Spark Core : Understanding RDD & Partitions in Spark

Let us see how to create an RDD in Spark.

 

RDD (Resilient Distributed Dataset): We can create RDD in 2 ways.

1. From Collections
• For small amount of data
• We can't use it for large amount of data
2. From Datasets 
• For huge amount of data
• Text, CSV, JSON, PDF, image etc.
• When data is large we should go with Dataset approach    

How to create an RDD ?

• Using collections
• val list = List(1, 2, 3, 4, 5, 6)
• val rdd = sc.parallelize(list)
• SC is Spark Context
• parallelize() method will convert input(collection in this case) into RDD
• Type of RDD will be based on the values assigned to collection, if we assign integers and RDD will be of type int

Let's see below Scala code :

# Created an RDD by providing a Collection(List) as input

scala> val rdd = sc.parallelize(List(1, 2, 3, 4, 5))

rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[0] at parallelize at <console>:23

# Printing RDD using collect() method

scala> rdd.collect()

res0: Array[Int] = Array(1, 2, 3, 4, 5)                                         

# Creating one more RDD(rdd1) using existing RDD(rdd) : Transformation

# map(x => x + 1) will increment each element in the given list by 1

# map() is a kind of transformation

scala> val rdd1 =rdd.map(x => x + 1)

rdd1: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[1] at map at <console>:23

# Printing first RDD(rdd), note that it won't change(Immutability)

scala> rdd.collect()

res1: Array[Int] = Array(1, 2, 3, 4, 5)

# Printing second RDD(rdd1)

scala> rdd1.collect()

res2: Array[Int] = Array(2, 3, 4, 5, 6)

Let's see more information about RDD from below code :

# Created rdd1 with a list of values 1, 2, 3, 4, 5

scala> val rdd1 = sc.parallelize(List(1, 2, 3, 4, 5))

rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[1] at parallelize at <console>:23

# Printed rdd1 using collect() 

scala> rdd1.collect()

res5: Array[Int] = Array(1, 2, 3, 4, 5)                                         

# Created another RDD, rdd2 and applied some transformation on the top of rdd1

scala> val rdd2 = rdd1.map(x => x + 1)

rdd2: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[2] at map at <console>:23

# Printed rdd2, note values have changed in new RDD as per the map() operation

scala> rdd2.collect()

res6: Array[Int] = Array(2, 3, 4, 5, 6)

# Apllied 3 transformations on the top of rdd2 and collected that RDD

scala> rdd2.map(x => x + 1).map(x => x +1).map(x => x + 1).collect()

res7: Array[Int] = Array(5, 6, 7, 8, 9)

Remember that starting point is RDD and ending point is collect(), in between we can have 'n' number of transformations. 

Example :

rdd

.map((x : Int) => {x + 1})

.map((x : Int) => {x + 1})

.map((x : Int) => {x + 1})

.map((x : Int) => {x + 1})

.map((x : Int) => {x + 1})

.collect()

We can do any transformation, see we have concatenated a string to each element.

scala> val rdd3 = rdd2.map(x => x + " Hi " + x)

rdd3: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[6] at map at <console>:23

scala> rdd3.collect()

res8: Array[String] = Array(2 Hi 2, 3 Hi 3, 4 Hi 4, 5 Hi 5, 6 Hi 6)

Scala provide below simplified feature to write code.

rdd.map((x : Int) => {x + 1}).collect()  ==> Spark follow Type Infer, so we can remove Int

rdd.map(x => {x + 1}).collect()  ==> {x + 1} is not a very big expression, removed { }

rdd.map(x => x + 1).collect()  ==> we can use _ and remove x

rdd.map(_ + 1).collect()  ==> "_" is place holder, rdd is having a value and that value will place in here 

This is called lambda expression in Scala.

Same logic in Python :

• Can use lambda and perform this operation
• Can write function (so that you can reuse this function)

>>> rdd1 = sc.parallelize([1, 2, 3, 4, 5])

>>> rdd1.collect()

[1, 2, 3, 4, 5]                                                                 

>>> 

>>> rdd1.map(lambda x : x + 1).collect()

[2, 3, 4, 5, 6]                                                                 

>>> 

>>> rdd1.map(lambda _: _ + 1).collect()

[2, 3, 4, 5, 6]                                                                 

>>>

(OR)

def myfunc(x):

    return x +1

rdd1.map(myfunc([1, 2, 3, 4, 5])).collect()

Partitions :

Partitioning is an integral concept in Spark that controls how the data is physically distributed across various nodes in the cluster during data processing.

• It depends on the number of processors you are using in your environment
• Currently my local system is having 2 processors, hence 2
• We can configure this partitions 

Use below methods to get the details of partitions.

scala> rdd3.getNumPartitions

res10: Int = 2

scala> rdd3.partitions.length

res11: Int = 2

Use glom().collect() to see the partition information. BTW, glom() is a transformation.

scala> rdd3.glom().collect()

res14: Array[Array[String]] = Array(Array(2 Hi 2, 3 Hi 3), Array(4 Hi 4, 5 Hi 5, 6 Hi 6))

• Array(Array(2 Hi 2, 3 Hi 3) went to 1st partition while creating RDD
• Array(4 Hi 4, 5 Hi 5, 6 Hi 6)) went to 2nd partition while creating RDD
• We can also manage which data goes to which partition, let's see that later

How to configure number of partitions while creating RDD ?

