"BIG DATA ANALYTICS"

(15CS82 CBCS as per VTU Syllabus)

Module-1
Hadoop Distributed File System (HDFS) Basics
Hadoop Distributed File System (HDFS) is a distributed file system which is designed to run on
commodity hardware. Commodity hardware is cheaper in cost. Since Hadoop requires processing
 power of multiple machines and since it is expensive to deploy costly hardware, we use commodity
hardware. When commodity hardware is used, failures are more common rather than an exception.
HDFS is highly fault-tolerant and is designed to run on commodity hardware.
HDFS provides high throughput access to the data stored. So it is extremely useful when you want
to build applications which require large data sets.
HDFS was originally built as infrastructure layer for Apache Nutch. It is now pretty much part of
Apache Hadoop project.

Fig 1.1 HDFS ARchitecture

HDFS has master/slave architecture. In this architecture one of the machines will be designated as
a master node (or name node). Every other machine would be acting as slave (or data node).
 NameNode/DataNode are java processes that run on the machines when Hadoop software is
installed.
 NameNode is responsible for managing the metadata about the HDFS Files. This metadata
includes various information about the HDFS File such as Name of the file, File Permissions,
FileSize, Blocks etc. It is also responsible for performing various namespace operations like
opening, closing, renaming the files or directories.
Whenever a file is to be stored in HDFS, it is divided into blocks. By default, blocksize is 64MB
(Configurable). These blocks are replicated (default is 3) and stored across various datanodes to
take care of hardware failures and for faster data transfers. NameNode maintains a mapping of
 blocks to DataNodes.

1
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

DataNodes serves the read and write requests from HDFS file system clients. They are also
responsible for creation of block replicas and for checking if blocks are corrupted or not. It sends
the ping messages to the NameNode in the form of block mappings.

How communication happens?

1. HDFS exposes Java/C API using which user can write an application to interact with HDFS.
Application using this API Interacts with Client Library (present on the same client machine).
2. Client (Library) connects to the NameNode using RPC. The communication between them
happens using ClientProtocol. Major functionality in ClientProtocol includes Create (creates a file
in name space), Append (add to the end of already existing file), Complete (client has finished
writing to file), Read etc.
3. Client (Library) interacts with DataNode directly using DataTransferProtocol. The
DataTransferProtocol defines operations to read a block, write to block, get checksum of block,
copy the block etc.
4. Interaction between NameNode and DataNode. It‘s always Data Node which initiates the
communication first and NameNode just responds to the requests intiated. The communication
usually involves DataNode Registration, DataNode sending heart beat messages, DataNode
sending blockreport, DataNode notifying the receipt of Block from a client or another DataNode
during replication of blocks.
Assumptions and Goals
1. Hardware Failure Hardware Failure is the norm rather than the exception. The entire HDFS
file system may consist of hundreds or thousands of server machines that stores pieces of file
system data. The fact that there are a huge number of components and that each component has a
non-trivial probability of failure means that some component of HDFS is always non-functional.
Therefore, detection of faults and automatically recovering quickly from those faults are core
architectural goals of HDFS.
2. Streaming Data Access: Applications that run on HDFS need streaming access to their data
sets. They are not general purpose applications that typically run on a general purpose file system.
HDFS is designed more for batch processing rather than interactive use by users. The emphasis is
on throughput of data access rather than latency of data access. POSIX imposes many hard
requirements that are not needed for applications that are targeted for HDFS. POSIX semantics in a
few key areas have been traded off to further enhance data throughout rates.
3. Large Data Sets: Applications that run on HDFS have large data sets. This means that a typical
file in HDFS is gigabytes to terabytes in size. Thus, HDFS is tuned to support large files. It should
 provide high aggregate data bandwidth and should scale to hundreds
h undreds of nodes in a single cluster. It
should support tens of millions of files in a single cluster.
4. Simple Coherency Model: Most HDFS applications need write-once-read-many access model
for files. A file once created, written and closed need not be changed. This assumption simplifies
data coherency issues and enables high throughout data access. A Map-Reduce application or a
Web-Crawler application fits perfectly with this model. There is a plan to support appending-writes
to a file in future.

