Big data technologies : A survey

Big Data technologies: A survey
Ahmed Oussous a
, Fatima-Zahra Benjelloun a
, Ayoub Ait Lahcen a,b,⇑
, Samir Belfkih a
a
LGS, National School of Applied Sciences (ENSA), Ibn Tofail University, Kenitra, Morocco
b
LRIT, Unité associée au CNRST URAC 29, Mohammed V University in Rabat, Morocco
a r t i c l e i n f o
Article history:
Received 2 January 2017
Revised 3 June 2017
Accepted 4 June 2017
Available online xxxx
Keywords:
Big Data
Hadoop
Big Data distributions
Big Data analytics
NoSQL
Machine learning
a b s t r a c t
Developing Big Data applications has become increasingly important in the last few years. In fact, several
organizations from different sectors depend increasingly on knowledge extracted from huge volumes of
data. However, in Big Data context, traditional data techniques and platforms are less efficient. They show
a slow responsiveness and lack of scalability, performance and accuracy. To face the complex Big Data
challenges, much work has been carried out. As a result, various types of distributions and technologies
have been developed. This paper is a review that survey recent technologies developed for Big Data. It
aims to help to select and adopt the right combination of different Big Data technologies according to
their technological needs and specific applications’ requirements. It provides not only a global view of
main Big Data technologies but also comparisons according to different system layers such as Data
Storage Layer, Data Processing Layer, Data Querying Layer, Data Access Layer and Management Layer.
It categorizes and discusses main technologies features, advantages, limits and usages.
Ó 2017 The Authors. Production and hosting by Elsevier B.V. on behalf of King Saud University. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2. Background. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2.1. Big Data definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2.2. Big Data applications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3. Big Data challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.1. Big Data management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.2. Big Data cleaning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.3. Big Data aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.4. Imbalanced systems capacities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.5. Imbalanced Big Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.6. Big Data analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.7. Big Data machine learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.7.1. Data Stream learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.7.2. Deep learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.7.3. Incremental and ensemble learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.7.4. Granular computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
http://dx.doi.org/10.1016/j.jksuci.2017.06.001
1319-1578/Ó 2017 The Authors. Production and hosting by Elsevier B.V. on behalf of King Saud University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
⇑ Corresponding author at: LGS, National School of Applied Sciences (ENSA), Ibn Tofail University, Kenitra, Morocco.
E-mail addresses: ahmed.oussous@outlook.com (A. Oussous), fatimazahra.benjelloun@outlook.com (F.-Z. Benjelloun), ayoub.aitlahcen@univ-ibntofail.ac.ma
(A. Ait Lahcen), samir.belfkih@univ-ibntofail.ac.ma (S. Belfkih).
Peer review under responsibility of King Saud University.
Production and hosting by Elsevier
Journal of King Saud University – Computer and Information Sciences xxx (2017) xxx–xxx
Contents lists available at ScienceDirect
Journal of King Saud University –
Computer and Information Sciences
journal homepage: www.sciencedirect.com
Please cite this article in press as: Oussous, A., et al. Big Data technologies: A survey. Journal of King Saud University – Computer and Information Sciences
(2017), http://dx.doi.org/10.1016/j.jksuci.2017.06.001

4. Big Data and Hadoop ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.1. Hadoop capabilities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.2. Data Storage Layer: HDFS and HBase. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.2.1. Hadoop Distributed File System (HDFS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.2.2. HBase. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.3. Data Processing Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.3.1. MapReduce programming model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.3.2. YARN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.3.3. Cascading: a MapReduce framework for complex flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.4. Data Querying Layer: Pig, JAQL and Hive. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.5. Data Access Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.5.1. Data Ingestion: Sqoop, Flume and Chukwa. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.5.2. Data streaming: storm and spark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.5.3. Storage Management: HCatalog. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.6. Data analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.7. Management Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.7.1. Coordination and Workflow: Zookeeper, Avro and Oozie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.7.2. System Deployment: Ambari, Whirr, BigTop and Hue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.8. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5. Hadoop distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.1. Cloudera. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.2. Hortonworks Data Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.3. Amazon Elastic MapReduce (EMR). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.4. MapR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.5. IBM InfoSphere BigInsights. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.6. GreenPlum’s Pivotal HD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.7. Oracle Big Data appliance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.8. Windows Azure HDInsight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.9. Choosing a Hadoop distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5.10. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
6. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
1. Introduction
Nowadays, large data volumes are daily generated at unprece-
dented rate from heterogeneous sources (e.g., health, government,
social networks, marketing, financial). This is due to many techno-
logical trends, including the Internet Of Things, the proliferation of
the Cloud Computing (Botta et al., 2016) as well as the spread of
smart devices. Behind the scene, powerful systems and distributed
applications are supporting such multiple connections systems
(e.g., smart grid systems (Chen et al., 2014a), healthcare systems
(Kankanhalli et al., 2016), retailing systems like that of Walmart
(Schmarzo, 2013), government systems (Stoianov et al., 2013), etc.).
Previously to Big Data revolution, companies could not store all
their archives for long periods nor efficiently manage huge data
sets. Indeed, traditional technologies have limited storage capacity,
rigid management tools and are expensive. They lack of scalability,
flexibility and performance needed in Big Data context. In fact, Big
Data management requires significant resources, new methods and
powerful technologies. More precisely, Big Data require to clean,
process, analyze, secure and provide a granular access to massive
evolving data sets. Companies and industries are more aware that
data analysis is increasingly becoming a vital factor to be compet-
itive, to discover new insight, and to personalize services.
Because of the interesting value that can be extracted from Big
Data, many actors in different countries have launched important
projects. USA was one of the leaders to catch Big Data opportunity.
In March 2012, the Obama Administration launched Big Data
Research and Development Initiative (Weiss and Zgorski, 2012)
with a budget of 200 million. In Japan, Big Data development
became one important axe of the national technological strategy
in July 2012 (Chen et al., 2014b). The United Nations issued a
report entitled Big Data for Development: Opportunities and Chal-
lenges (Letouzé, 2012). It aims to outline the main concerns about
Big Data challenges and to foster the dialogue about how Big Data
can serve the international development.
As a result of the different Big Data projects across the world,
many Big Data models, frameworks and new technologies were cre-
ated to provide more storage capacity, parallel processing and real-
time analysis of different heterogeneous sources. In addition, new
solutions have been developed to ensure data privacy and security.
Compared to traditional technologies, such solutions offer more flex-
ibility, scalability and performance. Furthermore, the cost of most
hardware storage and processing solutions is continuously dropping
due to the sustainable technological advance (Purcell, 2013).
To extract knowledge from Big Data, various models, programs,
softwares, hardwares and technologies have been designed and pro-
posed. They try to ensure more accurate and reliable results for Big
Data applications. However, in such environment, it may be time
consuming and challenging to choose among numerous technolo-
gies. In fact, many parameters should be considered: technological
compatibility, deployment complexity, cost, efficiency, perfor-
mance, reliability, support and security risks. There exist many Big
Data surveys in the literature but most of them tend to focus on algo-
rithms and approaches used to process Big Data rather than tech-
nologies (Ali et al., 2016; Chen and Zhang, 2014; Chen et al., 2014a)
(cf. Section 6). In this paper, we present a survey on recent technolo-
gies developed for Big Data. We categorize and deeply compare them
not only according to their usage, benefits, limits and features, but
also according to their different layers such as Data Storage Layer,
Data Processing Layer, Data Querying Layer, Data Access Layer and
Management Layer. This helps to better understand the connections
among various Big Data technologies and their functioning.
This paper is organized as follows. Section 2 defines Big Data
and presents some of its applications. Section 3 identifies and dis-
cusses some technical challenges faced in dynamic Big Data envi-
ronment. Section 4 presents Hadoop framework and compares
2 A. Oussous et al. / Journal of King Saud University – Computer and Information Sciences xxx (2017) xxx–xxx

some main modules developed on top of it (e.g., data storage, inte-
gration, processing and interactive querying). Section 5 presents
main Hadoop distributions: Cloudera, Hortonworks Data Platform
(HDP), Amazon Elastic MapReduce, MapR, IBM InfoSphere BigIn-
sights, GreenPlums Pivotal HD and Oracle Big Data appliance. Sec-
tion 6 presents related works.
2. Background
2.1. Big Data definition
Unlike traditional data, the term Big Data refers to large grow-
ing data sets that include heterogeneous formats: structured,
unstructured and semi-structured data. Big Data has a complex
nature that require powerful technologies and advanced algo-
rithms. So the traditional static Business Intelligence tools can no
longer be efficient in the case of Big Data applications.
Most data scientists and experts define Big Data by the follow-
ing three main characteristics (called the 3Vs) (Furht and
Villanustre, 2016):
Volume: Large volumes of digital data are generated continu-
ously from millions of devices and applications (ICTs, smart-
phones, products’ codes, social networks, sensors, logs, etc.).
According to McAfee et al. (2012), it is estimated that about
2.5 exabytes were generated each day in 2012. This amount is
doubling every 40 months approximately. In 2013, the total
digital data created, replicated, and consumed was estimated
by the International Data Corporation (a company which pub-
lishes research reports) as 4.4 Zettabytes (ZB). It is doubling
every 2 years. By 2015, digital data grew to 8 ZB (Rajaraman,
2016). According to IDC report, the volume of data will reach
to 40 Zeta bytes by 2020 and increase of 400 times by now
(Kune et al., 2016).
Velocity: Data are generated in a fast way and should be pro-
cessed rapidly to extract useful information and relevant
insights. For instance, Wallmart (an international discount
retail chain) generates more than 2.5 PB of data every hour from
its customers transactions. YouTube is another good example
that illustrates the fast speed of Big Data.
Variety: Big Data are generated from distributed various
sources and in multiple formats (e.g., videos, documents, com-
ments, logs). Large data sets consist of structured and unstruc-
tured data, public or private, local or distant, shared or
confidential, complete or incomplete, etc.
Emani et al. (2015) and Gandomi and Haider (2015) indicate
that more Vs and other characteristics have been added by some
actors to better define Big Data: Vision (a purpose), Verification
(processed data conforme to some specifications), Validation (the
purpose is fulfilled), Value (pertinent information can be extracted
for many sectors), Complexity (it is difficult to organize and analyze
Big data because of evolving data relationships) and Immutability
(collected and stored Big data can be permanent if well managed).
2.2. Big Data applications
Here are some examples of Big Data applications:
Smart Grid case: it is crucial to manage in real time the national
electronic power consumption and to monitor Smart grids opera-
tions. This is achieved through multiple connections among smart
meters, sensors, control centers and other infrastructures. Big Data
analytics helps to identify at-risk transformers and to detect abnor-
mal behaviors of the connected devices. Grid Utilities can thus
choose the best treatment or action. The real-time analysis of the
generated Big Data allow to model incident scenarios. This enables
to establish strategic preventive plans in order to decrease the cor-
rective costs. In addition, Energy-forecasting analytics help to bet-
ter manage power demand load, to plan resources, and hence to
maximize prots (Stimmel, 2014).
E-health: connected health platforms are already used to per-
sonalize health services (e.g., CISCO solution) (Nambiar et al.,
2013). Big Data is generated from different heterogeneous sources
(e.g., laboratory and clinical data, patients symptoms uploaded
from distant sensors, hospitals operations, pharmaceutical data).
The advanced analysis of medical data sets has many beneficial
applications. It enables to personalize health services (e.g., doctors
can monitor online patients symptoms in order to adjust prescrip-
tion); to adapt public health plans according to population symp-
toms, disease evolution and other parameters.It is also useful to
optimize hospital operations and to decrease health cost
expenditure.
Internet of Things (IoT): IoT (Chen et al., 2014b) represents one of
the main markets of big data applications. Because of the high vari-
ety of objects, the applications of IoT are continuously evolving.
Nowadays, there are various Big Data applications supporting for
logistic enterprises. In fact, it is possible to track vehicles positions
with sensors, wireless adapters, and GPS. Thus, such data driven
applications enable companies not only to supervise and manage
employees but also to optimize delivery routes. This is by exploit-
ing and combining various information including past driving
experience.Smart city is also a hot research area based on the
application of IoT data.
Public utilities: Utilities such as water supply organizations are
placing sensors in the pipelines to monitor flow of water in the
complex water supply networks. It is reported in the Press that
Bangalore Water Supply and Sewage Board is implementing a
real-time monitoring system to detect leakages, illegal connections
and remotely control valves to ensure equitable supply of water to
different areas of the city. It helps tp reduce the need for valve
operators and to timely identifying and fixing water pipes that
are leaking.
Transportation and logistics: (Rajaraman, 2016) Many public
road transport companies are using RFID (Radiofrequency Identifi-
cation) and GPS to track busesand explore interesting data to
improve there services. . . For instance, data collected about the
number of passengers using the buses in different routes are used
to optimize bus routes and the frequency of trips. various real-time
system has been implemented not only to provide passengers with
recommendations but also to offer valuable information on when
to expect the next bus which will take him to the desired destina-
tion. Mining Big Data helps also to improve travelling business by
predicting demand about public or private networks. For instance,
in India that has one of the largest railway networks in the world,
the total number of reserved seats issued every day is around
250,000 and reservation can be made 60 days in advance. Making
predictions from such data is a complicated issue because it
depends on several factors such as weekends, festivals, night train,
starting or intermediate station. By using the machine learning
algorithms, it is possible to mine and apply advanced analytics
on past and new big data collection. In fact advanced analytics
can ensure high accuracy of results regarding many issues.
Political services and government monitoring: Many government
such as India and United States are mining data to monitor political
trends and analyze population sentiments. There are many appli-
cations that combine many data sources: social network communi-
cations, personal interviews, and voter compositions. Such systems
enable also to detect local issues in addition to national issues. Fur-
thermore, governments may use Big Data systems to optimize the
use of valuable resources and utilities. For instance, sensors can be
placed in the pipelines of water supply chains to monitor water
A. Oussous et al. / Journal of King Saud University – Computer and Information Sciences xxx (2017) xxx–xxx 3