# Created a RDD (Mentioned num of partitions as 1)

scala> val rdd5 = sc.parallelize(List(1, 2, 3, 4, 5, 6), 1)

rdd5: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[0] at parallelize at <console>:23

# Printing RDD information

scala> rdd5.collect()

res0: Array[Int] = Array(1, 2, 3, 4, 5, 6)                                      

# Checking how many partitions created

scala> rdd5.getNumPartitions

res2: Int = 1

# Visualizing partition information 

scala> rdd5.glom().collect()

res3: Array[Array[Int]] = Array(Array(1, 2, 3, 4, 5, 6))

You can try to increase that partition number to 2, 3, 4, and so  on and see how it work using rdd.glom().collect(). We can always customize it with a number we want.

Let's say, while you are creating RDD, spark mentioned 'm' no of partitions itself. But you felt like that number 'm' is less/more based on your data. In this scenario, we can either increase/decrease this number by creating a new RDD with a new number of partitions(as partitions are immutable, we can't change existing RDD, hence need to create new RDD with required number of partitions). 

coalesce()  Vs repartition() :

• We can use above methods incase if we have to change the default partitions. In Apache Spark, coalesce() is a transformation function used to reduce the number of partitions in a DataFrame or RDD without performing a full shuffle. 
• It is often used for performance optimization, especially in scenarios where you need to decrease the number of partitions for efficient execution.
• Reduces Partitions – Unlike repartition(), which can increase or decrease partitions and involves a full shuffle, coalesce() only reduces the number of partitions without shuffling data across all nodes.
• coalesce() will ONLY merge the partitions when we try to reduce no of partitions, it won't move the data from one partition to other, it will just merge the partition. Moving data in a distributed environment is very costly operation.
• Efficient for Narrow Dependencies – Since it avoids full shuffling, coalesce() is more efficient when you are simply reducing the number of partitions.
• Best for Reducing Partitions after a Filter Operation – When many partitions become empty after a filter, coalesce() helps combine non-empty partitions efficiently.
• coalesce() will accept 2 params, first param is partition number, 2nd param is Boolean
• By default this flag will be false, means don't shuffle(just merge data)
• If we mention true, then instead of merge, data will be shuffled which will be a huge performance impact
• DO NOT set that to True
• repartition() will shuffle the data across the partitions. It is more time taken process, try to avoid using repartition(). Incase if you need to reduce the number of partition, then use coalesce() instead of repartition()

When to Use coalesce() ?

• After filtering a DataFrame where many partitions are empty.
• When writing to disk to avoid small file issues (e.g., df.coalesce(1).write.csv("output")).
• When improving performance by reducing unnecessary partitions.

Basic Actions and Transformations :

• We can use all basic actions like min(), max(), count() etc.
• Instead of using all the above actions, just we stats() as shown below.

import org.apache.spark.SparkContext

import org.apache.spark.SparkConf

val conf = new SparkConf().setAppName("Spark Basics").setMaster("local[*]")

val sc = SparkContext.getOrCreate(conf)

val rdd1 = sc.parallelize(List(1, 2, 3, 4, 5, 6, 7, 8, 9, 10))

rdd1.collect()

// Basic actions

rdd1.min()

rdd1.max()

rdd1.count()

//stats will do all basic transformations

rdd1.stats()

Output :

import org.apache.spark.SparkContext import org.apache.spark.SparkConf conf: org.apache.spark.SparkConf = org.apache.spark.SparkConf@31a471b9 sc: org.apache.spark.SparkContext = org.apache.spark.SparkContext@3ac95f39 rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[25] at parallelize at command-4386469042563736:7 res1: org.apache.spark.util.StatCounter = (count: 10, mean: 5.500000, stdev: 2.872281, max: 10.000000, min: 1.000000)

Map() transformation using lambda functions :

• map() operation comes under 1-1, means it will take one input and convert into another

// Databricks notebook source

// Spark context is an entry point for any operation in Spark

// We need to create it when we are using Databricks approach (but when you use terminal Spark Context will be automaiclly connected)

import org.apache.spark.SparkContext

import org.apache.spark.SparkConf

val conf = new SparkConf().setAppName("Spark Basics").setMaster("local[*]")

val sc = SparkContext.getOrCreate(conf)

val rdd1 = sc.parallelize(List(1, 2, 3, 4, 5, 6))

rdd1.collect()

// Appyling basic transformation

val rdd2 = rdd1.map( x => x + 1)

rdd2.collect()

// Another way of applying transformation using lambda expression

val rdd3 = rdd1.map( _ + 1)

rdd3.collect()

// Multiple transformations

val rdd4 = rdd1.map(x => x + 1).map(x => x + 1).map(x => x + 1).map(x => x + 1).map(x => x + 1)

rdd4.collect()

// Getting partition information

rdd4.getNumPartitions

map() Vs filter() :

import org.apache.spark.SparkContext

import org.apache.spark.SparkConf

val conf = new SparkConf().setAppName("example").setMaster("local[*]")

val sc = SparkContext.getOrCreate(conf)

val rdd1 = sc.parallelize(List(1, 2, 3, 4, 5, 6))

rdd1.collect()

// Output rdd1 = Array[Int] = Array(1, 2, 3, 4, 5, 6)

rdd1.map(x => x + 1).collect()

// Output = res5: Array[Int] = Array(2, 3, 4, 5, 6, 7)

rdd1.filter(x => x % 2 == 1).collect()

// Output = res6: Array[Int] = Array(1, 3, 5)

That's all for this blog. See you again!

Thanks,

Arun Mathe

Email ID : arunkumar.mathe@gmail.com

Contact ID : 9704117111