2
  BIG DATA ANALYTICS"
“ 

(15CS82 CBCS as per VTU Syllabus)

5. Moving computation is cheaper than moving data: A computation requested by an

application is most optimal if the computation can be done near where the data is located. This is
especially true when the size of the data set is huge. This eliminates network congestion and
increase overall throughput of the system. The assumption is that it is often better to migrate the
computation closer to where the data is located rather than moving the data to where the
application is running. HDFS provides interfaces for applications to move themselves closer to
where the data is located.
6. Portability across Heterogeneous Hardware and Software Platforms: HDFS should be
designed in such a way that it is easily portable from one platform to another. This facilitates
widespread adoption of HDFS as a platform
pl atform of choice for a large set of applications.

 Namenode and Datanode

HDFS has a master/slave architecture. An HDFS cluster consists of a single Namenode, a master
server that manages the filesystem namespace and regulates access to files by clients. In addition,
there are a number of Datanodes, one per node in the cluster, which manage storage attached to the
nodes that they run on. HDFS exposes a file system namespace and allows user data to be stored in
files. Internally, a file is split into one or more blocks and these blocks are stored in a set of
Datanodes. The Namenode makes filesystem namespace operations like opening, closing,
renaming etc. of files and directories. It also determines the mapping of blocks to Datanodes. The
Datanodes are responsible for serving read and write requests from filesystem clients. The
Datanodes also perform block creation, deletion, and replication upon instruction from the
 Namenode. The Namenode and Datanode are pieces of software that run on commodity machines.ma chines.
These machines are typically commodity Linux machines. HDFS is built using the Java language;
any machine that support Java can run the Namenode or the Datanode. Usage of the highly
 portable Java language means that HDFS can be deployed on a wide range of o f machines. A typical
deployment could have a dedicated machine that runs only the Namenode software. Each of the
other machines in the cluster runs one instance of the Datanode software. The architecture does not
 preclude running multiple Datanodes on the same machine but in a real-deployment that is never
the case. The existence of a single Namenode in a cluster greatly simplifies the architecture of the
system. The Namenode is the arbitrator and repository for all HDFS metadata. The system is
designed in such a way that user data never flows through the Namenode.

The File System Namespace

HDFS supports a traditional hierarchical file organization. A user or an application can create
directories and store files inside these directories. The file system namespace hierarchy is similar to
most other existing file systems. One can create anda nd remove files, move a file from one directory to
another, or rename a file. HDFS does not yet implement user quotas and access permissions.
HDFS does not support hard links and soft links. However, the HDFS architecture does not
 preclude implementing these features at a later time. The Namenode maintains the file system
namespace. Any change to the file system namespace and properties are recorded by the
 Namenode. An application can specify the number of replicas of a file that th at should be maintained
 by HDFS. The number of copies of a file is called the replication factor of that file. This
information is stored by the Namenode.

3
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

Data Replication
HDFS is designed to reliably store very large files across machines in a large cluster. It stores each
file as a sequence of blocks; all blocks in a file except the last block are the same size. Blocks
 belonging to a file are replicated for fault tolerance. The block size and replication factor are
configurable per file. Files in HDFS are write-once and have strictly one writer at any time. An
application can specify the number of replicas of a file. The replication factor can be specified at
file creation time and can be changed later. The Namenode makes all decisions regarding
replication of blocks. It periodically receives Heartbeat and a Blockreport from each of the
Datanodes in the cluster. A receipt of a heartbeat implies that the Datanode is in good health and is
serving data as desired. A Blockreport contains a list of all blocks on that Datanode.