flow in large networks. So it is possible for many countries to rely
on real-time monitoring system to detect leakages, illegal connec-
tions and remotely control valves to ensure equitable supply of
water to different areas of the city.
3. Big Data challenges
The mining of Big Data offers many attractive opportunities.
However, researchers and professionals are facing several chal-
lenges when exploring Big Data sets and when extracting value
and knowledge from such mines of information. The difficulties
lye at different levels including: data capture, storage, searching,
sharing, analysis, management and visualization. Furthermore,
there are security and privacy issues especially in distributed data
driven applications. Often, the deluge of information and dis-
tributed streams surpass our capability to harness. In fact, while
the size of Big Data keeps increasing exponentially, the current
technological capacity to handle and explore Big Data sets, is only
in the relatively lower levels of petabytes, exabytes and zettabytes
of data. In this section, we discuss in more details some technolog-
ical issues still opened for research.
3.1. Big Data management
Data scientists are facing many challenges when dealing with
Big Data. One challenge is how to collect, integrate and store, with
less hardware and software requirements, tremendous data sets
generated from distributed sources (Chen et al., 2014b;
Najafabadi et al., 2015a). Another challenge is Big Data manage-
ment. It is crucial to efficiently manage Big Data in order to facili-
tate the extraction of reliable insight and to optimize expenses.
Indeed, a good data management is the foundation for Big Data
analytics. Big Data management means to clean data for reliability,
to aggregate data coming from different sources and to encode
data for security and privacy. It means also to ensure efficient Big
Data storage and a role-based access to multiple distributed end-
points. In other words, Big Data management goal is to ensure reli-
able data that is easily accessible, manageable, properly stored and
secured.
3.2. Big Data cleaning
Those five steps (Cleaning, Aggregation, Encoding, Storage and
Access) are not new and are known in the case of traditional data
management. The challenge in Big Data is how to manage the com-
plexity of Big Data nature (velocity, volume and variety) (Khan
et al., 2014) and process it in a distributed environment with a
mix of applications. In fact, for reliable analysis results, it is essen-
tial to verify the reliability of sources and data quality before
engaging resources. However, data sources may contain noises,
errors or incomplete data. The challenge is how to clean such huge
data sets and how to decide about which data is reliable, which
data is useful.
3.3. Big Data aggregation
Another challenge is to synchronize outside data sources and
distributed Big Data plateforms (including applications, reposito-
ries, sensors, networks, etc.) with the internal infrastructures of
an organization. Most of the time, it is not sufficient to analyze
the data generated inside organizations. In order to extract valu-
able insight and knowledge, it is important to go a step further
and to aggregate internal data with external data sources. External
data could include third-party sources, information about market
fluctuation, weather forecasting and traffic conditions, data from
social networks, customers comments and citizen feedbacks. This
can help, for instance, to maximize the strength of predictive mod-
els used for analytics.
3.4. Imbalanced systems capacities
An important issue is related to the computer architecture and
capacity. Indeed, it is known that the CPU performance is doubling
each 18 months following the Moore’s Law, and the performance of
disk drives is also doubling at the same rate. However, the I/O
operations do not follow the same performance pattern. (e.i, ran-
dom I/O speeds have improved moderately while sequential I/O
speeds increase with density slowly) (Chen and Zhang, 2014). Con-
sequently, this imbalanced system capacities may slow accessing
data and affects the performance and the scalability of Big Data
applications. From another angle, we can notice the various devices
capacities over a network (e.i, sensors, disks, memories). This may
slow down system performance.
3.5. Imbalanced Big Data
Another challenge is classifying imbalanced dataset. This issue
has gained lots of attention in the last years. In fact, real-world
applications may produce classes with different distributions.
The first type of class that are under-presented with negligible
number of instances (known as the minority or positive class).
The second class that have an abundant number of instances
(known as the majority or negative class). Identifying the minority
classes is important in various fields such as medical diagnosis
(Nahar et al., 2013), software defects detection (Park et al., 2013),
Finances (Zhou, 2013), drug discovery (Zhou, 2013) or bio-
informatics (Yu et al., 2013).
The classical learning techniques are not adapted to imbalanced
data sets. This is because the model construction is based on global
search measures without considering the number of instances.
Indeed, global rules are usually privileged instead of specific rule
so the minority class are neglected during the model construction.
Thus, Standard learning techniques do not consider the difference
between the number of samples belonging to different classes (del
Río et al., 2014). However, the classes which are under-represented
may constitute important cases to identify.
In practice, many problem domains have more than two classes
with uneven distributions, such as protein fold classification and
weld flaw classification (Wang and Yao, 2012). These multi-class
imbalance problems pose new challenges that are not observed
in two-class problems. In fact, dealing with multi-class tasks with
different misclassification costs of classes is harder than dealing
with two-class ones. To solve this problem, different methods have
been developed and often categorized into two categories. The first
one extends some binary classification techniques to make them
applicable for multi-class classification problems, e.g., discriminant
analysis, decision trees, k-nearest neighbors, Naive Bayes, neural
networks, and support vector machines. The second category is
known as Decomposition and Ensemble Methods (DEM). It consists
of decomposing a multi-class classification problem into a set of
binary classification problems that can be solved by Binary Classi-
fiers (BCs), and then classifying a new observation by applying an
aggregative strategy on the BCs’ predictions (Zhou and Fujita,
2017; Zhou et al., 2017).
3.6. Big Data analytics
Big data brings big opportunities and transformative potential
for various sectors; on the other hand, it also presents unprece-
dented challenges to harnessing such large increasing volumes of
data. Advanced data analysis is required to understand the

relationships among features and explore data. For instance, data
analysis enables an organization to extract valuable insight and
monitor the patterns that may affect positively or negatively the
business. Other data driven applications need also real-time anal-
ysis, like navigation, social networks,Finance, biomedicine, astron-
omy, intelligent transport systems. Thus, advanced algorithms and
efficient methods of data mining are needed to get accurate results,
to monitor the changes in various fields and to predict future
observations. However, big data analysis is still challenging for
many reasons: the complex nature of Big Data including the 5Vs,
the need for scalability and performance to analyze such tremen-
dous heterogenous data sets with real-time responsiveness
(Wang et al., 2016; Tsai, 2016).
Nowadays, there are various analytical techniques including
data mining, visualization, statistical analysis, and machine learn-
ing. Many studies tackle this area by either enhancing the used
techniques, proposing new ones or testing the combination of var-
ious algorithms and technologies. Thus, Big Data pushed the devel-
opment of systems architectures, the hardware as well as
softwares. However, we still need analytical advancement to face
Big Data challenges and stream processing. One of the issues is
how to guarantee the timeliness of response when the volume of
data is very large? In the following sub-sections, we explore exam-
ples of the difficulties encountered when applying current analyt-
ical solutions: Machine learning, deep learning, incremental
approaches, granular computing.
3.7. Big Data machine learning
The objective of machine learning is to discover knowledge and
make intelligent decisions. It is used for many real word applica-
tions such as recommendation engines, recognition systems, infor-
matics and data mining, and autonomous control systems (Bishop,
2006). Generally, the field of Machine Learning (ML) is divided into
three subdomains: supervised learning, unsupervised learning, and
reinforcement learning. Reader may refer to Qiu et al. (2016) for
more details about the ML types.
3.7.1. Data Stream learning
Current real-world applications such as sensors networks,
credit card transactions, stock management, blog posts and net-
work traffic produce tremendous datasets. Data mining methods
are important to discover interesting patterns and to extract value
hidden in such huge datasets and streams.
However, traditional data mining techniques such as associa-
tion mining, clustering and classification lack of efficiency, scalabil-
ity and accuracy when applied to such Big Data sets in a dynamic
environment.
Because of the size, speed and variability of streams, it is not
feasible to store them permanently then to analyze them. Thus
researchers need to find new ways to optimize analytical tech-
niques, to process data instances in very limited amount of time
with limited resources (e.i, memory) and to produce in real-time
accurate results.
Furthermore, variability of streams brings unpredictable
changes (e.i, changing distribution of instances) in incoming data
streams. This concept drift affects the accuracy of classification
model trained from past instances. Therefore, several data mining
methods were adapted to include drift detection techniques and to
cope with changing environment. Classification and clustering are
the most studied ones.
Experiments on data streams demonstrated that changes in
underlying concept affects the performance of classifier model.
Thus, improved analytical methods are needed to detect and adapt
to the concept drifts (Jadhav and Deshpande, 2016).
As an example in the current unstable economic environment,
enterprises need an efficient Financial Distress Predict (FDP) sys-
tem. Such system is crucial to improve risk management and sup-
port banks in credit decisions. DFDP (Dynamic Financial Distress
Prediction) became an important branch of FDP research (Sun
et al., 2017). It improves corporate financial risk management. It
focuses on how to update the FDP model dynamically when the
new sample data batches gradually emerge and FDC (Financial Dis-
tress Concept drift) happens over time.
3.7.2. Deep learning
Nowadays, Deep learning constitutes an extremely active
research field in machine learning and pattern recognition. It plays
an important role in predictive analytics applications such as com-
puter vision, speech recognition and natural language processing
(Chen et al., 2014b).
Traditional machine-learning techniques and feature engineer-
ing algorithms, are limited in their ability to process natural data in
their raw form (Razzak et al., 2017). On the contrary, Deep Learning
is more powerful to resolve data analytical and learning problems
found in huge data sets. In fact, it helps to automatically extracting
complex data representations from large volumes of unsupervised
and uncategorized raw data.
Moreover, because deep learning is based on hierarchical learn-
ing and extraction of different levels of complex data abstractions,
it is suitable to simplify the analysis of large data volumes, seman-
tic indexing, data tagging, information retrieval, and discriminative
tasks such a classification and prediction (e.i, a feature extractor
that transformed the raw data (such as the pixel values of an
image)) into a suitable internal representation or feature vector
from which the learning subsystem, often a classifier, could detect
or classify patterns in the input. However, in spite of those advan-
tages, Big Data still presents significant challenges to deep learning
(Najafabadi et al., 2015b):
Huge volumes of Big Data: the training phase is not an easy task
for Big Data learning in general and Deep learning specifically.
This is because the iterative computations of the learning algo-
rithms are very difficult to be parallelized. Thus, there is still a
need to create efficient and scalable parallel algorithms to
improve training stage for Deep models
Heterogeneity: high volumes of data imposes a great challenge
for deep learning. It means to handle large number of examples
(inputs), large varieties of class types (outputs), and very high
dimensionality (attributes). Thus analytical solutions have to
deal with running-time complexity and model complexity. In
addition to that, such large data volumes make it not feasible
to train a deep learning algorithm with a central processor
and storage.
Noisy labels, and non-stationary distribution: because of the
disparate origins and heterogenous sources of Big Data, analyt-
ical researchers are still facing other challenges such as data
incompleteness, missing labeles and noisy labels.
High velocity: as we know, data are generating at extremely
high speed and should be processed in a real-time. In addition
to the high velocity, data are often non-stationary and presents
a changing distribution over time.
Because of those cited issues, Deep Learning solutions still lack
of maturity and need additional extensive research to optimize the
analytical results. In summary, research future works should con-
sider how to improve Deep Learning algorithms in order to tackle
streaming data analysis, high dimensionality, models scalability.
Studies have also to improve formulation of data abstractions, dis-
tributed computing, semantic indexing, data tagging, information

