark libraries and extensions

Spark Core

Spark Core is the base engine for large-scale parallel and distributed data processing. It is responsible for:
  • memory management and fault recovery
  • scheduling, distributing and monitoring jobs on a cluster
  • interacting with storage systems
Spark introduces the concept of an RDD (Resilient Distributed Dataset), an immutable fault-tolerant, distributed collection of objects that can be operated on in parallel. An RDD can contain any type of object and is created by loading an external dataset or distributing a collection from the driver program.
RDDs support two types of operations:
  • Transformations are operations (such as map, filter, join, union, and so on) that are performed on an RDD and which yield a new RDD containing the result.
  • Actions are operations (such as reduce, count, first, and so on) that return a value after running a computation on an RDD.
Transformations in Spark are “lazy”, meaning that they do not compute their results right away. Instead, they just “remember” the operation to be performed and the dataset (e.g., file) to which the operation is to be performed. The transformations are only actually computed when an action is called and the result is returned to the driver program. This design enables Spark to run more efficiently. For example, if a big file was transformed in various ways and passed to first action, Spark would only process and return the result for the first line, rather than do the work for the entire file.
By default, each transformed RDD may be recomputed each time you run an action on it. However, you may also persist an RDD in memory using the persist or cache method, in which case Spark will keep the elements around on the cluster for much faster access the next time you query it.


SparkSQL is a Spark component that supports querying data either via SQL or via the Hive Query Language. It originated as the Apache Hive port to run on top of Spark (in place of MapReduce) and is now integrated with the Spark stack. In addition to providing support for various data sources, it makes it possible to weave SQL queries with code transformations which results in a very powerful tool. Below is an example of a Hive compatible query:
// sc is an existing SparkContext.
val sqlContext = new org.apache.spark.sql.hive.HiveContext(sc)

sqlContext.sql("CREATE TABLE IF NOT EXISTS src (key INT, value STRING)")
sqlContext.sql("LOAD DATA LOCAL INPATH 'examples/src/main/resources/kv1.txt' INTO TABLE src")

// Queries are expressed in HiveQL
sqlContext.sql("FROM src SELECT key, value").collect().foreach(println)

Spark Streaming

Spark Streaming supports real time processing of streaming data, such as production web server log files (e.g. Apache Flume and HDFS/S3), social media like Twitter, and various messaging queues like Kafka. Under the hood, Spark Streaming receives the input data streams and divides the data into batches. Next, they get processed by the Spark engine and generate final stream of results in batches, as depicted below.
spark streaming
The Spark Streaming API closely matches that of the Spark Core, making it easy for programmers to work in the worlds of both batch and streaming data.


MLlib is a machine learning library that provides various algorithms designed to scale out on a cluster for classification, regression, clustering, collaborative filtering, and so on (check out Toptal’s article on machine learning for more information on that topic). Some of these algorithms also work with streaming data, such as linear regression using ordinary least squares or k-means clustering (and more on the way). Apache Mahout (a machine learning library for Hadoop) has already turned away from MapReduce and joined forces on Spark MLlib.


Interview Questions & Answers on Apache Spark [Part 3]

1.What is Apache Spark?
Spark is a fast, easy-to-use and flexible data processing framework. It has an advanced execution engine supporting cyclic data  flow and in-memory computing. Spark can run on Hadoop, standalone or in the cloud and is capable of accessing diverse data sources including HDFS, HBase, Cassandra and others.

2.Explain key features of Spark.

  • Allows Integration with Hadoop and files included in HDFS.
  • Spark has an interactive language shell as it has an independent Scala (the language in which Spark is written) interpreter
  • Spark consists of RDD’s (Resilient Distributed Datasets), which can be cached across computing nodes in a cluster.
  • Spark supports multiple analytic tools that are used for interactive query analysis , real-time analysis and graph processing

3.Define RDD.
RDD is the acronym for Resilient Distribution Datasets – a fault-tolerant collection of operational elements that run parallel. The partitioned data in RDD is immutable and distributed. There are primarily two types of RDD:

  • Parallelized Collections : The existing RDD’s running parallel with one another
  • Hadoop datasets: perform function on each file record in HDFS or other storage system

4.What does a Spark Engine do?
Spark Engine is responsible for scheduling, distributing and monitoring the data application across the cluster.

5.Define Partitions?
As the name suggests, partition is a smaller and logical division of data  similar to ‘split’ in MapReduce. Partitioning is the process to derive logical units of data to speed up the processing process. Everything in Spark is a partitioned RDD.

6.What operations RDD support?