Replica Placement the First Baby Steps

The selection of placement of replicas is critical to HDFS reliability and performance. This feature
distinguishes HDFS from most other distributed file systems. This is a feature that needs lots of
tuning and experience. The purpose of a rack-aware replica placement is to improve data
reliability, availability, and network bandwidth utilization. The current implementation for the
replica placement policy is a first effort in this direction. The short-term goals of implementing this
 policy are to validate it on production systems, learn more about its behavior and build a
foundation to test and research more sophisticated policies in the future. HDFS runs on a cluster of
computers that spread across many racks. Communication between two nodes on different racks
has to go through switches. In most cases, network bandwidth between two machines in the same
rack is greater than network bandwidth between two machines on different racks.
At startup time, each Datanode determines the rack it belongs to and notifies the Namenode of the
rack id upon registration. HDFS provides APIs to facilitate pluggable modules that can be used to
determine the rack identity of a machine. A simple but non-optimal policy is to place replicas
across racks. This prevents losing data when an entire rack fails and allows use of bandwidth from
multiple racks when reading data. This policy evenly distributes replicas in the cluster and thus
makes it easy to balance load on component failure. However, this policy increases the cost of
writes because a write needs to transfer blocks to multiple racks.
For the most common case when the replica factor is three, HDFS.s placement policy is to place
one replica on the local node, place another replica on a different node at the local rack, and place
the last replica on different node at a different rack. This policy cuts the inter-rack write traffic and
improves write performance. The chance of rack failure is far less than that of node failure; this
 policy does not impact data reliability and availability guarantees. But it reduces the aggregate
network bandwidth when reading data since a block is placed in only two unique racks rather than
three. The replicas of a file do not evenly distribute across the racks. One third of replicas are on
one node, two thirds of the replicas are on one rack; the other one third of replicas is evenly
distributed across all the remaining racks. This policy improves write performance while not
impacting data reliability or read performance. The implementation of the above policy is work-in-
 progress.
Replica Selection
HDFS tries to satisfy a read request from a replica that is closest to the reader. If there exists a
replica on the same rack as the reader node, then that replica is preferred to satisfy the read request.

4
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

If a HDFS cluster spans multiple data centers, then a replica that is resident in the local data center
is preferred over remote replicas.
SafeMode
On startup, the Namenode enters a special state called Safemode. Replication of data blocks does
not occur when the Namenode is in Safemode state. The Namenode receives Heartbeat and
Blockreport from the Datanodes. A Blockreport contains the list of data blocks that a Datanode
reports to the Namenode. Each block has a specified minimum number of replicas. A block is
considered safely-replicated when the minimum number of replicas of that data block has checked
in with the Namenode. When a configurable percentage of safely-replicated data blocks checks in
with the Namenode (plus an additional 30 seconds), the Namenode exits the Safemode state. It
then determines the list of data blocks (if any) that have fewer than the specified number of
replicas. The Namenode then replicates these blocks to other Datanodes.

The Persistence of File System Metadata

The HDFS namespace is stored by the Namenode. The Namenode uses a transaction log called the
EditLog to persistently record every change that occurs to file system metadata. For example,
creating a new file in HDFS causes the Namenode to insert a record into the EditLog indicating
this change. Similarly, changing the replication factor of a file causes a new record to be inserted
into the EditLog. The Namenode uses a file in its local file system to store the Edit Log. The entire
file system namespace, the mapping of blocks to files and filesystem properties are stored in a file
called the FsImage. The FsImage is a file in the Namenode‘s local file system too.
The Namenode has an image of the entire file system namespace and file Blockmap in memory.
This metadata is designed to be compact, so that a 4GB memory on the Namenode machine is
 plenty to support a very large number of files and directories. When the Namenode starts up, it
reads the FsImage and EditLog from disk, applies all the transactions from the EditLog into the in-
memory representation of the FsImage and then flushes out this new metadata into a new FsImage
on disk. It can then truncate the old EditLog because its transactions have been applied to the
 persistent FsImage. This process is called a checkpoint. In the current implementation, a
checkpoint occurs when the Namenode starts up. Work is in progress to support periodic
checkpointing in the near future.
The Datanode stores HDFS data into files in its local file system. The Datanode has no knowledge
about HDFS files. It stores each block of HDFS data in a separate file in its local file system. The
Datanode does not create all files in the same directory. Instead, it uses a heuristic to determine the
optimal number of files per directory. It creates subdirectories appropriately. It is not optimal to
create all local files in the same directory because the local file system might not be able to
efficiently support a huge number of files in a single directory. When a Datanode starts up, it scans
through its local file system, generates a list of all HDFS data blocks that correspond to each of
these local files and sends this report to the Namenode. This report is called the Blockreport.

