Modeling and Management of Big Data in Databases

Abstract

The work presented in this paper is motivated by the acknowledgment that a complete and updated systematic literature review (SLR) that consolidates all the research efforts for Big Data modeling and management is missing. This study answers three research questions. The first question is how the number of published papers about Big Data modeling and management has evolved over time. The second question is whether the research is focused on semi-structured and/or unstructured data and what techniques are applied. Finally, the third question determines what trends and gaps exist according to three key concepts: the data source, the modeling and the database. As result, 36 studies, collected from the most important scientific digital libraries and covering the period between 2010 and 2019, were deemed relevant. Moreover, we present a complete bibliometric analysis in order to provide detailed information about the authors and the publication data in a single document. This SLR reveal very interesting facts. For instance, Entity Relationship and document-oriented are the most researched models at the conceptual and logical abstraction level respectively and MongoDB is the most frequent implementation at the physical. Furthermore, 2.78% studies have proposed approaches oriented to hybrid databases with a real case for structured, semi-structured and unstructured data.

Source: Diana Martinez-Mosquera, Rosa Navarrete and Sergio Lujan-Mora, https://www.mdpi.com/2071-1050/12/2/634/htm
This work is licensed under a Creative Commons Attribution 4.0 License.

1. Introduction

The Big Data modeling term became widespread in 2011, as is visible in Figure 1. This figure shows searches in Google Trends related to Big Data modeling, which are intensified from 2011 onwards. Searches before 2004 are not presented, since Google Trends does not store earlier data. In recent years, researchers have consolidated their efforts to study new paradigms to deal with Big Data. Thus, novel Big Data modeling and management in databases approaches have emerged, in line with the new requirements. In consequence, new techniques in the database context have evolved towards Not Only SQL (NoSQL).

The work presented in this paper is motivated by the acknowledgment that a complete systematic literature review (SLR) that consolidates all the research efforts for Big Data modeling and management in databases is missing. An SLR is the best way to collect, summarize and evaluate all scientific evidence about a topic. It allows for the description of research areas shown the greatest and least interest by researchers. Considering the exposed issues, the SLR conducted in this work can contribute to solving this lack by collecting and analyzing details about the research published from 2010 to 2019. As a basis for our SLR, we adhered to the guidelines proposed by Kitchenham. Moreover, this paper presents a complete bibliometric analysis and summarizes existing evidence about research on Big Data modeling and management in databases. With the information attained from the analysis performed, we will identify trends and gaps in the published research to provide a background for new research. As result, 1376 papers were obtained from scientific libraries and 36 studies were selected as relevant. All the research efforts were mapped in order to respond the three research questions defined in this research. Our main goal is to consolidate the main works to provide an awareness of the trends and the gaps related to Big Data modeling.

The remainder of this paper is organized as follows. First is the Introduction section, containing the meaning of the different terms discussed in this study. Second, a Method section presents the process used to perform the planning, conducting and reporting of the SLR. The planning phase describes the identification of the need for a SLR study and the development of a review protocol, objectives and justification, research questions and strategy. The conducting phase presents the inclusion and selection criteria for the final corpus of selected studies.

Third, the Results section answers the research questions in three subsections. The first subsection, the Bibliometric Analysis, comprises the reporting stage, including authors' information, such as affiliation and country and relevant data about their works; for instance, publication information, number of citations, funding source, year, digital library, impact factor, ranking. The second subsection, the Systematic Literature Review, presents a mapping of the selected studies according to three key concepts in a concept matrix regarding to the dataset source, modeling and database. The third subsection, the Discussion, highlights relevant findings in the SLR study in order to identify existing trends and gaps. Finally, conclusions and future works are presented.

Figure 1. Trend of use of term Big Data.

1.1. Big Data Concepts

In this part, we describe the main concepts related to Big Data, in order to provide a general overview for the reader and a background of the terms discussed later.

1.1.1. A Brief History of Big Data

The production and processing of large volumes of data began to be of interest to researchers many years ago. By 1944, estimations for the size of libraries, which increased rapidly every year, were made in American universities. In 1997, at the Institute of Electrical and Electronics Engineers (IEEE) Conference on Visualization, the term "Big Data" was used for the first time during the presentation of a study about large datasets' visualization.