retrieval, criteria selection for extracting good data representa-
tions, and domain adaptation.
3.7.3. Incremental and ensemble learning
Incremental learning and ensemble learning constitute two
learning dynamic strategies. They are fundamental methods in
learning from big stream data with concept drift (Zang et al., 2014).
Incremental and ensemble learning are frequently applied to
data streams and big data. They tackle various difficulties such as
addressing data availability, limited resources. They are adapted
to many applications such as stock trend prediction and user pro-
filing. Applying incremental learning enable to produce faster clas-
sification or forecasting times while receiving new data.
Many traditional machine learning algorithms inherently sup-
port incremental learning, other algorithms can be adapted to
facilitate this. Examples of incremental algorithms include deci-
sions trees(IDE4, ID5R), decision rules, neuronal networks, neu-
ronal Gaussian RBF networks(Learn++, ARTMAP) or the
incremental SVM.
When comparing those types of algorithms, it is noticed that
incremental algorithms are faster. However, ensemble algorithms
are more flexible and can better adapt to concept drift. Moreover,
not all classification algorithms can be used in incremental learn-
ing, but almost every classification algorithms can be used in
ensemble algorithms (Zang et al., 2014). Thus, it is recommended
to use incremental algorithm in the absence of concept drift or if
the concept drift is smooth. On the contrary, ensemble algorithms
are recommended to ensure accuracy in the case of huge concept
drift or abrupt concept drift. Furthermore, if we have to deal with
relatively simple data-stream or a high level of real-time process-
ing, incremental learning is more suitable. However, ensemble
learning constitute a better choice in case of complicated or
unknown distribution of data streams.
3.7.4. Granular computing
Granular Computing (GrC) (Skowron et al., 2016) is not new, but
it has recently became more popular for its use in various Big Data
domains. It shows many advantages in the case of intelligent data
analysis, pattern recognition, machine learning and uncertain rea-
soning for huge size of data sets. Indeed, GrC play important roles
in the design of decision making models while ensuring acceptable
performance.
Technically, GrC constitute a general computation theory based
on granules such as classes, clusters, subsets, groups and intervals.
Thus, it may be used to build an efficient computational model for
complex Big Data applications such as data mining, document
analysis, financial gaming, organization and retrieval of huge data
bases of multimedia, medical data, remote sensing, biometrics.
Distributed systems require to support different users in under-
standing big data at different granularity levels. There is also a
need to analyze data and present results with different viewpoints.
to fulfil those requirements, GrC provides powerful tools for multi-
ple granularity and multiple viewing of data analysis. This enables
to better understand and analyze the complexity of various big
data sets. Moreover, GrC techniques can serve as effective process-
ing tools for real world intelligent systems and dynamic environ-
ment like FDS (Fuzzy Dynamic Decision Systems).GrC enables to
tackle the complex issue of evolving attributes and objects in
streams over time. Indeed, GrC plays an important role to find sim-
ple approximate solution while ensuring cost effectiveness and
improved description. For instance, the integration of GrC and
computational intelligence has become a hot area of research to
develop efficient decision-making models dedicated to resolve
complex problems of Big Data.
GrC can be implemented via various technologies such as: fuzzy
sets, rough sets, random sets, etc. Fuzzy set techniques provide a
novel way to investigate and represent the relation between a
set and its members. This is by considering the continuum degree
of belonging, namely membership functions (similar to human
recognition). Fuzzy information granulation is about a pool of
fuzzy granules derived by granulating objects, rather than a single
fuzzy granule.
In general, Fuzzy sets have been applied to various areas (Wang
et al., 2017) such as control systems, pattern recognition and
machine learning. Fuzzy sets enable us to represent and process
information at distinct levels of information granularity. More
specifically, Fuzzy set techniques play important role in all phases
of Big Data value chain: first in handling the uncertainties of raw
data, then annotating data and finally preparing specific granular
representation of data for artificial intelligent algorithms.
For instance, Huang et al. (2017a) has proved that their pro-
posed model outperform the static algorithms and related incre-
mental algorithms like Zhang’s method (Zhang et al., 2012).
Indeed, the model ensures a better efficiency and optimizes com-
putation. For that, the both solutions of Huang et al. (2017a) and
Huang et al. (2017b) are based on those fundamentals: first a
matrix method is used to construct and calculate rough approxi-
mations. Second, an incremental method is used to dynamically
update rough set approximation.
We notice from the literature that matrices are more and more
used for rough data analysis and approximations. This is because
matrix structure supports the intuitive description of huge data
sets, maintainability and optimized calculations. In fact, to extract
knowledge from huge evolving coming streams, the model updates
and do computations on just small relation matrices (sub-
matrices) instead of updating the whole relation matrix. Unlike
those approaches based on matrix operation, (Luo et al., 2016a)
model used probabilistic rough set model with the incremental
approach. Their goal is to model imprecise data with tolerance of
decision errors in terms of conditional probability and probabilistic
parameters.
Consequently, GrC techniques can improve the current big data
techniques while tackling big data challenges (e.i, challenges raised
by the 5Vs, by pre-processing data or by reconstructing the
problem at a certain granular level). However, it is worth noticing
that the role of GrC and fuzzy set techniques is to provide a
Table 1
Comparaison between HDFS and Hbase features.
Properties HDFS HBase
System HDFS is a distributed file system appropriate to store large files. HBase is a distributed non relational database built on the top of HDFS.
Query and search
performance
HDFS is not a general purpose file system. It does not provide
fast record lookup in files.
It enables fast record lookups (and updates) for large tables.
Storage HDFS stores large files (gigabytes to terabytes in size) across
Hadoop servers.
HBase internally puts the data in indexed ‘‘StoreFiles” that exist on HDFS for
high-speed lookups.
processing HDFS is suitable for High Latency operations batch processing. HBase is built for Low Latency operations.
Access Data is primarily accessed through MapRe- duce. HBase provides access to single rows from billions of records.
Input-ouput
operations
HDFS is designed for batch processing and hence does not
support random reads/writes operations.
HBase enables reads/writes operations. Data is accessed through shell
commands, client APIs in Java, REST, Avro or Thrift.

methodology for knowledge abstraction (granulation) and knowl-
edge representation. This is different from the role of other tech-
niques used for big data, like deep learning (Wang et al., 2017).
4. Big Data and Hadoop ecosystem
4.1. Hadoop capabilities
Apache Hadoop is a well known Big Data technology that has an
important supporting community. It has been designed to avoid
the low performance and the complexity encountered when pro-
cessing and analyzing Big Data using traditional technologies.
One main advantage of Hadoop is its capacity to rapidly process
large data sets, thanks to its parallel clusters and distributed file
system. In fact, unlike traditional technologies, Hadoop do not copy
in memory the whole distant data to execute computations.
Instead, Hadoop executes tasks where data are stored. Thus,
Hadoop relieves network and servers from a considerable commu-
nication load (Usha and Aps, 2014). For instance, it takes just few
seconds on Hadoop to query terabytes of data instead of 20 min
or more on classic SIEM. Another advantage of Hadoop is its ability
to run programs while ensuring fault-tolerance, usually encoun-
tered in distributed environment. To guarantee that, it prevent
data loss by replicating data on servers.
The power of Hadoop platform is based on two main sub-
components: the Hadoop Distributed File System (HDFS) and the
MapReduce framework (explained in the following sections). In
addition, users can add modules on top of Hadoop as needed
according to their objectives as well as their application require-
ments (e.g., capacity, performances, reliability, scalability, secu-
rity). In fact, Hadoop community has contributed to enrich its
ecosystem with several open source modules. In parallel, IT ven-
ders provide special enterprise hardening features delivered within
Hadoop distributions.
4.2. Data Storage Layer: HDFS and HBase
To store data, Hadoop relies on both its file system HDFS and a
non relational database called Apache HBase.
4.2.1. Hadoop Distributed File System (HDFS)
HDFS (White, 2012) is a data storage system. It supports up to
hundreds of nodes in a cluster and provides a cost-effective and
reliable storage capability. It can handle both structured and
unstructured data and hold huge volumes (i.e., stored files can be
bigger than a terabyte). However, users must be aware that HDFS
do not constitute a general purpose file system. This is because
HDFS was designed for high-latency operations batch processing.
In addition, it does not provide fast record lookup in files. HDFS
main advantage is its portability across heterogeneous hardware
and software platforms. In addition, HDFS helps to reduce network
congestion and increase system performance by moving computa-
tions near to data storage. It ensures also data replication for fault-
tolerance. Those features explain its wide adoption.
HDFS (Mall et al., 2016) is based on master–slave architecture.
It distributes large data across the cluster. In fact, the cluster has
a unique master (NameNode) that manages file system operations
and many slaves (DataNodes) that manage and coordinate data
storage on individual compute nodes. To provide data availability,
Hadoop lies on data replication.
4.2.2. HBase
HBase (Prasad and Agarwal, 2016) is a distributed non rela-
tional database. It is an open source project that is built on top of
HDFS. It is designed for low-latency operations. Hbase is based
on column-oriented key/value data model. It has the potential to
support high table-update rates and to scale out horizontally in
distributed clusters. HBase provides a flexible structured hosting
for very large tables in a BigTable-like format.
Tables store data logically in rows and columns (Coronel and
Morris, 2016). The benefit of such tables is that they can handle bil-
lions of rows and millions of columns. HBase allows many attri-
butes to be grouped into column families so that the elements of
a column family are all stored together. This approach is different
from a row-oriented relational database, where all columns of a
row are stored together. Thus, HBase is more flexible than rela-
tional databases. Instead, HBase has the advantage of allowing
users to introduce updates to better handle changing applications’
requirements. However, HBase has the limitation of not supporting
a structured query language like SQL.
Tables of HBase are called HStore and each Hstore has one or
more Map-Files stored in HDFS. Each table must have a defined
schema with a Primary Key that is used to access the Table. The
row is identified by table name and start key while columns may
have several versions for the same row key.
Hbase provides many features such us real-time queries, natu-
ral language search, consistent access to Big Data sources, linear
and modular scalability, automatic and configurable sharding of
tables (Dimiduk et al., 2013). It is included in many Big Data solu-
tions and data driven websites such as Facebook’s Messaging Plat-
form. HBase includes Zookeeper for coordination services and runs
a Zookeeper instance by default. Similarly to HDFS, HBase
(Maheswari and Sivagami, 2016) has a MasterNode that manages
the cluster and slaves that store parts of the tables and perform
operations on data. Table 1 summarizes the differences between
HDFS and HBase.
4.3. Data Processing Layer
MapReduce and YARN constitute two options to carry out data
processing on Hadoop. They are designed to manage job schedul-
ing, resources and the cluster. It is worth noticing that YARN is
more generic than MapReduce.
4.3.1. MapReduce programming model
MapReduce (Lydia and Swarup, 2015) is a framework composed
of a programming model and its implementation. It is one of the
first essential steps for the new generation of Big Data manage-
ment and analytics tools. MapReduce has an interesting benefit
for Big data applications. In fact, it simplifies the processing of mas-
sive volumes of data through its efficient and cost-effective mech-
anisms. It enables to write programs that can support parallel
processing.
In fact, MapReduce programming model uses two subsequent
functions that handle data computations: the Map function and
the Reduce function.
More precisely, a MapReduce program relies on the following
operations:
1. First, the Map function divides the input data (e.g., long text file)
into independent data partitions that constitute key-value
pairs.
2. Then, the MapReduce framework sent all the key-value pairs
into the Mapper that processes each of them individually,
throughout several parallel map tasks across the cluster. Each
data partition is assigned to a unique compute node. The Map-
per outputs one or more intermediate key-value pairs. At this
stage, the framework is charged to collect all the intermediate
key-value pairs, to sort and group them by key. So the result
is many keys with a list of all the associated values.

3. Next, the Reduce function is used to process the intermediate
output data. For each unique key, the Reduce function aggre-
gates the values associated to the key according to a predefined
program (i.e., filtering, summarizing, sorting, hashing, taking
average or finding the maximum). After that, it produces one
or more output key-value pairs.
4. Finally, the MapReduce framework store all the output Key-
value pairs in an output file.
Within MapReduce paradigm, the NameNode runs a Job-
Tracker instance in order to schedule the different jobs and
distribute tasks over the slave nodes. To insure execution reliabil-
ity, the JobTracker monitors the status of the slave nodes and
re-assigns tasks when they failed. Each of the slave nodes runs
a TaskTracker instance for the assigned tasks. A TaskTracker
instance executes the tasks as specified by the JobTracker and
monitors their execution. Each TaskTracker can use multiple
JVMs (Java Virtual Machines) to execute several maps or reduce
tasks in parallel.
Usually, a Hadoop cluster is composed of a client server,
multiple DataNodes and two types of NameNodes (primary and
secondary). The role of the client server is first to load data and
then to submit MapReduce jobs to the NameNode. The primary
Master NameNode is dedicated to coordinate and to manage
storage and computations. On the other hand, the secondary mas-
ter NameNode handles data replication and availability. A unique
physical server may handle the three roles (client, master and
slaves) in small clusters (less than 40 nodes). However, in medium
and large clusters, each role should be assigned to a single server
machine.
4.3.2. YARN
YARN is more generic than MapReduce. It provides a better scal-
ability, enhanced parallelism and advanced resource management
in comparison to MapReduce. It offers operating system functions
for Big Data analytical applications. Hadoop architecture has been
changed to incorporate YARN Ressource Manager. In general, YARN
works on the top of HDFS. This position enables the parallel execu-
tion of multiple applications. It allows also handling both batch
processing and real-time interactive processing. YARN is compati-
ble with Application Programming Interface (API) of MapReduce. In
fact, users have just to recompile MapReduce jobs in order to run
them on YARN.
Unlike MapReduce, YARN (White, 2012) enhances efficiency by
splitting the two main functionalities of the JobTracker into two
separate daemons: (1) ResourceManager (RM) that allocates and
manages resources across the cluster. (2) Application Master
(AM) framework with a library. It is designed to schedule tasks,
to match them with TaskTrackers and to monitor their progress.
AM negotiates also resources with RM and Node Manager. For
instance, it ensures task bookkeeping, maintains counters, restarts
failed or slow tasks. Thus, Job scheduling entity ensures lifecycle
management of all applications executed in a cluster.
4.3.3. Cascading: a MapReduce framework for complex flows
Cascading framework (Mazumder, 2016) is a rich Java API that
provides many components for fast and cost-effective Big Data
application development, testing and integration. Cascading has
interesting benefits.It allows managing advanced queries and han-
dling complex workflows on Hadoop clusters. It supports scalabil-
ity, portability, integration and test-driven development.
This API adds an abstraction level on the top of Hadoop to sim-
plify complex queries through a cascading concept. In fact, the
loaded data are processed and split by a series of functions to get
multiple streams called flows. Those flows form acyclic-directed
graphs and can be joined together as needed.
The pipe assembly defines the flow to run between the data
sources (Source Taps) and the output data (Sink Taps) that are con-
nected to the pipe. A pipe assembly may contain one or more
Tuples of a given size.
A cascading flow is written in Java and transformed during the
execution into classic MapReduce jobs. Flows are executed on
Hadoop clusters and are based on the following process:
A Flow instance is a workflow that first reads the input data
from one or many Source Taps, and then processes them by execut-
ing a collection of parallel or sequential operations as defined by
the pipe assembly. Then, it writes the output data into one or sev-
eral Sink Taps.
A Tuple represents a set of values (like a database record of SQL
table) that can be indexed with Fields and can be stored directly
into any Hadoop File format as key/value pair. A tuple should have
comparable types in order to facilitate Tuple comparison. Many
extensions were added to the Cascading framework to enhance
its capabilities, including (Nathan, 2013):
Pattern: used to build predictive big data applications. It pro-
vides many machine learning algorithms and enables translat-
ing Predictive Model Markup Language (PMML) documents
into applications on Hadoop.
Scalding: used as a dynamic programming language to solve
functional problems. It is based on Scala language with a simple
syntax. This extension is built and maintained by Twitter.
Cascalog: allows to develop application using Java or Clojure (a
dynamic programming language based on Lisp dialect). It sup-
ports Ad-hoc queries, by running a series of multiple MapRe-
duce jobs to analyze different sources (HDFS, databases and
local data). It provides higher level of abstraction than Hive or
Pig (cf. 4.4).
Lingual: provides an ANSI-SQL interface for Apache Hadoop and
supports a rapid migration of data and workloads to and from
Hadoop. Through Lingual, it is easier to integrate the existing
Business Intelligence tools and other applications.
4.4. Data Querying Layer: Pig, JAQL and Hive
Apache Pig (Mazumder, 2016) is an open source framework that
generates a high level scripting language called Pig Latin. It reduces
MapReduce complexity by supporting parallel execution of
Table 2
Hive, Pig and JAQL features.
Properties Data querying tools
Hive Pig Jaql
Language HiveQL (SQL-like) Pig Latin (script-based language) JAQL
Type of language Declarative (SQL dialect) Data flow Data flow
Data structures Suited for structured data Scalar and complex data types File-based data
Schema It has tables’ metadata stored in the database Schema is optionally defined at runtime Schema is optional
Data Access JDBC, ODBC PigServer Jaql web server
Developer Facebook Yahoo IBM