The Communication Protocol

All communication protocols are layered on top of the TCP/IP protocol. A client establishes a
connection to a well-defined and configurable port on the Namenode machine. It talks the
ClientProtocol with the Namenode. The Datanodes talk to the Namenode using the
DatanodeProtocol. The details on these protocols will be explained later on. A Remote Procedure

5
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

Call (RPC) abstraction wraps the ClientProtocol and the DatanodeProtocol. By design, the
 Namenode never initiates an RPC. It responds to RPC requests issued by a Datanode or a client.

Robustness
The primary objective of HDFS is to store data reliably even in the presence of failures. The three
types of common failures are Namenode failures, Datanode failures and network partitions.

Data Disk Failure, Heartbeats and Re-Replication

A Datanode sends a heartbeat message to the Namenode periodically. A network partition can
cause a subset of Datanodes to lose connectivity with the Namenode. The Namenode detects this
condition be a lack of heartbeat message. The Namenode marks these Datanodes as dead and does
not forward any new IO requests to these Datanodes. The data that was residing on those
Datanodes are not available to HDFS any more. This may cause the replication factor of some
 blocks to fall below their specified value. The Namenode determines all the blocks that need to be
replicated and starts replicating them to other Datanodes. The necessity for re-replication may arise
due to many reasons: a Datanode becoming unavailable, a corrupt replica, a bad disk on the
Datanode or an increase of the replication factor of a file.

Cluster Rebalancing
The HDFS architecture is compatible with data rebalancing schemes. It is possible that data may
move automatically from one Datanode to another if the free space on a Datanode falls below a
certain threshold. Also, a sudden high demand for a particular file can dynamically cause creation
of additional replicas and rebalancing of other data in the cluster. These types of rebalancing
schemes are not yet implemented.

Data Correctness
It is possible that a block of data fetched from a Datanode is corrupted. This corruption can occur
 because of faults in the storage device, a bad network or buggy software. The HDFS client
implements checksum checking on the contents of a HDFS file. When a client creates a HDFS file,
it computes a checksum of each block on the file and stores these checksums in a separate hidden
file in the same HDFS namespace. When a client retrieves file contents it verifies that the data it
received from a Datanode satisfies the checksum stored in the checksum file. If not, then the client
can opt to retrieve that block from another Datanode that has a replica of that block.

Metadata Disk Failure

The FsImage and the EditLog are central data structures of HDFS. A corruption of these files can
cause the entire cluster to be non-functional. For this reason, the Namenode can be configured to
support multiple copies of the FsImage and EditLog. Any update to either the FsImage or EditLog
causes each of the FsImages and EditLogs to get updated synchronously. This synchronous
updating of multiple EditLog may degrade the rate of namespace transactions per second that a
 Namenode can support. But this degradation is acceptable because HDFS applications are very
data intensive in nature; they are not metadata intensive. A Namenode, when it restarts, selects the
latest consistent FsImage and EditLog to use. The Namenode machine is a single point of failure

6
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

for the HDFS cluster. If a Namenode machine fails, manual intervention is necessary. Currently,
automatic restart and failover of the Namenode software to another machine is not supported.

Snapshots
Snapshots support storing a copy of data at a particular instant of time. One usage of the snapshot-
feature may be to roll back a corrupted cluster to a previously known good point in time. HDFS
current does not support snapshots but it will be supported it in future release.