Big Data is the buzzword of recent years, that is, a fashionable expression in information systems. The general population relates the term Big Data to its literal meaning of large volumes of data. However, Big Data is a generic term used to refer to large and complex datasets that arise from the combination of famous Big Data V's that characterize it.

1.1.2. Big Data Characterization

As mentioned before, Big Data does not refer only to high volumes of data to be processed. At the beginning of the Big Data studies, their volume, velocity and variety were considered as fundamental characteristics, which were known as the three Vs of Big Data. After advances in the research, new Vs, such as value and veracity, were established. Currently, there are authors who propose up to 42 characteristics needed to consider data as Big Data, therefore, they define 42 Vs for Big Data. For the purposes of our study, we will mention only ten Vs of Big Data, that are presented in a scientific study. Table 1 summarizes each characteristic, along with a brief description.

Table 1. Ten Vs Big Data.

1.1.3. Volume and Velocity

To deal with the Volume and Velocity characteristics of Big Data, ecosystems and architectural solutions, such as lambda and kappa, have been created. Both architectures propose a structure of layers to process Big Data; the main difference between them is that lambda proposes a layer for batch data processing and another for streaming data, while kappa proposes a single layer for both batch and streaming processing. This SLR focuses on data modeling, a concept related to the Variety characteristic, which is explained next.

1.1.4. Variety

Variety is a characteristic referring to the different types of data and the categories and management of a big data repository. As per this characteristic, Big Data has been classified into structured, semi-structured and unstructured data. The next subsections explain in detail each data type.

Structured Data

In Big Data, structured data are represented in tabular form, in spreadsheets or relational databases. To deal with this type of data, widely developed and known technologies and techniques are used. However, according to the report presented by the CISCO company, this type of data only constituted 10% of all existing data in 2014. Therefore, it is very important to analyze the 90% of the remaining data, corresponding to the semi-structured data and unstructured data that will be described below.

Semi-Structured Data

Semi-structured data are considered to be data that do not obey a formal structure, such as a relational database model. However, they present an internal organization that facilitates its processing; for instance, servers' logs in comma-separated values (csv) format, documents in eXtensible Markup Language (XML) format, JavaScript Object Notation (JSON) and Binary JSON (BSON) and so forth. Some authors may consider XML and JSON as structured.

Unstructured Data

Unstructured data are considered those that have either no predefined schema or no organization in their structure. Within this type of data are text documents, emails, sensor data, audio files, images files, video files, data from websites, chats, electronic health records, social media data and spatio-temporal data, among others. According to CISCO, the volume of unstructured data between 2017 and 2022 is expected to increase up to twelvefold.

To support the Variety, Volume and Velocity of Big Data, non-relational, distributed and open source data storage systems have been created. These systems include horizontal scalability, linearization, high availability and fault tolerance. Usually, these databases are known as NoSQL.

1.2. NoSQL

The existing paradigms for dealing with regular data are neither enough nor suitable to deal with Big Data requirements. For that reason, at the data storage level, the introduction of novel approaches, such as the NoSQL databases, is required. NoSQL refers to Not Only SQL, the term used for all the non-relational databases. NoSQL databases are considered schema-less, as they are designed to work without structure; however, in practice, there is a need for a self-sufficient model to define how data will be organized and retrieved from the database. To solve this requirement, some diverse NoSQL data models are proposed.

Data Models

A data model is a representation of the structure of the data for processing and organization. A data model is considered a primary element for storage, analysis and processing in storage systems.
Currently, storage systems are classified into two large groups, relational and non-relational. Within the relational, the well-known models are the Entity–Relationship (ER), Extended Entity Relationship (EER), Key-Cube and Multidimensional, among others. The objective of this article is not to present a deep study of the models considered as classic: they are well-known and do not need to be explained. We only develop a study of the models that are a novelty for Big Data.
For non-relational systems, there are the NoSQL databases; for them, the data models are classified into four main categories:

1. Column-oriented
2. Document-oriented
3. Graph
4. Key-value

Column-Oriented

In this model, data are represented in tabular form by columns and rows. The columns are identifiable by a partition key that is unique and mandatory and the rows by an optional clustering key. The primary key is the combination of the partition and clustering key. Basically, the schema of the tables consists of a set of columns, a primary key and a data type. For Database Management Systems (DBMS) that use the column-oriented data model, we can mention Accumulo, Amazon SimpleDB, Cassandra, Cloudata, Druid, Elassandra, Flink, HBase, Hortonworks, HPCC, Hypertable, IBM Informix, Kudu, MonetDB, Scylla and Splice Machine, among others.