  • Transformations
  • Actions

7.What do you understand by Transformations in Spark?
Transformations are functions applied on RDD, resulting into another RDD. It does not execute until an action occurs. map() and filer() are examples of transformations, where the former applies the function passed to it on each element of RDD and results into another RDD. The filter() creates a new RDD by selecting elements form current RDD that pass function argument.

8. Define Actions.
An action helps in bringing back the data from RDD to the local machine. An action’s execution is the result of all previously created transformations. reduce() is an action that implements the function passed again and again until one value if left. take() action takes all the values from RDD to local node.

9.Define functions of SparkCore.
Serving as the base engine, SparkCore performs various important functions like memory management, monitoring jobs, fault-tolerance, job scheduling and interaction with storage systems.

10.What is RDD Lineage?
Spark does not support data replication in the memory and thus, if any data is lost, it is rebuild using RDD lineage. RDD lineage is a process that reconstructs lost data partitions. The best is that RDD always remembers how to build from other datasets.

11.What is Spark Driver?
Spark Driver is the program that runs on the master node of the machine and declares transformations and actions on data RDDs. In simple terms, driver in Spark creates SparkContext, connected to a given Spark Master.

The driver also delivers the RDD graphs to Master, where the standalone cluster manager runs.

12.What is Hive on Spark?
Hive contains significant support for Apache Spark, wherein Hive execution is configured to Spark:
hive> set spark.home=/location/to/sparkHome;
hive> set hive.execution.engine=spark;

Hive on Spark supports Spark on yarn mode by default.

13.Name commonly-used Spark Ecosystems.

  • Spark SQL (Shark)- for developers
  • Spark Streaming for processing live data streams
  • GraphX for generating and computing graphs
  • MLlib (Machine Learning Algorithms)
  • SparkR to promote R Programming in Spark engine.

14.Define Spark Streaming.
Spark supports stream processing – an extension to the Spark API , allowing stream processing of live data streams. The data from different sources like Flume, HDFS is streamed and finally processed to file systems, live dashboards and databases. It is similar to batch processing as the input data is divided into streams like batches.

15.What is GraphX?
Spark uses GraphX for graph processing to build and transform interactive graphs. The GraphX component enables programmers to reason about structured data at scale.

16.What does MLlib do?
MLlib is scalable machine learning library provided by Spark. It aims at making machine learning easy and scalable with common learning algorithms and use cases like clustering, regression filtering, dimensional reduction, and alike.

17.What is Spark SQL?
SQL Spark, better known as Shark is a novel module introduced in Spark to work with structured data and perform structured data processing. Through this module, Spark executes relational SQL queries on the data. The core of the component supports an altogether different RDD called SchemaRDD, composed of rows objects and schema objects defining data type of each column in the row. It is similar to a table in relational database.

18.What is a Parquet file?
Parquet is a columnar format file supported by many other data processing systems. Spark SQL performs both read and write operations with Parquet file and consider it be one of the best big data analytics format so far.

19.What file systems Spark support?
• Hadoop Distributed File System (HDFS)
• Local File system
• S3

20.What is Yarn?
Similar to Hadoop, Yarn is one of the key features in Spark, providing a central and resource management platform to deliver scalable operations across the cluster . Running Spark on Yarn necessitates a binary distribution of Spar as built on Yarn support.

21.List the functions of Spark SQL.
Spark SQL is capable of:
• Loading data from a variety of structured sources
• Querying data using SQL statements, both inside a Spark program and from external tools that connect to Spark SQL through standard database connectors (JDBC/ODBC). For instance, using business intelligence tools like Tableau
• Providing rich integration between SQL and regular Python/Java/Scala code, including the ability to join RDDs and SQL tables, expose custom functions in SQL, and more

22.What are benefits of Spark over MapReduce?

  • Due to the availability of in-memory processing, Spark implements the processing around 10-100x faster than Hadoop MapReduce. MapReduce makes use of persistence storage for any of the data processing tasks.
  • Unlike Hadoop, Spark provides in-built libraries to perform multiple tasks form the same core like batch processing, Steaming, Machine learning, Interactive SQL queries. However, Hadoop only supports batch processing.
  • Hadoop is highly disk-dependent whereas Spark promotes caching and in-memory data storage
  • Spark is capable of performing computations multiple times on the same dataset. This is called iterative computation while there is no iterative computing implemented by Hadoop.

23.Is there any benefit of learning MapReduce, then?
Yes, MapReduce is a paradigm used by many big data tools including Spark as well. It is extremely relevant to use MapReduce when the data grows bigger and bigger. Most tools like Pig and Hive convert their queries into MapReduce phases to optimize them better.

24.What is Spark Executor?
When SparkContext connect to a cluster manager, it acquires an Executor on nodes in the cluster. Executors are Spark processes that run computations and store the data on the worker node. The final tasks by SparkContext are transferred to executors for their execution.