Data Blocks
HDFS is designed to support large files. Applications that are compatible with HDFS are those that
deal with large data sets. These applications write the data only once; they read the data one or
more times and require that reads are satisfied at streaming speeds. HDFS supports write-once-
read-many semantics on files. A typical block size used by HDFS is 64 MB. Thus, a HDFS file is
chopped up into 128MB chunks, and each chunk could reside in different Datanodes.

Staging
A client-request to create a file does not reach the Namenode immediately. In fact, the HDFS client
caches the file data into a temporary local file. An application-write is transparently redirected to
this temporary local file. When the local file accumulates data worth over a HDFS block size, the
client contacts the Namenode. The Namenode inserts the file name into the file system hierarchy
and allocates a data block for it. The Namenode responds to the client request with the identity of
the Datanode(s) and the destination data block. The client flushes the block of data from the local
temporary file to the specified Datanode. When a file is closed, the remaining un-flushed data in
the temporary local file is transferred to the Datanode. The client then instructs the Namenode that
the file is closed. At this point, the Namenode commits the file creation operation into a persistent
store. If the Namenode dies before the file is closed, the file is lost. The above approach has been
adopted after careful consideration of target applications that run on HDFS. Applications need
streaming writes to files. If a client writes to a remote file directly without any client side
 buffering, the network speed and the congestion in the network impacts throughput considerably.
This approach is not without precedence either. Earlier distributed file system, e.g. AFS have used
client side caching to improve performance. A POSIX requirement has been relaxed to achieve
higher performance of data uploads.

Pipelining
When a client is writing data to a HDFS file, its data is first written to a local file as explained
above. Suppose the HDFS file has a replication factor of three. When the local file accumulates a
 block of user data, the client retrieves a list of Datanodes from the Namenode. This list represents
the Datanodes that will host a replica of that block. The client then flushes the data block to the
first Datanode. The first Datanode starts receiving the data in small portions (4 KB), writes each
 portion to its local repository and transfers that portion to the second Datanode in the list. The
second Datanode, in turn, starts receiving each portion of the data block, writes that portion to its
repository and then flushes that portion to the third Datanode. The third Datanode writes the data to

7
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

10 line of pig latin = approx. 200 lines of Map-Reduce Java code

The compiler internally converts pig latin to MapReduce. It produces a sequential set of
MapReduce jobs, and that‘s an abstraction (which works like black box). PIG was initially
developed by Yahoo. It gives you a platform for building data flow for ETL (Extract, Transform
and Load), processing and analyzing huge data sets. In PIG, first the load command, loads the
data. Then we perform various functions on it like grouping, filtering, joining, sorting, etc. At last,
either you can dump the data on the screen or you can store the result back in HDFS

Apache Pig has several usage modes. The first is a local mode in which all processing is done on
the local machine. The non-local (cluster) modes are MapReduce and Tez. These modes execute
the job on the cluster using either the MapReduce engine or the optimized Tez engine. There are
also interactive and batch modes available; they enable Pig applications to be developed locally in
interactive modes, using small amounts of data, and then run at scale on the cluster in a production
mode.

USING APACHE HIVE

Facebook created HIVE for people who are fluent with SQL. Thus, HIVE makes them feel at home
while working in a Hadoop Ecosystem. Basically, HIVE is a data warehousing component which
 performs reading, writing and managing large data sets in a distributed environment using SQL-
like interface.

 HIVE + SQL = HQL

The query language of Hive is called Hive Query Language(HQL), which is very similar like SQL.
It has 2 basic components: Hive Command Line and JDBC/ODBC driver. The Hive Command
line interface is used to execute HQL commands. While, Java Database Connectivity (JDBC) and
Object Database Connectivity (ODBC) is used to establish connection from data storage.