Document-Oriented

In this model, data are stored in key-value pairs, value documents in XML, JSON or BSON formats. Each of the documents can have nested subdocuments, indexes, fields and attributes. As examples of DBMS that use the document-oriented data model, we can mention ArangoDB, Azure, BagriDB, Cloud Datastore, CouchDB, DocumentDB, Elastic, IBM Cloudant, MongoDB, NosDB, RavenDB, RethinkDB, SequoiaDB, ToroDB and UnQlite, among others.

Graph

This model consists of a graph that contains nodes and edges. A node represents an entity and an edge represents the relationship between entities. There are several graph structures: Undirected/directed, Labeled graphs, Attributed graphs, Bigdata, Multigraphs, Hypergraphs and Nested graphs, among others. Some examples of DBMS that use the graph data model are AllegroGraph, ArangoDB, Infinite Graph, GraphBase, HyperGraphDB, InfoGrid, Meronymy, Neo4j, Onyx Database, Titan, Trinity, Virtuoso OpenLink, Sparksee and WhiteDB.

Key-value

In this model, the data are represented by a key-value tuple. The key represents a unique identifier indexed to a value that represents data of arbitrary type, structure and size. Secondary keys and indexes are not supported. Aerospike, Azure Table Storage, BangDB, Berkeley DB, DynamoDB, GenieDB, KeyDB, Redis, Riak, Scalaris, Voldemort, among others are examples of DBMS that use the key-value data model.
Table 2 summarizes the main characteristics of NoSQL data models, such as its main concept, structure, techniques to create the data model, advantages and disadvantages.

Table 2. NoSQL characteristics.

1.3. Data Abstraction Levels

Generally, in the design of both relational and NoSQL databases, three levels of abstraction are used: conceptual, logical and physical. Data modeling is understood as the technique that records the features of data elements in a map that describes the data used in a process. Data modeling illustrates how the data elements are organized and related. Relational modeling methodologies have well established procedures, as a result of decades of research; however, for NoSQL databases the modeling methodologies, specifically for Big Data, are a novel topic that continues to be studied.

Data modeling at the conceptual level is closely related to the scope of the business process. Therefore, the conceptual model is technologic-agnostic and independent of the database to be used. Thus, already-known models for relational databases can be used in non-relational databases. At the logical level, the modeling is focused on the data model to be used. For NoSQL databases, the modeling is aimed at representing the data structure of the column-oriented, document-oriented, graph or key-value models, as described previously. On the physical level, the modeling will represent the own schema of the selected database; that is, the specific implementation of the NoSQL database, such as Cassandra or MongoDB.

2. Method

This SLR study was undertaken based on the guidelines proposed by Kitchenham, resulting in a three-phase division: (1) planning the SLR study, (2) conducting the SLR study and (3) reporting the SLR study. Figure 2 summarizes the phases in our SLR study.

Figure 2. Systematic Literature Review phases.

The activities included in the first phase were checking the existence of other SLR studies about the topic of interest; defining the review protocol to be applied in order to mitigate possible biases from involved researchers; conceiving objectives, justifications and research questions; and, finally, defining the strategy to be pursued.

In the second phase, the activities were to collect the primary studies from the sources defined by the planning step and subsequently apply the criteria for inclusion to select only those studies related to Big Data modeling.

At last, in the third phase, the corresponding activities were to perform a full bibliometric analysis to provide the information about the authors and publication data in a single document. Furthermore, each selected relevant study was mapped to several key concepts. The analysis results allowed us to answer the research questions.

Each phase is further elaborated in the next subsections.

2.1. Planning the SLR Study

The main goals of this phase are the identification of the need for a SLR study and the development of a review protocol.

2.1.1. Identification of need for a SLR study

Following the suggestions provided by some works, we searched for SLRs and similar publications related to Big Data modeling and management, to verify if there was a gap that could be covered with the SLR proposed in this work. We found four works dating 2015, 2016, 2017 and 2018, which are described below.

