INTRODUCTION TO SPARK:

SPARK:
-->It is a general purpose
In memory
Compute/process engine

======================== compute engine

Hadoop provides us with the below:
1. Hdfs ==> storage
2. Map reduce ==> process engine
3. YARN ==> resource manager

-->spark is alternative/replacement of map-reduce

Spark can work with the below:

Storage==> Local, hdfs, amazon s3 (cloud)
Process engine ==> Spark
Resource manager: YARN, MESOS, kubernetes

======================== In memory
Why spark is very fast?
MAP-REDUCE: For every MR job, mapper reads the data and writes the
data back to hdfs. It requires 2 disc operations for every mapper. In
general we have hundreds of mappers for each task.
-->Disc I/O is a tedious task due to which the computation more time.

SPARK: For every spark job, the processing happens IN-MEMORY ie

after the data is read from disc it processes the entire data in-memory
and writes the final result to disc.
-->Only 2 disc I/O due for a spark job due to which it is very fast
processing.
=============================General Purpose
Spark can be used for a wide range of activities like cleaning, querying,
streaming data which is all under one roof.

How is data stored in spark?

-->The data is stored in spark in the form of RDD .
-->RDD is a distributed in-memory collection
-->RDD (resilient distributed dataset)
===============================Resilient
What happens when an RDD fails/loses data ?
-->It means ability to get back to normal after a failure
-->We don't have replication as fault tolerance as we deal with in-
memory and it is a costly process.
-->By making use of RDD Lineage graph we try to retrieve for failed
RDD
RDD1
|==> Map
RDD2
| ==> Filter
RDD3
-->If RDD3 fails, then it gets back to its parent RDD2 and asks it to
peform "flatMap" and "filter" again to generate RDD3 which got failed
previously.

Why RDD are immutable?

-->If RDD are mutable ie if we store all the transformation & action on
single RDD.
-->what if if this RDD fails, we just lose all the data?
-->In order to make the RDD resilient we make it immutable
What are the various operations in spark?
1. Transformations:
-->They change the data from one form to another.
-->They are lazy ie they don't execute until an "action" is performed

2. Action:
-->They are in-charge of actually performing the computation
-->They are not lazy and execute when they are called

Why spark is lazy?

1. Spark computation are costly as it is in-memory
2. We need to make sure that we optimize the computation in best
possible way before giving it to spark
 3. They are lazy so they don't compute each & every statement
every time they were created
4. They prepare optimized plan sheet based (DAG)on which they
will compute when an "action" is called

Ex1. To load 2GB file into spark and print 1st line

Ex2. We have to perform filtering of data

-->It perform Predicate pushup to optimize the query plan ie it alters
the steps to be performed
-->spark-shell ==>to run spark commands in cloudera (we get an
interactive scala shell)
-->localhost:4040 ==>UI for spark (only active when spark-shell is
opened)
-->SC (spark context) is the entry point for a spark cluster. We need to
mention it if we have to run our queries on spark cluster and achieve
parallelism

-->we have to load the file, as it is lazy only DAG is created and loading
not yet done
-->when action is performed "collect" then our file is actually loaded
-->flatMap() ==> It takes each line as a input and transforms it
according to our operations. It is an array of arrays(each line stored as
array)
-->Array(Array(I,am,passionate,about,big,data),
Array(I,love,big,data)…..)
-->ie it takes each line as input and splits it based on " " and creates an
array for each line
-->finally all the arrays are merged into single array as seen below.

-->map: works on each input given. For n inputs we get n outputs.

-->reduceByKey: It sorts the data and computes for 2 records at a time.
(am,1)
(am,1) ==>(am,2)
(am,1) ==>(am,3)
(am,1) ==>(am,4)
SPARK PART 2:
What is "checkpointing" in spark?
-->To save the RDD state from lineage graph into a storage so that
resiiency of the RDD is achieved.

What are stages in spark?

-->Each job triggers multiple no. of tasks which can be considered
as a stage.

SPARK STAGES:
1. shuffleMap stage: It is an intermediatary stage in DAG which
provides data for another stage ie shuffling of data takes place.
2. ResultStage: It produces the output for the "action" called which
is the final result given by the spark
 Broadcast Variables:

-->It contains a list of values (small data) which is broadcasted ie

distributed to all the systems for a specific purpose to be met.