Secondly, Hive is highly scalable. As, it can serve both the purposes, i.e. large data set processing
(i.e. Batch query processing) and real time processing (i.e. Interactive query processing). It
supports all primitive data types of SQL. You can use predefined functions, or write tailored user
defined functions (UDF) also to accomplish your specific needs.

USING APACHE SQOOP TO ACQUIRE RELATIONAL DATA

Sqoop is a tool designed to transfer data between Hadoop and relational databases. You can use
Sqoop to import data from a relational database management system (RDBMS) into the Hadoop
Distributed File System (HDFS), transform the data in Hadoop, and then export the data back into
an RDBMS.

14
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

Sqoop can be used with any Java Database Connectivity (JDBC) – compliant database and has been
tested on Microsoft SQL Server, PostgresSQL, MySQL, and Oracle

When we submit Sqoop command, our main task gets divided into sub tasks which is handled by
individual Map Task internally. Map Task is the sub task, which imports part of data to the Hadoop
Ecosystem. Collectively, all Map tasks imports the whole data

Fig 2.2: Sqoop WorkFlow

Export also works in a similar manner.

When we submit our Job, it is mapped into Map Tasks which brings the chunk of data from
HDFS. These chunks are exported to a structured data destination. Combining all these exported
chunks of data, we receive the whole data at the destination, which in most of the cases is an
RDBMS (MYSQL/Oracle/SQL Server).

APACHE SOLR & LUCENE

Apache Solr and Apache Lucene are the two services which are used for searching and indexing in
Hadoop Ecosystem.

• Apache Lucene is based on Java, which also helps in spell checking.

• If Apache Lucene is the engine, Apache Solr is the car built around it. Solr is a complete
application built around Lucene.
• It uses the Lucene Java search library as a core for search and full indexing.

15
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

APACHE AMBARI

Ambari is an Apache Software Foundation Project which aims at making Hadoop ecosystem more
manageable. It includes software for provisioning, managing and monitoringApache Hadoop
clusters.

The Ambari provides:

1. Hadoop cluster provisioning:
▪ It gives us step by step process for installing Hadoop services across a number of
hosts.
▪ It also handles configuration of Hadoop services over a cluster.
2. Hadoop cluster management:
▪ It provides a central management service for starting, stopping and re-configuring
Hadoop services across the cluster.
3. Hadoop cluster monitoring:
▪ For monitoring health and status, Ambari provides us a dashboard.
▪ The Amber Alert framework is an alerting service which notifies the user,
whenever the attention is needed. For example, if a node goes down or low disk
space on a node, etc.

USING APACHE FLUME TO ACQUIRE DATA STREAMS

Apache Flume is an independent agent designed to collect, transport, and store data into HDFS.
Often data transport involves a number of Flume agents that may traverse a series of machines and
locations. Flume is often used for log files, social media-generated data, email messages, and just
about any continuous data source.
As shown in Figure 2.3, a Flume agent is composed of three components.

Fig 2.3. Flume agent with source, channel, and sink

16
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

Fig 2.6 YARN architecture with two clients (MapReduce and MPI).

YARN APPLICATION FRAMEWORKS

YARN presents a resource management platform, which provides services such as scheduling,
fault monitoring, data locality, and more to MapReduce and other frameworks. Figure 7 illustrates
some of the various frameworks that will run under YARN. Note that the Hadoop version 1
applications (e.g., Pig and Hive) run under the MapReduce framework.

Fig 2.7 Example of the Hadoop version 2 ecosystem.

21
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

Distributed-Shell
As described earlier in this chapter, Distributed-Shell is an example application included with the
Hadoop core components that demonstrates how to write applications on top of YARN. It provides
a simple method for running shell commands and scripts in containers in parallel on a Hadoop
YARN cluster.

Hadoop MapReduce
MapReduce was the first YARN framework and drove many of YARN‘s requirements. It is
integrated tightly with the rest of the Hadoop ecosystem projects, such as Apache Pig, Apache
Hive, and Apache Oozie.