Ribeiro, Silva and Rodrigues da Silva completed a survey in 2015 focused on data modeling and data analytics. Although not an SLR study, the work describes some concepts that are relevant to the Big Data models. The authors identified the four main data models for Big Data - key-value, document, wide-column and graph - also described in our work. They also presented a brief summary of the abstraction levels, concepts, languages, modeling tools and database tools support. However, their study is not as detailed as ours, nor do they present a bibliometric analysis. For instance, there is a lack of information about the data models used at the conceptual, logical and physical levels, the techniques used for transforming towards the different abstraction levels, the research trends, which data set sources, types and models for Big Data are the most studied and so forth. Furthermore, our SLR is up to date on August 2019. Nevertheless and similarly to us, the study pays special attention to the fact that Big Data modeling and management in databases must be considered for research, documentation and development, as they demonstrate the data modeling necessity as a means to improve the development process in Big Data. However, they do not cover the criteria that we have mentioned before.

Sousa and Val Cura cover the 2012 to 2016 timeframe. They present an SLR study about logical modeling for NoSQL databases. The authors nominate 12 articles and classify them under conceptual, logical and physical levels. They also identify modeling proposals for NoSQL databases, for NoSQL databases' migration and layers' proposals. We do not consider it as a complete work, since in our research we examined 1376 articles about Big Data modeling and management. Furthermore, they conclude that no research about data model conversion from conceptual to logical level existed at this time, even though our findings revealed the existence of several studies related to it.

Davoudian, Chen and Liu present a thorough study of all the concepts and techniques used in NoSQL databases; the data models used in Big Data are described but in our work we also present a deep study on Big Data modeling methods. This is considered as a relevant work but it does not show a bibliometric analysis of all authors, conferences and journals, among other relevant information to know the trends and gaps in this topic of research. Additionally, our work focuses on examining each of the studies conducted in research to give researchers a guide to current approaches and future directions.

Wu, Sark and Zhu identify some NoSQL databases and focuses to compare them according to their data model and the theorem, which indicates that a distributed system can only guarantee two of the following three properties simultaneously: Consistency, Availability and Partition tolerance (CAP). This work also does not consist of an SLR like the one presented in this work.

Furthermore, other recent surveys related to Big Data have been published; for instance, one describing the state-of-the-art about methodologies developed for multimedia Big Data analytics and the challenges, techniques, applications, strategies and future outlook. Another study presents and analyzes in detail the current stage of Big Data environments and platforms and available garbage collection algorithms. These works neither cover the scope of our research questions for Big Data modeling and management, nor achieve the same level of detail and precision.

In the next subsection, we detail the development of our review protocol, asserting our objectives and justification and the research questions.

2.1.2. Development of a Review Protocol

A review protocol is essential in order to mitigate any possible bias from researchers and it must be defined before conducting the SLR. Thus, during this stage, we formed the applied method. First, we proposed specific development goals and the respective justification for our work. Then, we formulated three research questions with the intent of summarizing the existing evidence about Big Data modeling and management. Finally, we elaborated a strategy to conduct this SLR study effectively.

Objectives and Justification

The first objective is to present information about the most relevant research about Big Data modeling and management in a comprehensive bibliometric analysis. This study contains a number of studies from the digital libraries and details the authors, their institutional affiliations, countries and publication details, such as the publication year and their impact factors in the Journal Citation Reports (JCR) and the Scimago Journal Rank (SJR) for journals and in the CORE Ranking for Conferences.
Based on our findings, the second objective was to detect the different approaches for Big Data modeling used in the different studies in order to determine trends and gaps within the three key concepts, source, modeling and database. The SLR study conducted in this research can focus all the research related to Big Data modeling into a single document, to benefit the industry, the academy and the community.

Research Questions

This stage comprises the most important phase of the protocol development. Hence, we took particular care while following Kitchenham's suggestions. Firstly, we identified three actors within the population: (1) researchers, (2) information analysts and (3) software developers who research, document and implement solutions for Big Data modeling and management in databases. Secondly, we considered collecting all the approaches related to data modeling oriented to Big Data. Thirdly, as outcomes, we intend to summarize the findings and determine the trends and gaps in the studied topic. This study raised the following research questions:

Research Question (RQ1): How has the number of published papers about Big Data modeling and management changed over time?