25.Name types of Cluster Managers in Spark.
The Spark framework supports three major types of Cluster Managers:

  • Standalone: a basic manager to set up a cluster
  • Apache Mesos: generalized/commonly-used cluster manager, also runs Hadoop MapReduce and other applications
  • Yarn: responsible for resource management in Hadoop

26.What do you understand by worker node?
Worker node refers to any node that can run the application code in a cluster.

27.What is PageRank?
A unique feature and algorithm in graph, PageRank is the measure of each vertex in the graph. For instance, an edge from u to v represents endorsement of v’s importance by u. In simple terms, if a user at Instagram is followed massively, it will rank high on that platform.

28.Do you need to install Spark on all nodes of Yarn cluster while running Spark on Yarn?
No because Spark runs on top of Yarn.

29.Illustrate some demerits of using Spark.
Since Spark utilizes more storage space compared to Hadoop and MapReduce, there may arise certain problems. Developers need to be careful while running their applications in Spark. Instead of running everything on a single node, the work must be distributed over multiple clusters. 

30.How to create RDD?
Spark provides two methods to create RDD:
• By parallelizing a collection in your Driver program. This makes use of SparkContext’s ‘parallelize’ method
val data = Array(2,4,6,8,10)
val distData = sc.parallelize(data)

• By loading an external dataset from external storage like HDFS, HBase, shared file system

Interview Questions & Answers on Apache Spark [Part 2]

Q1: Say I have a huge list of numbers in RDD(say myrdd). And I wrote the following code to compute average:
def myAvg(x, y):
 return (x+y)/2.0;
avg = myrdd.reduce(myAvg);
What is wrong with it? And How would you correct it?
Ans: The average function is not commutative and associative;
I would simply sum it and then divide by count.
def sum(x, y):
 return x+y;
total = myrdd.reduce(sum);
avg = total / myrdd.count();
The only problem with the above code is that the total might become very big thus over flow. So, I would rather divide each number by count and then sum in the following way.
cnt = myrdd.count();
def devideByCnd(x):
 return x/cnt;
myrdd1 =;
avg = myrdd.reduce(sum);

Q2: Say I have a huge list of numbers in a file in HDFS. Each line has one number.And I want to compute the square root of sum of squares of these numbers. How would you do it?
# We would first load the file as RDD from HDFS on spark
numsAsText = sc.textFile("hdfs://namenode:9000/user/kayan/mynumbersfile.txt");
# Define the function to compute the squares
def toSqInt(str):
 v = int(str);
 return v*v;
#Run the function on spark rdd as transformation
nums =;

#Run the summation as reduce action
total = nums.reduce(sum)

#finally compute the square root. For which we need to import math.
import math;
print math.sqrt(total);

Q3: Is the following approach correct? Is the sqrtOfSumOfSq a valid reducer?

numsAsText =sc.textFile("hdfs://namenode:9000/user/kalyan/mynumbersfile.txt");
def toInt(str):
 return int(str);
nums =;
def sqrtOfSumOfSq(x, y):
 return math.sqrt(x*x+y*y);
total = nums.reduce(sum)
import math;
print math.sqrt(total);
Ans: Yes. The approach is correct and sqrtOfSumOfSq is a valid reducer.

Q4: Could you compare the pros and cons of the your approach (in Question 2 above) and my approach (in Question 3 above)?
You are doing the square and square root as part of reduce action while I am squaring in map() and summing in reduce in my approach.
My approach will be faster because in your case the reducer code is heavy as it is calling math.sqrt() and reducer code is generally executed approximately n-1 times the spark RDD.
The only downside of my approach is that there is a huge chance of integer overflow because I am computing the sum of squares as part of map.

Q5: If you have to compute the total counts of each of the unique words on spark, how would you go about it?

#This will load the bigtextfile.txt as RDD in the spark
lines = sc.textFile("hdfs://namenode:9000/user/kalyan/bigtextfile.txt");

#define a function that can break each line into words
def toWords(line):
      return line.split();

# Run the toWords function on each element of RDD on spark as flatMap transformation.
# We are going to flatMap instead of map because our function is returning multiple values.

words = lines.flatMap(toWords);

# Convert each word into (key, value) pair. Her key will be the word itself and value will be 1.
def toTuple(word):
     return (word, 1);

wordsTuple =;

# Now we can easily do the reduceByKey() action.
def sum(x, y):
    return x+y;

counts = wordsTuple.reduceByKey(sum)