Apache Tez
One great example of a new YARN framework is Apache Tez. Many Hadoop jobs involve the
execution of a complex directed acyclic graph (DAG) of tasks using separate MapReduce stages.
Apache Tez generalizes this process and enables these tasks to be spread across stages so that they
can be run as a single, all-encompassing job. Tez can be used as a MapReduce replacement for
 projects such as Apache Hive and Apache Pig. No changes are needed to the Hive or Pig
applications.

Apache Giraph
Apache Giraph is an iterative graph processing system built for high scalability. Facebook, Twitter,
and LinkedIn use it to create social graphs of users. Giraph was originally written to run on
standard Hadoop V1 using the MapReduce framework, but that approach proved inefficient and
totally unnatural for various reasons.. In addition, using the flexibility of YARN, the Giraph
developers plan on implementing their own web interface to monitor job progress.

Hoya: HBase on YARN

The Hoya project creates dynamic and elastic Apache HBase clusters on top of YARN. A client
application creates the persistent configuration files, sets up the HBase cluster XML files, and then
asks YARN to create an ApplicationMaster. YARN copies all files listed in the client‘s
application-launch request from HDFS into the local file system of the chosen server, and then
executes the command to start the Hoya ApplicationMaster. Hoya also asks YARN for the number
of containers matching the number of HBase region servers it needs.

Apache Spark
Spark was initially developed for applications in which keeping data in memory improves
 performance, such as iterative algorithms, which are common in machine learning, and interactive
data mining. Spark differs from classic MapReduce in two important ways. First, Spark holds

22
 "BIG DATA ANALYTICS"
(15CS82 CBCS as per VTU Syllabus)

intermediate results in memory, rather than writing them to disk. Second, Spark supports more than
 just MapReduce functions; that is, it greatly expands the set of possible analyses that can be
executed over HDFS data stores. It also provides APIs in Scala, Java, and Python.
Since 2013, Spark has been running on production YARN clusters at Yahoo!. The advantage of
 porting and running Spark on top of YARN is the common resource management and a single
underlying file system

Apache Storm
Traditional MapReduce jobs are expected to eventually finish, but Apache Storm continuously
 processes messages until it is stopped. This framework is designed to process unbounded streams
of data in real time. It can be used in any programming language. The basic Storm use-cases
include real-time analytics, online machine learning, continuous computation, distributed RPC
(remote procedure calls), ETL (extract, transform, and load), and more. Storm provides fast
 performance, is scalable, is fault tolerant, and provides processing guarantees. It works directly
under YARN and takes advantage of the common data and resource management substrate.

Apache REEF: Retainable Evaluator Execution Framework

YARN‘s flexibility sometimes requires significant effort on the part of application implementers.
The steps involved in writing a custom application on YARN include building your own
ApplicationMaster, performing client and container management, and handling aspects of fault
tolerance, execution flow, coordination, and other concerns. The REEF project by Microsoft
recognizes this challenge and factors out several components that are common to many
applications, such as storage management, data caching, fault detection, and checkpoints.
Framework designers can build their applications on top of REEF more easily than they can build
those same applications directly on YARN, and can reuse these common services/libraries. REEF‘s
design makes it suitable for both MapReduce and DAG-like executions as well as iterative and
interactive computations.

Hamster: Hadoop and MPI on the Same Cluster

The Message Passing Interface (MPI) is widely used in high-performance computing (HPC). MPI
is primarily a set of optimized message-passing library calls for C, C++, and Fortran that operate
over popular server interconnects such as Ethernet and InfiniBand. Because users have full control
over their YARN containers, there is no reason why MPI applications cannot run within a Hadoop
cluster. The Hamster effort is a work-in-progress that provides a good discussion of the issues
involved in mapping MPI to a YARN cluster.

Apache Flink: Scalable Batch and Stream Data Processing

Apache Flink is a platform for efficient, distributed, general-purpose data processing. It features
 powerful programming abstractions in Java and Scala, a high-performance run time, and automatic

23