MapReduce jobs and workflows on Hadoop. Through its interactive
environment, Pig like Hive, simplifies exploring and processing in
parallel massive data sets using HDFS (e.g., complex data flow for
ETL, various data analysis). Pig allows also interaction with exter-
nal programs like shell scripts, binaries, and other programming
languages. Pig has its own data model called Map Data (a map is
a set of key-value pairs). (Krishnan, 2013).
Pig Latin has many advantages.It is based on an intuitive syntax
to support an easy development of MapReduce jobs and workflows
(simple or nested flows). It reduces the development time while
supporting parallelism (Sakr, 2016d). Thus, users can rely on Pig
Latin language and several operators to upload and process data.
Pig Latin is an alternative to Java programming language with
scripts similar to a Directed Acyclic Graph (DAG). In fact, in such
DAC, operators that process data constitute nodes while data flows
are presented by edges (Loganathan et al., 2014). On the contrary
to SQL, Pig does not require a schema and can process semi-
structured and unstructured data. It supports more data formats
than Hive. Pig can run on both the local environment in a single
JVM and the distributed environment on a Hadoop cluster.
JAQL (Beyer et al., 2011) is a declarative language on top of
Hadoop that provides a query language and supports massive data
processing. It converts high level queries into MapReduce jobs. It
was designed to query semi-structured data based on JSONs (Java-
Script Object Notation) format. However, it can be used to query
other data formats as well as many data types (e.g., XML,
comma-separated values (CSV) data, flat files). So, JAQL like Pig
does not require a data schema. JAQL provides several in-built
functions, core operators and I/O adapters. Such features ensure
data processing, storing, translating and data converting into JSON
format.
Apache Hive (Shaw et al., 2016) is a data warehouse system
designed to simplify the use of Apache Hadoop. In contrast to
MapReduce, that manages data within files via HDFS, Hive enables
to represent data in a structured database that is more familiar for
users. In fact, Hive’s data model is mainly based on tables. Such
tables represent HDFS directories and are divided into partitions.
Each partition is then divided into buckets.
Moreover, Hive provides a SQL-like language called HiveQL
(Sakr, 2016a) that enable users to access and manipulate
Hadoop-based data stored in HDFS or HBase. Therefore, Hive is
suitable for many business applications.
Hive (Bansal et al., 2016) is not suitable for real-time transac-
tions. In fact, it is based on a low-latency operations. Like Hadoop,
Hive is designed for large scale processing so even small jobs may
take minutes. Indeed, HiveQL transparently converts queries (e.g.,
ad hoc queries, joins, and summarization) into MapReduce jobs
that are processed as batch tasks.
Hive enables also plugging in traditional mappers and reducers
when it is not feasible or inefficient to express them in HiveQL.
Unlike most SQL having schema-on-write feature, Hive
(Loganathan et al., 2014) has schema-on-read and supports mul-
tiple schemas, which defers the application of a schema until
you try to read the data. Though the benefit here is that it loads
faster, the drawback is that the queries are relatively slower.
Hive lacks full SQL support and does not provide row-level
inserts, updates or delete. This is where HBase worth investing.
Table 2 summarizes and compares some characteristics of Hive,
Pig, and JAQL.
4.5. Data Access Layer
4.5.1. Data Ingestion: Sqoop, Flume and Chukwa
Apach Sqoop (Vohra, 2016) is an open source software.It pro-
vides a command-line interface (CLI) that ensures an efficient
transfer of bulk data between Apache Hadoop and structured data-
stores (such as relational databases, enterprise data warehouses
and NoSQL databases). Sqoop offers many advantages. For
instance,it provides fast performance, fault tolerance and optimal
system utilization to reduce processing loads to external systems.
The transformation of the imported data is done using MapReduce
or any other high-level language like Pig, Hive or JAQL (Jain, 2013).
It allows easy integration with HBase, Hive and Oozie. When Sqoop
imports data from HDFS, the output will be in multiple files. These
files may be delimited text files, binary Avro or SequenceFiles con-
taining serialized data. The process of Sqoop Export will read a set
of delimited text files from HDFS in parallel, parse them into
records, and insert them as new rows in a target database table.
Flume (Hoffman, 2015) is designed to collect, aggregate and
transfer data from external machines to HDFS. It has a simple flex-
ible architecture and handles streaming of data flows. Flume is
based on a simple extensible data model to handle massive dis-
tributed data sources. Flume provides various features including
fault-tolerance, tunable reliability mechanism as well as failure-
recovery service. Though that Flume complements well Hadoop,
it is an independent component that can work on other platforms.
It is known for its capacity to run various processes on a single
machine. By using Flume, users can stream data from various
and high volume sources (like Avro RPC source and syslog) into
sinks (such as HDFS and HBase) for real-time analysis (Hoffman,
2013). In addition, Flume provides a query processing engine that
can transform each new data batch before it is channeled to the
specified sink.
Chukwa (Shireesha and Bhutada, 2016) is a data collection sys-
tem built on the top of Hadoop. Chukwa’s goal is to monitor large
distributed systems. It uses HDFS to collect data from various data
providers, and MapReduce to analyze the collected data. It inherits
Hadoop’s scalability and robustness. It provides an interface to dis-
play, monitor and analyze results
Chukwa offers a flexible and powerful platform for Big Data. It
enables analysts to collect and analyze Big Data sets as well as to
monitor and display results. To ensure flexibility, Chukwa is struc-
tured as a pipeline of collection, processing stages as well as
defined interfaces between stages.
Chukwa is based on four main components: first, it relies on
data agents on each machine to emit data. Next, collectors are
Table 3
A comparaison between Flume and Chukwa.
Properties Projects
Chukwa Flume
Real-time It acquires data for periodic real-time analysis (within minutes) It focuses on continuous real-time analysis (within seconds)
Architecture batch system Continuous stream processing system
Manageability It distributes information about data flows broadly among its services It maintains a central list of ongoing data flows, stored
redundantly using Zookeeper
Reliability The agents on each machine are responsible for deciding what data to send.
Chukwa uses an end-to-end delivery model that can leverage local on-disk log files
for reliability
Robust/Fault tolerant with tunable reliability mechanisms and
failover and recovery mechanisms. Flume adopts a hop-by-hop
model

used to collect data from agents and write it to a stable storage.
MapReduce jobs are used to parse and archive data. Users can
rely on a friendly interface (HICC) to display results and data.
It has a web-portal style (De carvalho et al., 2013; Rabkin and
Katz, 2010).
While Flume and Chukwa share similar goals and features, they
have some differences resumed in Table 3.
4.5.2. Data streaming: storm and spark
Storm (Mazumder, 2016) is an open source distributed system
that has the advantage of handling real time data processing in
contrast with Hadoop, which is designed for batch processing.
In comparison to flume, Storm shows better efficiency in imple-
menting complex processing requirements by relying on the Tri-
dent API.
Storm is based on a topology composed of a complete network
of spouts, bolts, and streams. A spout is a source of streams. A bolt
is used to process input streams in order to produce output
streams. Hence, Storm is suitable to perform transformations on
streams using ‘‘spouts” and ‘‘bolts”.
The ISpout interface of Storm can potentially support any
incoming data. In fact, by using Storm (Mazumder, 2016), users
can ingest data from various real time synchronous and asyn-
chronous systems (e.i, JMS, Kafka, Shell and Twitter). Based on
Bolts,Storm enables to write data to any output system. Storm pro-
vides the IBolt interface that supports any type of output systems
like JDBC (to store data to any relational database), Sequence Files,
Hadoop components like HDFS, Hive, HBase, and other messaging
system.
The storm cluster and Hadoop cluster are apparently similar.
However, in Storm, it is possible to run different topologies for dif-
ferent storm tasks. Instead, in Hadoop platform, the only option
consists in implementing Map Reduce jobs for the corresponding
applications. One main difference between Map Reduce jobs and
topologies is the following. MapReduce job stops but a topology
continues to process messages either all the time or until user ter-
minate (Acharjya and Ahmed, 2016a).
Storm is an easy-to-use, rapid, scalable and fault-tolerant sys-
tem, if one or more processes fails, Storm will automatically restart
it. If the process fails repeatedly, Storm will reroute it to another
machine and restart it there. It can be used for many cases (Lyko
et al., 2016) such us real-time analytics, online machine learning,
continuous computation and distributed RPC.
Storm (Lublinsky et al., 2013) is used to prepare results that can
then be analyzed by other Hadoop tools. It can process million
tuples per second. Like MapReduce, Storm provides a simplified
programming model, which hides the complexity of developing
distributed applications.
Apache Spark (Acharjya and Ahmed, 2016a) is an open source
distributed processing framework that was created at the UC
Berkeley AMPLab. Spark is like Hadoop but it is based on in-
memory system to improve performance. It is a recognized analyt-
ics platform that ensures a fast, easy-to-use and flexible comput-
ing. Spark handles complex analysis on large data sets. Indeed,
Spark runs programs up to 100x faster than Hive and Apache
Hadoop via MapReduce in-memory system. Spark is based on the
Apache Hive codebase. In order to improve system performance,
Spark swap out the physical execution engine of Hive. In addition,
Spark offers APIs to support a fast application development in var-
ious languages including Java, Python and Scala (Karau, 2013).
Spark is able to work with all files storage systems that are sup-
ported by Hadoop.
Spark’s data model (Sakr, 2016b) is based on the Resilient Dis-
tributed Dataset (RDD) abstraction. RDDs constitutes a read-only
collection of objects stored in system memory across multiple
machines. Such objects are available without requiring a disk
access. Furthermore, they can be rebuilt if a partition is lost.
The Spark project consists of multiple components for task
scheduling, memory management, fault recovery, interacting with
storage systems, etc. The components are listed as follows:
Spark SQL (Sakr, 2016b): One important feature of Spark SQL is
that it unifies the two abstractions: relational tables and RDD.
So programmers can easily mix SQL commands to query exter-
nal data sets with complex analytics. Concretely, users can run
queries over both imported data from external sources (like
Parquet files an Hive Tables) and data stored in existing RDDs.
In addition, Spark SQL allows writing RDDs out to Hive tables
or Parquet files. It facilitates fast parallel processing of data
queries over large distributed data sets for this purpose. It uses
a query languages called HiveQL. For a fast application develop-
ment, Spark has developed the Catalyst framework. This one
enable users via Spark SQL to rapidly add new optimizations.
Spark streaming (Azarmi, 2016b): Spark Streaming is another
component that provides automatic parallelization, as well as
scalable and fault-tolerant streaming processing. It enables
users to stream tasks by writing batch like processes in Java
and Scala. It is possible to integrate batch jobs and interactive
queries. It runs each streaming computation as a series of short
batch jobs on in-memory data stored in RDDs.
Table 4
A comparaison between Strom and Spark.
Properties Projects
Sprak Storm
Foundation UC Berkeley BackType, Twitter
Type Open source Open source
Implementation language Scala Coljure
Supported languages Java, Python, R, Scala Any
Execution model Batch, streaming Streaming
Latency Spark has latency of just few seconds (Deponding on batch size) Strom has latecy of sub-seconds
Management style Spark writes data to the storage and requires stateful
computations
Storm rools on it own or uses trident and requires stateless
computation
Fault Tolerance Support only exactly once processing mode Supports ‘exaclty once’, ‘at least once’ and ‘at most once’ processing
mode.
Stream sources HDFS Spout
Stream Computation Windows Operations Bolts
Stream Primitives Dstream Tuple
Provisioning Basic monitoring using ganglia Apache Ambari
Resources Manger
Integration
Messos and Yarn Mesos
Hadoop Distr HDP, CDH, MapR HDP