Rationale: Our interest is to consolidate, through a bibliometric analysis, all the research efforts for the topic, providing researchers with the ability to know all the information about the authors and the publication data in a single document. Thus, the reader can know how the studies, in our topic of interest, have grown over time, who were the authors who provided significant contributions towards the subject, which are the most cited studies and which countries are most interested in this research topic, as well as which journals and conferences are involved in this topic and which scientific libraries have the major share of studies about Big Data modeling and management. In addition, we wanted to know whether these researches were mostly funded by industry or the academy.

Research Question (RQ2): Are there any research studies that focus on approaches for semi-structured and unstructured data and what techniques to apply?

Rationale: Our goal is to find out whether the studies are focused on semi-structured and unstructured data, which, according to the data specified in the Big Data Concepts subsection, comprised most of the available data. In addition, we intend to present what models the researchers propose at each modeling abstraction level and to determine three key concepts: source, modeling and database:

• For source: The dataset sources and data types;
• For modeling: The data abstraction levels, the data model proposed at conceptual, logical and physical levels, the techniques used for transformations between abstraction levels, the applied modeling language, the modeling methodology and the proposed tools for automatic model transformation;
• For database: The database type and the evaluation and performance comparison between models.
  

Research Question (RQ3): What are the trends and gaps in Big Data modeling and management?

Rationale: Based on the data obtained in RQ2, our main interest is to present the solutions proposed by researchers in this topic in a consolidated work. The objective is to allow researchers to focus their efforts on the gaps and solutions that allow for standardization over the currently existent or novel methods.

Strategy

The strategy to conduct an exhaustive compilation of studies on the topic of interest included four actions:
1. Finding primary studies from scientific digital libraries, mainly considering whether: (1) they contain indexed research documents, (2) there is a high frequency of databases update and (3) they publish related research about our topic of interest. The sources listed below comply with the desired requirements, in order to focus our systematic review of relevant research:

• IEEE Xplore
• ScienceDirect
• Scopus
• Web of Science (WoS)
  

Moreover, according to a comprehensive study, which evaluated the quality of 28 scientific search systems, Google Scholar is inappropriate as principal search system, while ScienceDirect, Scopus and WoS are suitable to evidence synthesis in an SLR.

2. Applying the inclusion criteria to the primary studies, in order to select those studies related to Big Data modeling and management, we conducted a search of a specific terms-matching process within the articles' titles, abstracts and keywords. Based on our research questions, two major search terms were derived: big data and model. The terms were selected after combining several options, in order to get a significant number of articles and these terms covered the majority of studies that addressed our research subject.
Due to the fact that the selected digital libraries do not share a common search syntax, we enumerated all the search strings applied in each one. The word "model" has been used because some studies use this term when referring to modeling:

• IEEE Xplore - ((("Document Title":"big data" and "data model") OR "Abstract":"big data" and "data model") OR "Index Terms":"big data" and "data model")
• ScienceDirect - Title, abstract, keywords: "big data" and "data model"
• Scopus - TITLE-ABS-KEY ("big data" AND "data model")
• WoS - TS = ("big data" AND "data model"). TS regards to Topic fields that include titles, abstracts and keywords.
  

For the inclusion criteria, only studies written in English and published in conferences or journals were considered. Although no date-limiting factor was defined in our search criteria, it was observed that no results prior to 2010 were returned by any selected scientific library. These results match with Figure 1, where a report from Google Trends demonstrates that the term Big Data started to become popular in 2011;
3. Reviewing the studies for a second time through a reading of the papers' content allowed us to discard the ones not relevant to the context of our topic of interest;
4. The snowballing technique was applied to locate additional relevant articles according to existing references from within the already-reviewed studies.

2.2. Conducting the SLR Study

The search of the SLR study was conducted in August 2019. Figure 3 provides a representation of the selection process applied to the studies. Duplicated studies were discarded from the potentially relevant studies stage.

Figure 3. Selection Process

2.2.1. Inclusion Stage

In this stage, we selected studies related to Big Data modeling and checked their titles, abstracts and keywords according to the previously planned strategy. We considered only English studies that addressed our research questions and published in conferences or journals. Our main objective is to identify the different approaches to data modeling and management in data stores in a general way for the different types of data at the three abstraction levels. As a result, 1 Chinese article and 27 articles corresponding to books, book chapters, letters, notes or editorials, were discarded. Additionally, we also discarded 1259 articles that, although mentioning data models, referred to specific applications or not related to data persistence but to data ingestion, data lakes or data analytics. From this stage, 117 studies were accepted.