Problem: From above ex. We can see redundant words (in, the, a
etc) which are not actually not very helpful.
We collect a list of such words and broadcast them and filter them
out from using "filter" function.

-->similar to map side join in hive

What are Accumulators?

-->It is a shared copy of value kept in driver machine.
-->It is updated by executors but none can read them
-->Similar to counter in map-reduce

How MR jobs were managed before YARN ?

-->Map-Reduce jobs were handled by Job Tracker and Task tracker.

-->Job Tracker:
1. Clients submit the requests to it which it then scheduled to be
distributed to slaves.
2. It schedules the jobs, priortizes jobs and allocates resources to
them
3. It monitors the performance of jobs
4. It single handedly takes care of all these tasks

-->Task Tracker:
1. Inside each block where we have tasks to be performed, it is
tracked by task tracker

Limitations:
1. Scalability: If cluster size increases then load on Job Tracker will
be high and its performance dips.
2. Resource Under-Utilization : Instances where MR slots are fixed
and they are not used to their full potential
3. Only MR jobs can be handled (no graphic processing etc)

Why YARN was introduced?

YARN (Yet Another Resource Negotiator)

-->As the job tracker has to manage huge no. of tasks and resources
which it failed to perform well. We introduced a separate
component just to schedule and manage resouces which is YARN

-->Components of YARN:
1. Resource Manager : Does only scheduling (load is now reduced)
2. Application Master: Does the monitoring and negotiates
resources
3. Node Manager: Manages all the tasks which are assigned to it.

Architecture:

1. Client submits the requests to run its programs to "Resource

Manager"
2. RM creates a container (memory + cpu) and AM on a node to
take care of the job
-->Each job will have an AM
3. AM will request RM for resources (containers) to run its tasks
-->RM will have a priority chart based on which it allocates
-->RM sends container location to AM
4. AM will then start its tasks in containers it got from RM
5. Tasks monitoring is done by AM and handles issues if any
6. Results from tasks are given to node manager
 How Limitations were overcome ?
1. Scalability : Each job can have one AM and no. of jobs can be
easily increased
2. Resource Utilization: No fixed MR slots and containers can provide
dynamic allocation
3. Apart from MR job we can run spark, giraph jobs as well

What happens when job is very small?

Uberization:

-->AM will not request for additional resources from RM

-->It will run the job in its own container

YARN on Spark Cluster:

How to run sparks jobs on spark cluster?

1. Interactive shell: spark-shell
2. Submit jobs: we write code and run jobs using Eclipse

How spark jobs run on spark cluster?

It follows master-slave architecture
1. Master (Driver) : Schedules, monitors, distributes tasks
2. Slave (Executor) : Executes code locally

Client(sends request) --> Driver -->Executors (

)

Which component executes Where?

-->Executors always run on cluster machines (data nodes, in-
memory)
-->Driver runs on either client (client mode) or cluster (cluster
mode)
-->Cluster mode is preferrable as it doesn't depend on client
machines (mainly real-time production)

Who controls the spark cluster?

It can be done by few of below "cluster managers"
1. YARN , Mesos, Kubernetes

YARN Architecture in Spark:

1. Spark session : Entry point into spark cluster where driver

provides location of executors where processing will happen

Yarn in client mode:

1. When spark job launched then Spark session created

2. Then Yarn RM creates AM on node manager
3. AM requests resources if req and launch executors in them
4. Driver and executors can communicate directly
Yarn in cluster mode:

1.The only difference is that Driver runs on AM

How to convert collections into RDD?

-->when we have Array, List, Set of values.
-->we can use "parallelize" method to convert them into RDD
-->val listOrder=List("mobile", "garment","purse")
sc.parallelize(listOrder)

What are types of transformations?

Narrow Transformations: Data shuffling is not required ie moving

data from one machine to another
Ex. Map(), flatMap() , filter()
-->when we perform map(), all the data resides in one machine and
we need not move it across various machines.
Val a="Vaishu is a big data developer" map(a=>a.split(" ") ==>
Array[vaishu is a big data developer]
Wide Transformations: Data shuffling in involved ie data is moved
across machines.
-->It works mainly on pair RDD
Ex. reduceByKey, groupByKey()

P1=({vaish,1},{big,1},{love,1})
P2=({love,1},{vaish,1})
|
|
P1={{vaish,2},{big,1}}
P2=({love,2})
Now, data has to be moved across the partitions to compute the
final result

What are stages in spark?