MLlib (Landset et al., 2015): MLlib is a distributed machine
learning framework built on top of Spark. For performance,
MLlib provides various optimized machine learning algorithms
such us classification, regression, clustering, and collaborative
filtering. Like Mahout, MLlib is useful for machine learning cat-
egories. They offer algorithms for topic modeling and frequent
pattern mining. Mlib supports also regression Models. However,
Mahout does not support such model. MLlib is relatively young
in comparison to Mahout.
GraphX, Wendell2014: GraphX constitutes a library for manipu-
lating graphs and executing graph-parallel computations. Like
Spark Streaming and Spark SQL, GraphX extends the features
of Spark RDD API. Thus users can create a directed graph with
arbitrary properties attached to each vertex and edge. GraphX
offers different operators that support graphs manipulation
(e.g., subgraph and mapVertices). It provides also a library of
graph algorithms (e.g., PageRank and triangle counting).
Table 4 summarizes and compares some characteristics of
Storm and Spark.
4.5.3. Storage Management: HCatalog
Apache HCatalog (Wadkar and Siddalingaiah, 2014b) provides a
table and storage management service for Hadoop users. It enables
interoperability across data processing tools (like Pig, Hive and
MapReduce). This is achieved through a shared schema and data
type mechanisms. It provides an interface to simplify read and
write data operations for any data format (e.g., RCFile, CSV, JSON
and SequenceFiles formats) for which a Hive SerDe (serlializer-
deserializer) can be written. For that, The system administrator
provides the InputFormat, OutputFormat and the SerDe.
The abstracted table of HCatalog provides a relational view of
data in HDFS and allows to view disparate data formats in a tabular
format. So users do not have to know where and how data is
stored. Furthermore, HCatalog supports users with other services.
It notifies data availability and provides a REST interface to permit
access to Hive Data Definition Language(DDL) operations (Wadkar
and Siddalingaiah, 2014b). It also provides a notification service
that notifies workflow tools (like Oozie) when new data becomes
available in the warehouse.
4.6. Data analytics
Apache Mahout (Mazumder, 2016) is an open source machine
learning software library. Mahout can be added on top of Hadoop
to execute algorithms via MapReduce. It is designed to work also
on other platforms.
Mahout (Dinsmore, 2016) is essentially a set of Java libraries.It
has the benefit of ensuring scalable and efficient implementation
of large scale machine learning applications and algorithms over
large data sets. Indeed, Mahout library provides analytical capabil-
ities and multiple optimized algorithms. For instance, it offers
libraries for clustering (like K-means, fuzzy K-means, Mean Shift),
classification, collaborative filtering (for predictions and compar-
isons), frequent pattern mining and text mining (for scanning text
and assigning contextual data). Additional tools include topic mod-
eling, dimensionality reduction, text vectorization, similarity mea-
sures, a math library, and more. The various companies those who
have implemented scalable machine learning algorithms are Goo-
gle, IBM, Amazon, Yahoo, Twitter and Facebook (Acharjya and
Ahmed, 2016a).
As confirmed by Hsieh et al. (2016) and Acharjya and Ahmed
(2016a), by integrating Mahout, users do not have to worry about
algorithms development. Instead, they can concentrate on their
main analytical problem and choose the appropriate analytical
models for their applications.
Like Apache Hive (which provides an SQL-like interface to query
data in Hadoop distributed file system), Mahout translates
machine learning tasks expressed in Java into MapReduce jobs
(Manoochehri, 2013).
R (Team, 2000) is a programming language for statistical com-
puting, machine learning and graphics. R is free, open-source soft-
ware distributed and maintained by the R-project that relies on a
community of users, developers and contributors. R programming
language includes a well-developed, simple and effective function-
alities, including conditionals, loops, user-defined recursive func-
tions and input and output facilities. Many Big Data distributions
(like Cloudera, Hortonworks and Oracle) use R to performe
analytics.
One drawbacks of R is its limited capacity to handle extremely
large datasets because of the one node memory limitations. In fact,
R like other high-level languages leads to memory overload
because it is based on temporary copies instead of referencing
existing objects. Furthermore, R programs are executed in a single
thread, and the data must be stored in RAM. Thus, data structures
should be no larger than 10–20 percent of a computer’s available
RAM.
To support a scalable Big Data analysis and resolve memory
limitations, R has developed several packages such us ff package
(offers file-based access to data sets that are too large to be loaded
into memory, along with a number of higher-level functions), big-
memory Package (memory parallel analysis and data mining of
massive data sets), snow Package (supports simple parallel com-
puting) and Teradata Aster R which runs on the Teradata Aster
Table 5
A Comparison between Mahout and R.
Properties Analytical Tools
Apache Mahout R
Type Open source Open source
Programming
language
Java R language
Arichitecture Mostly MapReduce, porting to spark In-memory system
Supported
Platform
All Hadoop distributions and other platforms Hadoop Cloudera Hortonworks Oracle
Features – Its data model is based on Resilient Distributed Datasets (RDD’s). – APIs
for rapid application development). – Support SQL, HiveQL and Scala
through Spark-SQL. – Efficient query execution by Catalyst framework.
– High level tools to interact with data. – Efficient query execution by
Catalyst framework. – High level tools to interact with data.
– Programming language. – Libraries with optimized algorithm for
machine learning algorithms and graphs.
Key Benefits – New users can get started with common use cases quickly. – It
translates machine learning tasks expressed in Java into MapReduce jobs.
– Limited performance in case of very large data sets (One-node memory).
– Supports statistics and machine learning algorithm. – Flexibility to
develop programs. – Package for more options.

Discovery Platform (Brown, 2014), it aims to facilitate the distribu-
tion of data analysis over a cluster of machines and to overcome
one-node memory limitations in R applications. pdDR (program-
ming with Big Data in R) project (Raim, 2013) is another solution
that aims to solve this problem, it enables high-level distributed
data parallelism in R.
R provides a more complete set of classification models (regard-
ing the types and depth of algorithms) in comparison to Mahout
(Ames et al., 2013). However, R is not a rapid solution when com-
pared to other environment because of its object-oriented pro-
gramming that case memory management problems. Indeed, it
may be more practical to use Mahout, Spark, SAS or other frame-
works to ensure a better performance of extensive computations
(Ames et al., 2013).
Ricardo is part of the eXtreme Analytics Platform (XAP) project
of IBM Almaden Research Center designed to handle deep analytics
problems. It combines the features of Hadoop with those of R as
two integrated partners and components. In fact, Ricardo handles
many types of advanced statistical analysis through R functionali-
ties (like K-means, clustering, time-series, SVM classification). It
leverages also the parallelism of Hadoop DMS.
At the same time, Ricardo provides large-scale data manage-
ment capabilities through Hadoop. It supports also Jaql that is a
high-level query language. With such combination Ricardo enables
R to submit aggregation-processing queries (written in Jaql) to
Hadoop. In fact, Ricardo decomposes data analysis algorithms.
Small data parts are executed by R while large data parts are exe-
cuted by Hadoopp/Jaql DMS. This technique minimizes data trans-
fer across the system and ensures system performance. This is for
simple trading, but to support complex trading algorithms, multi-
ple iterations are performed over data sets with trading between R
and Hadoop at each iteration. Experiments showed that Ricardo
improves R’s performance and facilitates operations such us data
exploration, model building, model evaluation over massive data
sets. Table 5 summarizes and compares some characteristics of
Mahout and R.
4.7. Management Layer
4.7.1. Coordination and Workflow: Zookeeper, Avro and Oozie
Zookeeper (Lublinsky et al., 2013) is an open source service
designed to coordinate applications and clusters in Hadoop envi-
ronment. It provides several benefits. For instance, Zookeeper sup-
ports high performance and data availability. It simplifies also
distributed programming and ensures reliable distributed storage.
It is implemented in Java and provides APIs for Java and C-based
programs. Zookeeper is a distributed application based on a cli-
ent–server architecture. Zookeeper’s server can run across several
clusters. Zookeeper has a file system structure that mirrors classic
file system tree architectures. Through its simple interface, Zoo-
keeper enables also to implement fast, scalable and reliable cluster
coordination services for distributed systems. For instance, it pro-
vides the configuration management service that allows a dis-
tributed setup, the naming service to find machines within large
cluster, the replicated synchronization service to protect data and
nodes from lost, the locking service that enables a serialized access
to a shared resource as well as the automatic system recovery from
failures. ZooKeeper is based on an in-memory data management.
Thus, it ensures distributed coordination at a high speed. Zoo-
keeper is increasingly used within Hadoop to provide high avail-
ability for the ResourceManager. It is used also by HBase to
ensure servers management, bootstrapping, and coordination
(George, 2011).
Unlike other components, Apache ZooKeeper (Junqueira and
Reed, 2013) can be used outside Hadoop platform. ZooKeeper is
used by Twitter, Yahoo and other companies within their dis-
tributed systems for configuration management, sharding, locking
and other purposes. It is used also by In IBM’s Big Insights and
Apache Flume.
Apache Avro is a framework for modeling, serializing and mak-
ing Remote Procedure Calls (RPC) (Shapira et al., 2015). Avro
defines a compact and fast binary data format to support data-
intensive applications, and provides support for this format in a
variety of programming languages such us Java, Scala, C, C++ and
Python (Maeda, 2012). Avro ensures efficient data compression
and storages at various nodes of Apache Hadoop.
Within Hadoop, Avro passes data from one program or language
to another (e.g., from C to Pig). Since data is stored with its schema
(self-describing), Avro is compatible with scripting languages.
There is a data serialization system at the core of Avro. Avro sche-
mas can contain both simple and complex types. Avro uses JSON as
an explicit schema or dynamically generates schemas of the exist-
ing Java objects.
Avro offers similar functionality of systems such as Thrift, Pro-
tocol Buffers, etc. However, Avro differs from those systems by
ensuring: (i) Dynamic typing (data processing without code gener-
ation), (ii) untagged data, and (iii) No manually-assigned field IDs
when a schema changes (Lublinsky et al., 2013).
Apache Oozie (Islam and Srinivasan, 2015) is a workflow sched-
uler system designed to run and manage jobs in Hadoop clusters. It
is a reliable, extensible and scalable management system that can
handle efficient execution of large volume of workflows. The work-
flow jobs take the form of a Directed Acyclical Graphs (DAGs).
Oozie can support various types of Hadoop jobs including MapRe-
duce, Pig, Hive, Sqoop and Distcp jobs (Kamrul Islam and
Srinivasan, 2014). One of the main components of Oozie is the
Oozie server. This server is based on two main components: a
Workflow Engine that stores and runs different types of workflow
jobs, and a Coordinator Engine that runs recurrent workflow jobs
triggered by a predefined schedule (White, 2012). Oozie enables
to track the execution of the workflows. In fact, users can cus-
tomize Oozie in order to notify the client about the workflow
and execution status via Http callbacks (e.g., workflow is complete,
workflow enters or exits an action node). Currently, Oozie supports
Derby by default in addition to other databases such us HSQL,
MySQL, Oracle and PostgreSQL. Oozie provides a collection of APIs
library and a command-line interface (CLI) that is based on a client
component (Lublinsky et al., 2013).
4.7.2. System Deployment: Ambari, Whirr, BigTop and Hue
Apache Ambari (Wadkar and Siddalingaiah, 2014a) is designed
to simplify Hadoop management thanks to an intuitive interface. It
supports provisioning, managing, and monitoring Apache Hadoop
clusters through an easy-to-use management Web User Interface.
The interface is based on RESTful APIs. Ambari supports many
Hadoop components such us: HDFS, MapReduce, Hive, HCatalog,
HBase, ZooKeeper, Oozie, Pig and Sqoop. Moreover, Ambari ensures
security over Hadoop clusters using Kerberos authentication proto-
col. It provides also role-based user authentication, authorization,
and auditing functions to manage integrated LDAP and Active
Directory.
Apache Whirr (Sammer, 2012) simplify the creation and
deployment of clusters in cloud environments such as Amazon’s
AWS. It provides a collection of libraries for running cloud services.
The operator can run Whirr as a command-line tool either locally
or within the cloud. Whirr is used to spin up instances and to
deploy and configure Hadoop. In addition, Apache Whirr supports
provisioning of Hadoop as well as Cassandra, ZooKeeper, HBase,
Valdemort (key-value storage), and Hama clusters on the cloud
environments.
BigTop (Lovalekar, 2014) supports Hadoop ecosystem. It aims to
develop packaging and verify Hadoop-related projects such as

those developed by the Apache community. The goal is to evaluate
and to ensure the integrity and the reliability of the system as a
whole rather than to evaluate each sub-module individually.
Hue (Chullipparambil, 2016) is a web application for interacting
with Hadoop and its ecosystem. Hue aggregates the most common
Hadoop components into a single interface. Its main goal is to
enable programmers to use Hadoop without worrying about the
underlying complexity or using a command line. HUE helps to
browse the system, create and manage user accounts, monitor
cluster health, create MapReduce jobs, and provides a front end
for Hive called Beeswax. Beeswax has Wizards to help create Hive
tables, load data, run and manage Hive queries, and download
results in Excel format. Hue (Landset et al., 2015) is compatible
with any version of Hadoop and is available in all of the major
Hadoop distributions.
4.8. Summary
To sum up, this section discussed the Hadoop ecosystem includ-
ing Hadoop Distributed File System (HDFS), HBaseNoSQL database,
Hive data warehouse solution, and the Pig query language for ad
hoc analytical requirements. The section provides an overview of
other components of Hadoop environment including Hcatalog,
Zookeeper, Avro, Oozie, Flume, Sqoop.
One of the major drawbacks to adopting MapReduce paradigm
is that the development cycle of MapReduce jobs can take long
time and may be complex for some use cases. Thus, advanced
scripting languages (like Pig, JAQL) and query languages have been
developed to simplify exploring and processing massive datasets.
Such improvements reduce the number of code lines and enhance
programmer productivity. Furthermore, they offer an intuitive and
easy-to use interface to handle access management to different
sources of Big Data. In fact, they can handle the storage and pro-
cessing of various data sets that can be local, distributed over large
networks or on the cloud. Another important feature is that they
can channel data between data sources to sinks like Hbase and
HDFS.
Various methods were also developed to simplify Big Data anal-
ysis. For instance R is an easy to use solution that can perform
advanced analysis on large data sets via Hadoop. It provides a more
complete set of classification models (regarding the types and
depth of algorithms) in comparison to Mahout.
However, R clearly has some limitations. In fact, R is not a rapid
solution when compared to other environment. This is due to its
object-oriented programming that cause some memory manage-
ment problems. For that reason, it is recommended to use Mahout,
Spark, SAS or other rapid frameworks when the performance of
extensive computations is a priority.
The current study enable us to recommend to organizations and
users to carefully choose among the various available Big Data ana-
lytical tools (either open source or proprietary ones). The selection
should be based on the importance of each factor: nature of data
sets (e.i, volumes, streams, distribution), complexity of analytical
problems, algorithms and analytical solutions used, systems capa-
bilities, security and privacy issues, the required performance and
scalability in addition to the available budget.
We believe also that it is difficult to afford some commercial
tools at a personal level because of the prices and licensing issues.
As a another recommendation,users must be aware that relying
on open source solutions may lead to outdating and modifications
problem. However, open source systems supports the development
and the innovation at a large scale. Organizations have to pay
attention when choosing very recent technologies still in produc-
tion. They have a limited maturity and may lack the support of
developer communities or academic researchers.
5. Hadoop distributions
Various IT vendors and communities work to improve and
enrich Hadoop infrastructure, tools and services. Sharing Big Data
innovations through open source modules is helpful and promotes
Big Data technologies. However, the downside is that users may
end up with an Hadoop platform composed of various versions of
modules from different sources. Because each Hadoop module
has its own curve of maturity, there is a risk of versions incompat-
ibility inside Hadoop platform. In addition, the integration of vari-
ous technologies on the same platform increases security risks.
Usually each module is tested. However, most of the time the com-
bination of technologies from different sources may bring hidden
risks that are not fully investigated nor tested.
To face those issues, many IT Vendors such us IBM, Cloudera,
MapR and Hortonworks have developed their own modules and
packaged them into distributions. One of the goals is to ensure
compatibility, security and performance of all combined modules.
Most of the available Hadoop distributions have been enriched
gradually. They include various services such us: distributed stor-
age systems, resource management, coordination services, interac-
tive searching tools, advanced intelligence analysis tools, etc.
Furthermore, Hadoop distribution providers offer their own com-
mercial support.
5.1. Cloudera
Cloudera (Azarmi, 2016b) is one of the most used Hadoop dis-
tributions. It enables deploying and managing an Enterprise Data
Hub powered by Hadoop. It provides many benefits such as a cen-
tralized administration tool, a unified batch processing, an interac-
tive SQL, as well as a role-based access control.
In addition to that, Cloudera solutions can be integrated to a
wide range of existing infrastructure and can handle disparate
workloads and data formats in a single system. Cloudera proposes
an easy way for browsing and querying data in Hadoop. In fact, it is
possible to realize a real-time interactive querying and visualize
the result in a convenient way. In addition, several tools are avail-
able to support security and data management.
One of the principle Cloudera modules is Impala (Sakr, 2016c).
It constitutes an interesting query language module that is com-
patible with Hadoop. Impala structures data at rest on a columnar
data format. It allows to handle interactive and real-time analysis
on Big Data. On the contrary to Hive, Impala do not use MapReduce
framework. Instead, it uses its own in-memory processing engine
to ensure fast queries over large data sets. Thus, Impala is faster
than Hive when returning querying results. Indeed, Impala like
AMPLab Shark project (Manoochehri, 2013) can directly use data
from existing HDFS and HBase sources. Thus, it minimizes data
movement and hence reduces the execution time of ‘‘Big queries”.
Impala duplicates storage for fault-tolerance and enables the inte-
gration with major Business Intelligence tools. It includes also the
native Hadoop security that is based on Kerberos for authentication
and Apache Sentry for role-based authorization (Menon, 2014).
Cloudera (Prasad and Agarwal, 2016) provides also a flexible
model that supports structured as well as unstructured data.
Cloudera is faster than Hive. For instance, it executes queries at
least 10 times faster than Hive/MapReduce. It has been confirmed
that in comparison to HiveQL (Hive Query Language), Cloudera
ensures 7 to 45 times performance gain for queries with at least
one join. Even the aggregation queries have been speed-up by
approximately 20–90 times. Cloudera also outperforms HiveQL or
MapReduce in terms of real-time responsiveness: in fact, Cloudera
Enterprise version reduces response time of queries to seconds
instead of minutes using HiveQL or MapReduce.