2.2.2. Selection Stage

At this stage, a quick review of the full content of every study allowed us to select only those studies related to Big Data modeling. This resulted in the acceptance of 31 studies, the rejection of 70 papers and the filtering of 16 duplicated works. At this phase, we eliminated the duplicated papers.
After scanning the whole content of these selected 31 studies, we also included five new papers after the snowballing review. Finally, 36 studies made up our final corpus to report the SLR study.

2.3. Reporting the SLR Study

The objective of this step is to answer the research questions raised in the review protocol. For this purpose, this study is divided into three parts. In the first part, we perform a bibliometric analysis to answer RQ1; in the second part, we present the literary review with the most relevant data of the approaches in a concept matrix to answer RQ2; and in the third part, we discuss the trends and gaps to answer RQ3. Section 3 presents in detail the results collected from the activities described in this phase.

3. Results

In this section, we answer the research questions via the below activities:

1. A bibliometric analysis, to gather information about the authors and the publication data, the authors and countries with more contributions in the subject, the impact of the selected studies and how the research has grown throughout time, as well as the journals and conferences proceedings where the studies were published;

2. A literature review to map the studies according to three key concepts - source, modeling and database - in a concept matrix. In the source concept, we analyze the dataset sources and data types. In the modeling concept, we analyze the data abstraction levels, the data models proposed at the conceptual, logical and physical levels, the techniques used to perform transformations between abstraction levels, the applied modeling language, the modeling methodology and the proposed modeling tools. At the database concept, we analyze the type and conduct an evaluation and performance comparison between models;

3. A discussion to identify trends and gaps in Big Data modeling and management.

3.1. Bibliometric Analysis

3.2. Systematic Literature Review

3.3. Discussion

In this section, we answer RQ3. Firstly, we refer to the relevant information attained from the bibliometric analysis. From this analysis, it is possible to verify the growth of this research topic. Since 2015, the number of studies has increased around 100 times and 2018 is the year with the most publications. Scopus is the source holding the highest number of relevant studies. The countries with the greatest contribution in the topic of Big Data modeling are the USA and China and more than 50% of the studies are funded.

Those data suggest existent interest in the topic in the research community. Although many researches refer to the Big Data analytics side, the engineering that takes care of the platforms, extraction techniques and loading and transforming data until reaching the required storage, is of no lesser importance, since these comprise the starting point required by analytics processes in order to produce value from the data.

Next, we will take the information attained from the SLR as a reference to discover the trends and gaps in the topic of interest.

3.3.1. Trends

To summarize the observed trends, Figure 17 helps as a reference to present the three main concepts that this analysis is focused on: source, modeling and database.

Figure 17. Trends for Big Data Modeling and Management.

Source

For the Source aspect, the most researched data sources are unstructured data, specifically website information.

Modeling

For the Modeling aspect, seven main trends were identified within the analyzed corpus:

1. Most studies present their proposals at the conceptual, logical and physical abstraction levels;
2. ER model is the most used in the approaches at the conceptual abstraction level, followed by the multidimensional model and, thirdly, XML;
3. At the logical abstraction level, the most researched model is document-oriented, followed by column-oriented and graph-oriented;
4. At the physical abstraction level, implementations focus more on the MongoDB DBMS, followed by Neo4j and Cassandra. Thus, the following de facto standards have emerged, MongoDB for the document-oriented data model, Cassandra for column-oriented and Neo4j for graph data model. These data are supported by statistical information from DB-Engines Ranking - Trend Popularity of the Solid IT company as of December 2019;
5. The most proposed modeling methodology is data-driven;
6. There is not a clear tendency towards a data modeling approach but there are proposals with UML and XML;
7. No data modeling tool is defined as a trend but some studies used EMF;
8. Regarding the different fields of application, ER is commonly used at the conceptual abstraction level in the different case studies. At the logical level, approaches for the migration from relational to document-oriented, graph and column-oriented models are proposed. Specifically, for spatio-temporal and transmission data, solutions for the graph model are proposed. And, for XML and JSON formats from web servers, solutions for all NoSQL data models are suggested.
   

Database

For the Database aspect, 55.55% of the studies focus on homogeneous database types, so the trend towards a persistent polyglot system is not observed.