-->Based on DAG we can see the stages of execution in spark.
-->Each wide transformation invokes a separate stage as data
shuffling is required
-->If we have 2 stages then we have 1 wide transformation for sure
-->Stage 1: Processes the input and writes it to the disk
-->Stage 2: From disk wide transformation takes it and gives the
final output

Difference btw reduceByKey() and reduce() :

-->reduceByKey() : It is a transformation ie one rdd is converted

into another.
-->It works on only pair rdd
Pair RDD: It has tuple of 2 elements Ex. (vaish,1) ==>like (key,value)
pair

-->reduce() : it is an action ie rdd is converted into local variable.

-->It gives us a single output value.
Ex. Val a=1 to 100 a.reduce((x,y)=>x+y)
Output:5050 (it just gives sum of all numbers)

reduceByKey() Reduce()
1. It is a 1. It is an action
transformation 2. It gives a single
2. It works on output value
only pair RDD
3. It gives
multiple
output values

Why reduceByKey() is transformation and not action?

-->reduceByKey() gives us a list of output and not a single output

-->It can be used again for other transformation

reduceByKey() vs groupByKey() :

reduceByKey() groupByKey()
1. Each partition 1. Each partition will
will process just group the similar
the data and inputs and give it to
give the output reducer machine
to reducer 2. More processing at
machine reducer
2. Less 3. Less parallelism
processing at 4. Does not perform
reducer local aggregation
 3. More 5. If data is huge then
parallelism reduce can't single
4. Performs local handedly work on it
aggregation due to which we
should never use
groupByKey()

reduceByKey():

groupByKey:
 Practical: groupByKey()

File size ~ 350mb Tasks=11 (350/12)

-->File compressed and sent to reducers. As we can see only 2 reducers

are doing all the work and all others are idle. It is not an idle situation
as we need to make the best use of all tasks. Due to this groupByKey()
is not recommended
What is a Pair RDD:

1. RDD with a tuple of 2 elements

("vaish",10) (20,30)
2. Pair RDD is not a map as map can have only unique keys
3. It works on transformations like reduceByKey() and groupByKey()

What happens when multiple actions are called?

Val mapIn=X.map(x=>x,1)
Val redIn= mapIn.reduceByKey(x,y=>x+y)

Val resultFinal=redIn.collect() ==> action 1 (processes entire data)

Val finalResult=redIn.count ==>action2 (spark stores the results
upto last transformation and uses it directly)

-->Every action triggers a job in spark

-->Every job processes the data right from beginning
 -->** But spark optimizes it and holds the results of processing
done previously which can be used for the next action directly
rather than processing from the beginning.

Properties in spark:

1. For data collections

1.1 sc.defaultparallelism => It gives us the default no. of partitions
created for the rdd's

1. For Hdfs data

2.1 sc.defaultMinPartitions => It determines min no. of partitions
returned by rdd
Ex. File~ 100mb (actually 1 partition)
But due to above property it returns 2
2.2 sc.getNumPartitions => It gives no. of partitions for a particular
rdd
This property takes precedence over defaultMinPartitions

1. Repartition vs Coalsence

Repartition Coalesce
1. Used to either 1.Used to only decrease no.
increase/decrease no. of of partitions
partitions for rdd's 2.If we increase partitions
Val no changes will be made
newRdd=inputFile.repartition(1 and same old value remains
0)
Val Val
newRdd=inputFile.repartition(1 newRdd=inputFile.coalesce(
) 2)
2. It is a wide
transformation as data
shuffling is involved
1. Use repartition for 1. Use coalesce for
INCREASE of partitions DECREASE of
partitions
It combines partitions
on same machines due
to which shuffling is
less and optimized

Before:
Node1: p1 p2
Node 2: p3 p4
Node3: p5 p6

After:
Node1: p1(p1+p2)
Node2: p3(p3+p4)
Node3:p5(p5+p6)

2. Don't use
REPARTIION for
DECREASE
Repartition tries to
equally distribute the
data among all
partitions due to
 which more shuffling
is required

**
1. No. of actions = No. of jobs= No. of AM
2. No. of tasks= No. of partitions
3. 1 block size in local = 1rdd partition=32mb
4. 1 rdd= n partitions
5. 1 task = 1 partition