In spite of all those cited advantages, Cloudera has some disad-
vantages (Prasad and Agarwal, 2016). For instance, it is not suitable
for querying streaming data such as streaming video or continuous
sensor data. In addition to that, all joins operations are performed
in memory that is limited by the smallest memory node present in
the cluster. Furthermore, Cloudera robustness can be affected by
the single point failure during query execution. Indeed, it quits
the entire query if any host that is executing the query fails. Cloud-
era Enterprise RTQ does not support internal indexing for files and
does not allow to delete individual rows.
5.2. Hortonworks Data Platform
The Hortonworks Data Platform (HDP) (Azarmi, 2016a) is built
on Apache Hadoop to handle Big Data storage, querying and pro-
cessing. It has the advantage of being a rapid, cost-effective and
scalable solution. It provides several services for management,
monitoring and data integration.in addition to that, HDP has been
positioned as a key integration platform since it provides open
source management tools and supports connections with some BI
platforms.
HDP ensures a distributed storage through DHFS and the non-
relational database Hbase. It allows a distributed data processing
based on the MapReduce, quering data through Hue and running
scripts using Pig. HDP includes Oozie to manage and schedule
workflows, as well as Hcatalog to handle Metadata services. Many
tools are also available in HDP, including webHDFS, Sqoop, Talend
Open Source, Ambari and Zookeeper.
5.3. Amazon Elastic MapReduce (EMR)
Amazon Elastic MapReduce (Amazon EMR) (Sakr, 2016a) is a
web-based service built on Hadoop framework. It has the benefit
of providing an easy, rapid and effective processing of huge data
sets. Furthermore, it simplifies running Hadoop and related Big
Data applications on AWS. It removes the cost and complexity of
managing Hadoop installation. In addition, it allows resizing on
demand the Amazon clusters by extending or shrinking resources.
Thus, it is possible to easily extract valuable insight from big data
sources without caring about the Hadoop complexity (Aher and
Kulkarni, 2015).
This solution is popular in many industries and supports differ-
ent goals such as log analysis, web indexing, data warehousing,
machine learning, financial analysis, scientific simulation, and
bioinformatics. It can handle many data source and types, includ-
ing clickstream logs, scientific data, etc. Another advantage is that
users can connect EMR to several tools like S3 for HDFS, backup
recovery for HBase, Dynamo support for Hive. It includes many
interesting free components such us Pig and Zookeeper.
5.4. MapR
MapR (Kobielus, 2012) is a commercial distribution for Hadoop
designed for enterprises. It has been enhanced to provide a better
reliability, performance and ease of use of Big Data storage, pro-
cessing and especially analysis with machine learning algorithms.
It provides a set of components and projects that can be integrated
to a wide range of Hadoop ecosystem. MapR does not use HDFS.
Indeed, MapR has developed its owns MapR File Systems (MapR-
FS) in order to increase performance and enable easy backups.
The MapR-FS has the advantage of being compatible with NFS.
Thus, data can be easily transferred between them. MapR is based
on the standard Hadoop programming model.
5.5. IBM InfoSphere BigInsights
IBM InfoSphere BigInsights is designed to simplify the use of
Hadoop in the enterprise environment. It has the required poten-
tial to fulfill enterprise needs in terms of Big Data storage, process-
ing, advanced analysis and visualization (Zikopoulos et al., 2011).
The Basic Edition of IBM InfoSphere BigInsights includes HDFS,
Hbase, MapReduce, Hive, Mahout, Oozie, Pig, ZooKeeper, Hue,
and several other open source tools.
IBM InfoSphere BigInsights Enterprise Edition (Zikopoulos et al.,
2012) provides additional important services: performance capa-
bilities, reliability feature, built-in resiliency, security management
and optimized fault-tolerance. It supports advanced Big Data anal-
ysis through adaptive algorithms (e.g., for text processing). In addi-
tion, IBM provides a data access layer that can be connected to
different data sources (like DB2, Streams, dataStage, JDBC, etc.). It
also leverages IBM Infosphere Streams (2013), another tool belong-
ing to the Infosphere set. This IBM distribution has other advan-
tages: first, the possibility to directly store data streams into
BigInsights clusters. Second, it supports real-time analytics on data
streams. This is achieved through a sink adapter and a source adap-
ter to read data from clusters (Di Martino et al., 2014). IBM facili-
tates also visualization through Dashboards and Big Sheets (a
spreadsheet-like interface for manipulating data in clusters).
5.6. GreenPlum’s Pivotal HD
Pivotal HD (Hurwitz et al., 2013) provides advanced database
services (HAWQ) with several components, including its own par-
allel relational database. The platform combines an SQL query
engine that provides Massively Parallel Processing (MPP)
(Gollapudi, 2013), as well as the power of the Hadoop parallel pro-
cessing framework. Thus, the Pivotal HD solution can process and
analyze disparate large sources with different data formats. The
platform is designed to optimize native querying and to ensure
dynamic pipelining.
In addition, Hadoop Virtualization Extensions (HVE) tool sup-
ports the distribution of the computational work across many vir-
tual servers. Free features are also available for resource and
workflow management through Yarn and Zookeeper. To support
an easy management and administration, the platform provides a
command center to configure, deploy, monitor and manage Big
Data applications. For easier data integration, Pivotal HD proposes
its own DataLoader besides the open source components Sqoop
and Flume.
5.7. Oracle Big Data appliance
Oracle Big Data Appliance (Dijcks, 2012) combines, in one sys-
tem, the power of optimized industry-standards hardware, Oracle
software experience as well as the advantages of Apache Hadoop’s
open source components. Thus, this solution includes the open
source distribution of Cloudera CDH and Cloudera Manager
(Segleau et al., 2013).
Oracle Big Data Appliance is presented as a complete solution
that provides many advantages: scalable storage, distributed
computing, convenient user interface, end-to-end administration,
easy-to-deploy system and other features. It supports also the
management of intensive Big Data projects.
The Oracle appliance (Murthy et al., 2011) lies on the power of
the Oracle Exadata Database Machine as well as the Oracle
Exalytics Business Intelligence Machine. The data is loaded into
the Oracle NoSQL database. It provides Big Data connectors for
high-performance and efficient connectivity. It includes also an
open source oracle distribution of R to support advanced analysis.

The Oracle Big Data Enterprise can be deployed using Oracle Linux
and Oracle Java Hotspot virtual machine Hotspot.
5.8. Windows Azure HDInsight
Windows Azure HDInsight (Sarkar, 2014; Nadipalli, 2015) is a
cloud platform developed by Microsoft and powered by Apache
Hadoop framework. It is designed for Big Data management on
the cloud to store, process and analysis any type of large data
sources. It provides simplicity, convenient management tools,
and open source services for Cloud Big Data projects.Furthermore,
it simplifies the processing and intensive analysis of large data sets
in a convenient way. It integrates several Microsoft tools such as
PowerPivot, PowerView and BI features.
5.9. Choosing a Hadoop distribution
As presented above, several distributions are available to sup-
port the use of Apach Hadoop. They are built on Hadoop platform
with open source or proprietary components. In our view, the
choice of a particular distribution depends on many factors, includ-
ing use cases, business needs, Big Data analysis goals, problems to
solve, as well as the existing infrastructure. Lublinsky et al. (2013)
recommends to consider the following parameters when choosing
among distributions:
The technical characteristics: Hadoop version, available compo-
nents, proposed functionalities and features (scalability, data
availability, parallel processing, performance, connectivity with
the existing applications and so on).
The convenience level: easy tools for installation and configura-
tion, user-friendly management interface, possibility to upgrade
versions and integrate patches.
The maintenance needs: clusters’ management, the multi- cen-
ters support, the disaster-recovery support, and so on.
The available budget as well as the cost of the selected solution.
The cost includes the investments related to the deployment,
maintenance and future upgrades and licenses.
The support proposed to simplify the solution’s integration with
the existing infrastructure.
Table 6 shows the main characteristics of the major Hadoop dis-
tributions: Cloudera, Hortonworks and MAPR.
5.10. Summary
In sum, this section outlines some Big Data distributions. It is
noticed that considerable progress has been made to improve
Hadoop distributions. Thus, various tools and modules are avail-
able to handle Big Data projects in terms of:
Data integration: distributions provides powerful open source
tool to support data uploading and integration into the system.
Some of them are based on their own DataLoader.
Distributed storage: distributions are based either on the stan-
dard HDFS as a distributed le system or on their own file sys-
tem. Distributions that incorporate their own file systems
provides more convenience and easier back-ups. The file system
is paired most of the time to a non-relational database and or
rarely to a parallel relational database.
Centralized management: we notice that most of the available
distributions offer tools for workflow and resource manage-
ment in parallel to services for coordination and system control.
In addition, some of them provide advanced tools to manage the
enterprise data hub. Considerable work has been achieved to
offer reliable monitoring and diagnosis service of the entire
platform through Heatmap, alerts, logs, audits, reports and more.
Rapid and interactive analysis: some distributions offers
machine learning tools, support complex algorithm for scalable
advanced analytics.The usefulness of such distributions lies in
offering most of the time their own version for searching and
querying data in real-time (e.g., using API query, or specific
query language). Some of them support even SQL interface to
facilitate querying data. This is a strong point that helps SQL
familiar users to take advantage from new advances in Big
Data. As performance is one of the key issues in Big Data
environment, distributions rely either on MapReduce frame-
works or a specific in-memory processing engine to achieve a
better performance and guarantee rapid queries.
Security: most Hadoop distributions support native Hadoop
security based on Kerberos for authentification and Apache
Table 6
Cloudera, Hortonworks and MapR features.
Cloudera Hortonworks MAPR
Founded Year Mars 2009 June 2011 2009
License Multiple versions: Open source and Licensed Open source Licensed
GUI Yes Yes Yes
Execution
environement
Local or Cloud Local or Cloud Local or Cloud (Amazon)
Metadata
architecture
Centralized Centralized Distributed
Replication Data Data Data + metadata
Management tools Cloudera Manager Ambari MapR Control System
File System Access HDFS, read-only NFS HDFS, read-only NFS HDFS, read/write NFS (POSIX)
SQL Support Impala Stinger Drill
Security Supports default Kerberos based authentication for
Hadoop services.
Supports default Kerberos based authentication for
Hadoop services
Supports default Kerberos
based authentication for
Hadoop services.
Deployement Deployement with Whirr toolkit. Complex
deployment compared to AWS Hadoop or MapR
Hadoop.
Deployement with Ambari. Simple Deployment. Through AWS Management
Console.
Maintenance The maintenance and upgrade requires efforts. Job
schulding is done through Oozie.
A set of operational capabilities that provide visibility
of the health of the clusters.
Easy to maintain as cluster is
managed through AWS
Management Console and
AWS toolkit.
Cost Cloudera Standard is free. Cloudera entreprise version
is proprietary, needs to be purchased separately. Costs
are applicable based on components and tools
adopted
HDP is the only completely open Hadoop data platform
available. All solutions in HDP are developed as
projects through the Apache Software Foundation.
There are no proprietary extension
Billing is done through AWS
on hourly basis.