# Now, print
Q6: In a very huge text file, you want to just check if a particular keyword exists. How would you do this using Spark?
lines = sc.textFile("hdfs://namenode:9000/user/kalyan/bigtextfile.txt");
def isFound(line):
 if line.find(“mykeyword”) > -1:
  return 1;
 return 0;
foundBits =;
sum = foundBits.reduce(sum);
if sum > 0:
 print “FOUND”;
 print “NOT FOUND”;

Q7: Can you improve the performance of this code in previous answer?
Ans: Yes. 
The search is not stopping even after the word we are looking for has been found. Our map code would keep executing on all the nodes which is very inefficient.
We could utilize accumulators to report whether the word has been found or not and then stop the job. Something on these line:
import thread, threading
from time import sleep
result = "Not Set"
lock = threading.Lock()
accum = sc.accumulator(0)
def map_func(line):
 #introduce delay to emulate the slowness
 if line.find("Adventures") > -1:
  return 1;
 return 0;
def start_job():
 global result
  sc.setJobGroup("job_to_cancel", "some description")
  lines = sc.textFile("hdfs://namenode:9000/user/kalyan/wordcount/input/big.txt");
  result =;
 except Exception as e:
  result = "Cancelled"
def stop_job():
 while accum.value < 3 :
supress = lock.acquire()
supress = thread.start_new_thread(start_job, tuple())
supress = thread.start_new_thread(stop_job, tuple())
supress = lock.acquire()

Interview Questions & Answers on Apache Spark [Part 1]

Q1: When do you use apache spark? OR  What are the benefits of Spark over Mapreduce?
  1. Spark is really fast. As per their claims, it runs programs up to 100x faster than Hadoop MapReduce in memory, or 10x faster on disk. It aptly utilizes RAM to produce the faster results.
  2. In map reduce paradigm, you write many Map-reduce tasks and then tie these tasks together using Oozie/shell script. This mechanism is very time consuming and the map-reduce task have heavy latency.
  3. And quite often, translating the output out of one MR job into the input of another MR job might require writing another code because Oozie may not suffice.
  4. In Spark, you can basically do everything using single application / console (pyspark or scala console) and get  the results immediately. Switching between 'Running something on cluster' and 'doing something locally' is fairly easy and straightforward. This also leads to less context switch of the developer and more productivity.
  5. Spark kind of equals to MapReduce and Oozie put together.

Q2: Is there are point of learning Mapreduce, then?
Ans: Yes. For the following reason: 
  1. Mapreduce is a paradigm used by many big data tools including Spark. So, understanding the MapReduce paradigm and how to convert a problem into series of MR tasks is very important.
  2. When the data grows beyond what can fit into the memory on your cluster, the Hadoop Map-Reduce paradigm is still very relevant.
  3. Almost, every other tool such as Hive or Pig converts its query into MapReduce phases. If you understand the Mapreduce then you will be able to optimize your queries better.

Q3: When running Spark on Yarn, do I need to install Spark on all nodes of Yarn Cluster?
Since spark runs on top of Yarn, it utilizes yarn for the execution of its commands over the cluster's nodes.
So, you just have to install Spark on one node.

Q4: What are the downsides of Spark?
Spark utilizes the memory. The developer has to be careful. A casual developer might make following mistakes:
  1. She may end up running everything on the local node instead of distributing work over to the cluster.
  2. She might hit some webservice too many times by the way of using multiple clusters.

The first problem is well tackled by Hadoop Map reduce paradigm as it ensures that the data your code is churning is fairly small a point of time thus you can make a mistake of trying to handle whole data on a single node.
The second mistake is possible in Map-Reduce too. While writing Map-Reduce, user may hit a service from inside of map() or reduce() too many times. This overloading of service is also possible while using Spark.

Q5: What is a RDD?
The full form of RDD is resilience distributed dataset. It is a representation of data located on a network which is
  1. Immutable - You can operate on the rdd to produce another rdd but you can’t alter it.
  2. Partitioned / Parallel - The data located on RDD is operated in parallel. Any operation on RDD is done using multiple nodes.
  3. Resilience - If one of the node hosting the partition fails, another nodes takes its data.

RDD provides two kinds of operations: Transformations and Actions.

Q6: What is Transformations?
Ans: The transformations are the functions that are applied on an RDD (resilient distributed data set). The transformation results in another RDD. A transformation is not executed until an action follows.

The example of transformations are:
  1. map() - applies the function passed to it on each element of RDD resulting in a new RDD.
  2. filter() - creates a new RDD by picking the elements from the current RDD which pass the function argument.

Q7: What are Actions?
An action brings back the data from the RDD to the local machine. Execution of an action results in all the previously created transformation. The example of actions are:
  1. reduce() - executes the function passed again and again until only one value is left. The function should take two argument and return one value.
  2. take() - take all the values back to the local node form RDD.

Related Posts Plugin for WordPress, Blogger...