Sentry coupled to LDAP for fine grained role based access con-
trol. This constitute one of their strength as security is a major
issue when dealing with distributed Big Data belonging to
various proprietaries. Some distributions offer in parallel addi-
tional auditing tools and special features to ensure security on
the Cloud on the bases of many options such us: VPN IPsec
for encrypted connections, virtual private cloud instances, hard-
ware isolation, list of network access control and security
groups. In spite of those security solutions, users must be aware
that dealing with security and privacy issues at different levels
(i.e., data level, storage level and analysis level) is still
complicated and rises many challenges (i.e., how to find a bal-
ance between security and privacy rules while extracting value
and knowledge from continuous streams).
Visualization: hopefully, current distributions offer various
dynamic visualisation tools (e.g., dashboards, Big Sheets,reports,
graphs). Thus, they facilitate expressive and intuitive visualiza-
tion of data, query results, and they support easy Results mon-
itoring. Some Big Data applications go a step forward by
formulating recommendations to support decision making in
various fields (heath, security, transport and so on).
Cloud computing services: we notice that other researchers and
professionals worked to offer distributions that can be seam-
lessly integrated to cloud platforms. So they offer many cloud
services. This is important for web applications that need
important computing resources. It is also useful to enable com-
panies to externalize applications’ maintenance and other man-
agement overhead to Cloud providers at an affordable cost.
As a summary, the studied distributions and tools enable to
extract valuable insight and knowledge from Big Data in its differ-
ent forms. However, such distributions have clearly some limita-
tions and may differ in their offerings and capacities. We notice
that many researchers are still going forward to optimize all the
faces of the used technologies and approaches in order to face
the increasing multi-streams and Big Data challenges. As a sum-
mary, the studied distributions and tools enable to extract valuable
insight and knowledge from Big Data in its different forms. How-
ever, such distributions have clearly some limitations and may dif-
fer in their offerings and capacities. We notice that many
researchers are still going forward to optimize all the faces of the
used technologies and approaches in order to face the increasing
multi-streams and Big Data challenges.
Most of Big Data surveys give an overview on Big Data applica-
tions, opportunities and challenges. Others discuss also techniques
and methodologies used in Big data environment and how they can
help to improve performance, scalability and results accuracy.
Wang et al. (2016) introduces an overview on Big Data includ-
ing four issues, namely: (i) concepts, characteristics and processing
paradigms of Big Data; (ii) the state-of-the-art techniques for deci-
sion making in Big Data; (iii) decision making applications of Big
Data in social science; and (iv) the current challenges of Big Data
as well as possible future directions. Benjelloun et al. (2015) pre-
sents several Big Data projects, opportunities, examples and mod-
els in many sectors such as healthcare, commerce, tourism and
politics. It gives also examples of technologies and solutions devel-
oped to face Big Data challenge. Chen and Zhang (2014) presents
also a survey about Big Data applications, opportunities, challenges
and used techniques. In addition, it discuss methodologies to han-
dle Big Data challenges (e.g., granular computing, cloud computing,
bio-inspired computing, and quantum computing). Chen et al.
(2014a) introduces the general background of big data and real-
world applications. Unlike the previous papers, it discuss Big Data
technologies by focusing on the four phases of the value chain of
big data, i.e., data generation, data acquisition, data storage, and
data analysis. For each phase, they present the general background,
discuss the technical challenges, and review the latest advances.
But it still does not give reader a sufficient comparison and under-
standing of the current Big Data technologies. Salleh and
Janczewski (2016) provides a literature review on security and pri-
vacy issues of big data, while (Benjelloun and Ait Lahcen, 2015)
presents Big Data security challenges and a state of the art in meth-
ods, mechanisms and solutions used to protect data-intensive
information systems.
Sangeetha and Prakash (2017) reviews in more detail big data
mining algorithms, data slicing techniques and clustering tech-
niques. It also discusses their advantages and drawbacks as well
as their performance and quality measurement. Wu et al. (2014)
summarizes the key challenges for Big Data mining. It discuses
some key research initiatives and the authors’ national research
projects in this field. Qiu et al. (2016) gives a brief review of con-
ventional and advanced machine learning algorithms. It presents
also the challenges of learning with big data and the corresponding
recent solutions. Wang (2016) introduces methods in machine
learning, main and new technologies in Big Data, and some appli-
cations of machine learning in Big Data. Challenges of machine
learning applications in Big Data are also discussed.
Skourletopoulos et al. (2017) presents a review of the current big
data research. Thus, it explores applications, opportunities and
challenges, as well as the state-of-the-art techniques and underly-
ing models that exploit cloud computing technologies, such as the
Big Data-as-a-Service (BDaaS) or Analytics-as-a-Service (AaaS).
It is worth mentioning that most of Big Data surveys do not
focus on technologies and present algorithms and approaches used
to process Big Data. For example, Radha and Rao (2016) presents
the characteristics of Big Data applications and state of-the-art
tools and techniques to handle data-intensive applications. Khan
et al. (2014) Surveys and classifies the various attributes of Big
Data, including its nature, definitions, rapid growth rate, volume,
management, analysis, and security. This study also proposes a
data life cycle that uses the technologies and terminologies of Big
Data.
We notice that other articles focus just on one layer of the Big
Data system (i.e., storage techniques, analytics tools or visualisa-
tion tools), in fact (Acharjya and Ahmed, 2016b) presents the tools
used to analyze Big Data. Tsai (2016) presents and discusses big
data analytics. Some important open issues and further research
directions related to Big Data analytics are also presented.
The survey of Raghav et al. (2016) explains the major prerequi-
sites and challenges that should be addressed by the recent explo-
ration and visualization systems. It also describes about the
different techniques and tools currently used for the visualization
of large sets of data and their capabilities to support massive vol-
ume of data from variety of data sources.
Siddiqa et al. (2016) gives a comprehensive investigation of
state-of-the-art storage technologies available for big data.
Another paper (Oussous et al., in press) establish a precise picture
about NoSQL Databases for Big Data as well as the advantages and
disadvantages of the main NoSQL data models and frameworks.
Lee (2017) illustrates the application of data analytics using
merchant review data. The impacts of big data on key business per-
formances are then evaluated. Six technical and managerial chal-
lenges are discussed.
Rajaraman (2016) explains what is big data, how it is analysed,
and give some case studies illustrating the potentials and pitfalls of
big data analytics. Fang et al. (2015) presents an overview of big
data initiatives, technologies and research in industries and acade-
mia, and discusses challenges and potential solutions.
Ali et al. (2016) Highlights the potential and applications of Big
Data technologies for the development of many fields. It provides a
background on Big Data techniques. It presents also a broad view of
approaches used for human development such as Big Data

analytics. It discusses some technical challenges involved in imple-
menting big data.
Luo et al. (2016b) reviews and discuss big data application in
four major biomedical subdisciplines: (1) bioinformatics, (2) clini-
cal informatics, (3) imaging informatics, and (4) public health
informatics.
Thus, most Big Data surveys have only focused on discussing Big
Data opportunities, applications, challenges and issues. Others pre-
ferred to survey and study the algorithms and techniques used in
such context (i.e., data mining and machine learning). Only few
surveys treat Big Data technologies regarding the aspects and
layers that constitute a real-world Big Data system. In fact, most
of the time, such surveys focus and discusses Big Data technologies
from one angle (i.e., Big Data analytics, Big data mining, Big Data
storage, Big Data processing or Big data visualisation). On the con-
trary, in our paper, we tried to treat the subject from the different
angles. For that, we introduce fundamentals about Big Data, oppor-
tunities and applications. Then we present the main challenges
encountered when facing the complexity of imbalanced data sets
and difficulties arisen by the 5Vs of Big Data. After that, we present
the main solutions. We especially compare the different solutions
and technologies dedicated to Big Data. This is done for each layer
of a general Big Data application (i.e., storage layer, access, query-
ing, analysis, management layer). We paid attention to determine
their advantages, features and facilities as well as their limits.
The comparison and tables use many metrics such as scalability,
results reliability (fault-tolerance), as well as the performance gain.
Thus, the reader can have a comprehensive view about the compo-
nents of any Big Data processing system and a deep understanding
of the technological issues and advantages of the main solutions.
Our aim is to help the reader to select the most suitable solution
for each layer according to its case.
6. Conclusion
Current Big Data platforms are supported by various processing,
analytical tools as well as dynamic visualization. Such platforms
enable to extract knowledge and value from complex dynamic
environment. They also support decision making through recom-
mendations and automatic detection of anomalies, abnormal
behavior or new trends.
In this paper, we have studied Big Data characteristics and dee-
ply discussed the challenges raised by Big Data computing systems.
In addition to that, we have explained the value of Big Data mining
in several domains. Besides, we have focused on the components
and technologies used in each layer of Big Data platforms. Different
technologies and distributions have been also compared in terms
of their capabilities, advantages and limits. We have also catego-
rized Big Data systems based on their features and services pro-
vided to final users. Thus, this paper provides a detailed insight
into the architecture, strategies and practices that are currently fol-
lowed in Big Data computing. In spite of the important develop-
ments in Big Data field, we can notice through our comparison of
various technologies, that many short comings exist. Most of the
time, they are related to adopted architectures and techniques.
Thus, further work needs to be carried out in several areas such
as data organization, domain specific tools and platform tools in
order to create next generation Big Data infrastructures. Hence,
technological issues in many Big Data areas can be further studied
and constitute an important research topic.
References
Acharjya, D., Ahmed, K.P., 2016a. A survey on big data analytics: challenges, open
research issues and tools. Int. J. Adv. Comput. Sci. App. 7, 511–518.
Acharjya, D., Ahmed, K.P., 2016b. A survey on big data analytics: challenges, open
research issues and tools. Int. J. Adv. Comput. Sci. App. 7, 511–518.
Aher, S.B., Kulkarni, A.R., 2015. Hadoop mapreduce: a programming model for large
scale data processing. Am. J. Comput. Sci. Eng. Surv. (AJCSES) 3, 01–10.
Ali, A., Qadir, J., urRasool, R., urRasool, R., Sathiaseelan, A., Zwitter, A., Crowcroft, J.,
2016. Big data for development: applications and techniques. Big Data Anal. 1, 2.
Ames, A., Abbey, R., Thompson, W., 2013. Big Data Analytics Benchmarking SAS, R,
and Mahout. SAS Technical Paper.
Azarmi, B., 2016a. The big (data) problem. In: Scalable Big Data Architecture.
Springer, pp. 1–16.
Azarmi, B., 2016b. Scalable Big Data Architecture. Springer.
Bansal, H., Mehrotra, S., Chauhan, S., 2016. Apache Hive Cookbook. Packt Publ..
Benjelloun, F.-Z., Ait Lahcen, A., 2015. Big data security: challenges,
recommendations and solutions. In: Handbook of Research on Security
Considerations in Cloud Computing. IGI Global, pp. 301–313.
Benjelloun, F.-Z., Ait Lahcen, A., Belfkih, S., 2015. An overview of big data
opportunities, applications and tools. In: Intelligent Systems and Computer
Vision (ISCV), 2015 (pp. 1–6). IEEE.
Beyer, K.S., Ercegovac, V., Gemulla, R., Balmin, A., Eltabakh, M., Kanne, C.-C., Ozcan,
F., Shekita, E.J., 2011. Jaql: a scripting language for large scale semistructured
data analysis. In: Proceedings of VLDB Conference.
Bishop, C.M., 2006. Pattern recognition. Mach. Learn. 128, 1–58.
Botta, A., de Donato, W., Persico, V., Pescapé, A., 2016. Integration of cloud
computing and internet of things: a survey. Future Gener. Comput. Syst. 56,
684–700.
Brown, M.S., 2014. Data Discovery For Dummies, Teradata Special Edition. John
Wiley Sons Inc..
Chen, C.P., Zhang, C.-Y., 2014. Data-intensive applications, challenges, techniques
and technologies: a survey on big data. Inf. Sci. 275, 314–347.
Chen, M., Mao, S., Liu, Y., 2014a. Big data: a survey. Mobile Networks App. 19, 171–
209.
Chen, M., Mao, S., Zhang, Y., Leung, V.C., 2014b. Big Data: Related Technologies,
Challenges and Future Prospects. Springer.
Chullipparambil, C.P., 2016. Big Data Analytics Using Hadoop Tools (Ph.D. thesis).
San Diego State University.
Coronel, C., Morris, S., 2016. Database Systems: Design, Implementation,
Management. Cengage Learning.
De carvalho, O.M., Roloff, E., Navaux, P.O., 2013. A Survey of the state-of-the-art in
event processing. In: 11th Workshop on Parallel and Distributed Processing
(WSPPD).
Di Martino, B., Aversa, R., Cretella, G., Esposito, A., Kołodziej, J., 2014. Big data (lost)
in the cloud. Int. J. Big Data Intell. 1, 3–17.
Dijcks, J.P., 2012. Oracle: Big Data for the Enterprise. Oracle White Paper.
Dimiduk, N., Khurana, A., Ryan, M.H., Stack, M., 2013. HBase in Action. Manning
Shelter Island.
Dinsmore, T.W., 2016. Streaming analytics. In: Disruptive Analytics. Springer, pp.
117–144.
Emani, C.K., Cullot, N., Nicolle, C., 2015. Understandable big data: a survey. Comput.
Sci. Rev. 17, 70–81.
Fang, H., Zhang, Z., Wang, C.J., Daneshmand, M., Wang, C., Wang, H., 2015. A survey
of big data research. IEEE Network 29, 6–9.
Furht, B., Villanustre, F., 2016. Introduction to big data. In: Big Data Technol. App..
Springer International Publishing, Cham, pp. 3–11.
Gandomi, A., Haider, M., 2015. Beyond the hype: big data concepts, methods, and
analytics. Int. J. Inf. Manage. 35, 137–144.
George, L., 2011. HBase: The Definitive Guide. O’Reilly Media Inc..
Gollapudi, S., 2013. Getting Started with Greenplum for Big Data Analytics. Packt
Publishing Ltd..
Hoffman, S., 2013. Apache Flume: Distributed Log Collection for Hadoop. Packt
Publishing Ltd..
Hoffman, S., 2015. Apache Flume: Distributed Log Collection for Hadoop. Packt
Publishing Ltd..
Hsieh, M.-Y., Li, G.-L., Liao, M.-H., Chou, W.-K., Li, K.-C., 2016. Accurate analysis of a
movie recommendation service with linked data on hadoop and mahout. In:
Frontier Computing. Springer, pp. 551–560.
Huang, Y., Li, T., Luo, C., Fujita, H., Horng, S.-J., 2017a. Dynamic variable precision
rough set approach for probabilistic set-valued information systems.
Knowledge-Based Syst. 122, 131–147.
Huang, Y., Li, T., Luo, C., Fujita, H., Horng, S.-J., 2017b. Matrix-based dynamic
updating rough fuzzy approximations for data mining. Knowledge-Based Syst.
119, 273–283.
Hurwitz, J., Nugent, A., Halper, F., Kaufman, M., 2013. Big Data for Dummies. (1st
ed.). For Dummies.
Islam, M.K., Srinivasan, A., 2015. Apache Oozie: The Workflow Scheduler for
Hadoop. O’Reilly Media Inc..
Jadhav, A., Deshpande, L., 2016. A survey on approaches to efficient classification of
data streams using concept drift. Int. J. 4.
Jain, A., 2013. Instant Apache Sqoop. Packt Publishing Ltd..
Junqueira, F., Reed, B., 2013. ZooKeeper: Distributed Process Coordination. O’Reilly
Media Inc..
Kamrul Islam, M., Srinivasan, A., 2014. Apache Oozie The Workflow Scheduler for
Hadoop. O’Reilly Media Inc..
Kankanhalli, A., Hahn, J., Tan, S., Gao, G., 2016. Big data and analytics in healthcare:
introduction to the special section. Inf. Syst. Front. 18, 233–235.
Karau, H., 2013. Fast Data Processing with Spark. Packt Publishing Ltd..
Khan, N., Yaqoob, I., Hashem, I.A.T., Inayat, Z., Mahmoud Ali, W.K., Alam, M., Shiraz,
M., Gani, A., 2014. Big data: survey, technologies, opportunities, and challenges.
Sci. World J.

Kobielus, J.G., 2012. The forrester wave: Enterprise hadoop solutions, q1 2012.
Forrester.
Krishnan, K., 2013. Data Warehousing in the Age of Big Data. Morgan Kaufmann
Publishers Inc., San Francisco, CA, USA.
Kune, R., Konugurthi, P.K., Agarwal, A., Chillarige, R.R., Buyya, R., 2016. The anatomy
of big data computing. Software: Pract. Experience 46, 79–105.
Landset, S., Khoshgoftaar, T.M., Richter, A.N., Hasanin, T., 2015. A survey of open
source tools for machine learning with big data in the hadoop ecosystem. J. Big
Data 2, 1.
Lee, I., 2017. Big data: dimensions, evolution, impacts, and challenges. Bus.
Horizons.
Letouzé, E., 2012. Big Data for Development: Challenges Opportunities. UN Global
Pulse.
Loganathan, A., Sinha, A., Muthuramakrishnan, V., Natarajan, S., 2014. A systematic
approach to Big Data. Int. J. Comput. Sci. Inf. Technol. 4, 869–878.
Lovalekar, S., 2014. Big Data: an emerging trend in future. Int. J. Comput. Sci. Inf.
Technol. 5.
Lublinsky, B., Smith, K.T., Yakubovich, A., 2013. Professional Hadoop Solutions. John
Wiley Sons.
Luo, C., Li, T., Chen, H., Fujita, H., Yi, Z., 2016a. Efficient updating of probabilistic
approximations with incremental objects. Knowledge-Based Syst. 109, 71–83.
Luo, J., Wu, M., Gopukumar, D., Zhao, Y., 2016b. Big data application in biomedical
research and health care: a literature review. Biomed. Inf. Insights 8, 1.
Lydia, E.L., Swarup, M.B., 2015. Big data analysis using hadoop components like
flume, mapreduce, pig and hive. Int. J. Sci. Eng. Comput. Technol. 5, 390.
Lyko, K., Nitzschke, M., Ngomo, A.-C.N., 2016. Big data acquisition. In: New Horizons
for a Data-Driven Economy. Springer, pp. 39–61.
Maeda, K., 2012. Comparative survey of object serialization techniques and the
programming supports. J. Commun. Comput. 9, 920–928.
Maheswari, N., Sivagami, M., 2016. Large-scale data analytics tools: apache hive,
pig, and hbase. In: Data Sci. Big Data Comput.. Springer, pp. 191–220.
Mall, N.N., Rana, S., et al., 2016. Overview of big data and hadoop. Imperial J.
Interdiscip. Res. 2.
Manoochehri, M., 2013. Data Just Right: Introduction to Large-scale Data
Analytics. Addison-Wesley.
Mazumder, S., 2016. Big data tools and platforms. In: Big Data Concepts, Theories,
and Applications. Springer, pp. 29–128.
McAfee, A., Brynjolfsson, E., et al., 2012. Big Data: the management revolution.
Harvard Bus. Rev. 90, 60–68.
Menon, R., 2014. Cloudera Administration Handbook. Packt Publishing Ltd..
Murthy, B., Goel, M., Lee, A., Granholm, D., Cheung, S., 2011. Oracle Exalytics in-
Memory Machine: A brief Introduction.
Nadipalli, R., 2015. HDInsight Essentials. Packt Publishing Ltd..
Nahar, J., Imam, T., Tickle, K.S., Chen, Y.-P.P., 2013. Computational intelligence for
heart disease diagnosis: a medical knowledge driven approach. Expert Syst.
App. 40, 96–104.
Najafabadi, M.M., Villanustre, F., Khoshgoftaar, T.M., Seliya, N., Wald, R.,
Muharemagic, E., 2015a. Deep learning applications and challenges in big
data analytics. J. Big Data 2, 1.
Najafabadi, M.M., Villanustre, F., Khoshgoftaar, T.M., Seliya, N., Wald, R.,
Muharemagic, E., 2015b. Deep learning applications and challenges in big
data analytics. J. Big Data 2, 1.
Nambiar, R., Bhardwaj, R., Sethi, A., Vargheese, R., 2013. A look at challenges and
opportunities of Big Data analytics in healthcare. In: In: 2013 IEEE International
Conference on Big Data. IEEE, pp. 17–22.
Nathan, P., 2013. Enterprise Data Workflows with Cascading. O’Reilly Media Inc..
Oussous, A., Benjelloun, F.-Z., AitLahcen, A., Belfkih, S., Comparison and
classification of nosql databases for big data. Int. J. Big Data Intell. (In press).
Park, B.-J., Oh, S.-K., Pedrycz, W., 2013. The design of polynomial function-based
neural network predictors for detection of software defects. Inf. Sci. 229, 40–57.
Prasad, B.R., Agarwal, S., 2016. Comparative study of big data computing and
storage tools: a review. Int. J. Database Theory App. 9, 45–66.
Purcell, B.M., 2013. Big Data using cloud computing. Holy Family Univ. J. Technol.
Res.
Qiu, J., Wu, Q., Ding, G., Xu, Y., Feng, S., 2016. A survey of machine learning for big
data processing. EURASIP J. Adv. Signal Process. 2016, 1–16.
Rabkin, A., Katz, R.H., 2010. Chukwa: a system for reliable large-scale log collection.
In: LISA. vol. 10. pp. 1–15.
Radha, K., Rao, B.T., 2016. A study on big data techniques and applications. Int. J.
Adv. Appl. Sci. 5, 101–108.
Raghav, R., Pothula, S., Vengattaraman, T., Ponnurangam, D., Raghav, R., Pothula, S.,
Vengattaraman, T., Ponnurangam, D., 2016. A survey of data visualization tools
for analyzing large volume of data in big data platform. In: International
Conference on Communication and Electronics Systems (ICCES). IEEE, pp. 1–6.
Raim, A.M., 2013. Introduction to Distributed Computing with pbdR at the UMBC
High Performance Computing Facility. Technical Report HPCF-2013-2, UMBC
High Performance Computing Facility, University of Maryland, Baltimore County.
Rajaraman, V., 2016. Big data analytics. Resonance 21, 695–716.
Razzak, M.I., Naz, S., Zaib, A., 2017. Deep learning for medical image processing:
Overview, challenges and future. arXiv preprint arXiv:1704.06825.
del Río, S., López, V., Benítez, J.M., Herrera, F., 2014. On the use of mapreduce for
imbalanced big data using random forest. Inf. Sci. 285, 112–137.
Sakr, S., 2016. Big data 2.0 processing systems: a survey. Springer Briefs in
Computer Science.
Sakr, S., 2016b. General-purpose big data processing systems. In: Big Data 2.0
Processing Systems. Springer, pp. 15–39.
Sakr, S., 2016c. Introduction. In: Big Data 2.0 Processing Systems: A Survey. Springer
International Publishing, Cham, pp. 1–13.
Sakr, S., 2016d. Large-scale stream processing systems. In: Big Data 2.0 Processing
Systems. Springer, pp. 75–89.
Salleh, K.A., Janczewski, L., 2016. Technological, organizational and environmental
security and privacy issues of big data: A literature review. Procedia Comput.
Sci. 100, 19–28.
Sammer, E., 2012. Hadoop Operations. O’Reilly Media Inc..
Sangeetha, J., Prakash, V.S.J., 2017. A survey on big data mining techniques. Int. J.
Comput. Sci. Inf. Secur. 15, 482.
Sarkar, D., 2014. Introducing HDInsight. In: Pro Microsoft HDInsight. Springer, pp.
1–12.
Schmarzo, B., 2013. Big Data: Understanding How Data Powers Big Business. John
Wiley Sons.
Segleau, D., Laker, K., Stackowiak, R., Mishra, G., Harding, D., Mohiuddin, K., Sun, H.,
Hornick, M., Nelson, B., Macdonald, B., Plunkett, T., 2013. Oracle Big Data
Handbook. McGraw-Hill, Oracle Press.
Shapira, G., Seidman, J., Malaska, T., Grover, M., 2015. Hadoop Application
Architectures. O’Reilly Media Inc..
Shaw, S., Vermeulen, A.F., Gupta, A., Kjerrumgaard, D., 2016. Hive architecture. In:
Practical Hive. Springer, pp. 37–48.
Shireesha, R., Bhutada, S., 2016. A study of tools, techniques, and trends for big data
analytics. IJACTA 4, 152–158.
Siddiqa, A., Karim, A., Gani, A., 2016. Big data storage technologies: a survey.
Frontiers 1.
Skourletopoulos, G., Mavromoustakis, C.X., Mastorakis, G., Batalla, J.M., Dobre, C.,
Panagiotakis, S., Pallis, E., 2017. Big data and cloud computing: a survey of the
state-of-the-art and research challenges. In: Advances in Mobile Cloud
Computing and Big Data in the 5G Era. Springer, pp. 23–41.
Skowron, A., Jankowski, A., Dutta, S., 2016. Interactive granular computing.
Granular. Computing 1, 95–113.
Stimmel, C.L., 2014. Big Data Analytics Strategies for the Smart Grid. CRC Press.
Stoianov, N., Urueña, M., Niemiec, M., Machnik, P., Maestro, G., 2013. Integrated
security infrastructures for law enforcement agencies. Multimedia Tools App.,
1–16
Sun, J., Fujita, H., Chen, P., Li, H., 2017. Dynamic financial distress prediction with
concept drift based on time weighting combined with adaboost support vector
machine ensemble. Knowledge-Based Syst. 120, 4–14.
Team, R.C., 2000. R Language Definition. R foundation for statistical computing,
Austria.
Tsai, C.-W., Lai, C.-F., Chao, H.-C., Vasilakos, A.V., 2016. Big data analytics. In: Big
Data Technologies and Applications. Springer, pp. 13–52.
Usha, D., Aps, A.J., 2014. A survey of Big Data processing in perspective of hadoop
and mapreduce. International Journal of Current Engineering and Technology.
Vohra, D., 2016. Using apache sqoop. In: Pro Docker. Springer, pp. 151–183.
Wadkar, S., Siddalingaiah, M., 2014a. Apache Ambari. In: Pro Apache Hadoop.
Springer, pp. 399–401.
Wadkar, S., Siddalingaiah, M., 2014b. Hcatalog and hadoop in the enterprise. In: Pro
Apache Hadoop. Springer, pp. 271–282.
Wang, H., Xu, Z., Fujita, H., Liu, S., 2016. Towards felicitous decision making: an
overview on challenges and trends of big data. Inf. Sci. 367, 747–765.
Wang, H., Xu, Z., Pedrycz, W., 2017. An overview on the roles of fuzzy set techniques
in big data processing: trends, challenges and opportunities. Knowledge-Based
Syst. 118, 15–30.
Wang, L., 2016. Machine learning in big data. Int. J. Adv. Appl. Sci. 4, 117–123.
Wang, S., Yao, X., 2012. Multiclass imbalance problems: analysis and potential
solutions. IEEE Trans. Syst. Man Cybern. Part B (Cybern.) 42, 1119–1130.
Weiss, R., Zgorski, L., 2012. Obama Administration Unveils Big Data Initiative:
Announces 200 Million in New RD Investments. Office of Science and
Technology Policy, Washington, DC.
White, T., 2012. Hadoop: The Definitive Guide. O’Reilly Media Inc..
Wu, X., Zhu, X., Wu, G.-Q., Ding, W., 2014. Data mining with big data. IEEE Trans.
Knowl. Data Eng. 26, 97–107.
Yu, H., Ni, J., Zhao, J., 2013. Acosampling: An ant colony optimization-based
undersampling method for classifying imbalanced dna microarray data.
Neurocomputing 101, 309–318.
Zang, W., Zhang, P., Zhou, C., Guo, L., 2014. Comparative study between incremental
and ensemble learning on data streams: case study. J. Big Data 1, 1–16.
Zhang, J., Li, T., Ruan, D., Liu, D., 2012. Rough sets based matrix approaches with
dynamic attribute variation in set-valued information systems. Int. J.
Approximate Reasoning 53, 620–635.
Zhou, L., 2013. Performance of corporate bankruptcy prediction models on
imbalanced dataset: the effect of sampling methods. Knowledge-Based Syst.
41, 16–25.
Zhou, L., Fujita, H., 2017. Posterior probability based ensemble strategy using
optimizing decision directed acyclic graph for multi-class classification. Inf. Sci.
400, 142–156.
Zhou, L., Wang, Q., Fujita, H., 2017. One versus one multi-class classification fusion
using optimizing decision directed acyclic graph for predicting listing status of
companies. Inf. Fusion 36, 80–89.
Zikopoulos, P., Eaton, C., et al., 2011. Understanding Big Data: Analytics for
Enterprise Class Hadoop and Streaming Data. McGraw-Hill Osborne Media.
Zikopoulos, P., Parasuraman, K., Deutsch, T., Giles, J., Corrigan, D., et al., 2012.
Harness the Power of Big Data The IBM Big Data Platform. McGraw Hill
Professional.

Big data technologies : A survey

More Related Content

What's hot (20)

Similar to Big data technologies : A survey (20)

Recently uploaded (20)

Big data technologies : A survey