3.3.2. Gaps

Figure 18 allowed us to identify also three main concepts for analyzing the gaps: source, modeling and database.

Figure 18. Gaps for Big Data Modeling and Management.

Source

For the Source aspect, two main gaps were identified within the analyzed corpus:

1. One of the main objectives of Big Data is to use the data to generate value; this value will depend on the needs of the business. For this reason, it is considered very important that studies are validated with real use cases, since only with the use of real datasets can whether value is being generated be verified. As result of our SLR, only ten studies, corresponding to 27.78%, use real datasets; 16.67% of these works present their case studies with data from websites, 5.56% use data from sensors and 2.78% involve electronic documents' data and images' metadata;
2. Additionally, other Big Data main features must be guaranteed, such as volume, velocity and veracity. According to the results of Table 6, not all of these features are justified in the approaches;
3. To comply with the variety, studies should consider that data can come in any format: structured, semi-structured or unstructured. Therefore, the proposed approaches should address any of these types. Only 2.78% have proposed a solution for all three types of data. The remaining 47.22% of the studies only present approaches for unstructured data, 30.56% only for structured data, 13.89% for semi-structured data, 2.78% combine structured and semi-structured data and 2.78% do not specify any type.
   

Modeling

For the Modeling aspect, the five main gaps found in our study correspond to standardization and data modeling methodology. Standards are considered important, since they guarantee a universal, uniform interpretation, readability of the data model, portability between database engines, platforms and applications, among other facilities for project managers, analysts, designers, developers and end-users of the databases. In summary, the results are as follows:

1. There is no standardization regarding the definition of mapping rules for the transformation between models at the conceptual, logical and physical data abstraction levels. Thus, the 36 studies propose different approaches for the transformation, making it difficult for users to choose the most appropriate one;
2. No NoSQL system has emerged as a standard or as a de facto standard yet;
3. There is no clearly defined use of some standardized language or method for modeling data at the logical and physical levels. According to the results of our SLR, only at the conceptual level can ER be considered as a trend, maybe because the conceptual level is technologic-agnostic;
4. As mentioned in the Modeling Language subsection, for NoSQL databases the new paradigm for the modeling process is query-driven. However, only five studies are focused on this methodology.
5. For Big Data analysis, the implementation of efficient systems is required. In the data-driven methodology, the models are designed before assessing what queries will be performed. The limitation of this solution is that all data will be stored, when only a limited fraction are needed to answer the required queries. On the contrary, in query-driven methodologies, a set of queries must be expressed, evaluated and integrated before modeling, so that planning can focus on the answers to the necessary queries. For solutions where real-time data are used, data-driven models have the risk of being impractical because of the diverse nature of data streams.
6. A research demonstrates that the treatment of data, in terms of time–cost, with the use of query-driven requires less processing time than data-driven. This, in turn, leads to smaller amounts of consumed energy and, therefore, longer life of equipment.
7. According to another study, the use of a query-driven methodology allowed users to make queries in a natural user language that focuses on relevant regions of interest;
   

Techniques such as reverse engineering are not taken into account; thus, this detail could complement the studies. This method is widely used in relational databases on projects that work with existent databases and when no documented models exist (or there is a risk of these models being outdated).

Database

For the Database aspect, the main gap found is related to the variety characteristics of Big Data. To manage this feature correctly, NoSQL databases must be targeted to PPS. A PPS can be obtained through storage in hybrid databases; that is, it supports different storage technologies. Only eight studies have proposed solutions oriented to hybrid databases and, out of these, only one study demonstrated this with a real case that models and manages both structured, semi-structured and unstructured data and stores them in different databases.

4. Conclusions

As a limitation, we know that we were not able to collect all the existent research about the topic, due to the vast number of synonyms and constraints in the major search terms. However, the high quantity of primary studies - a set of 1376 papers was obtained - provides a trustworthy and complete data set.

The SLR study on Big Data modeling and management conducted in this paper has identified 36 relevant studies. These works have been selected based on the inclusion and selection process of research papers from digital scientific libraries until August 2019; there are no published papers prior to 2010.

We have raised three research questions that were answered through bibliometric analysis, an SLR and a discussion. The results of the bibliometric study provide a lot of relevant information. For instance, from 2015 onwards, the number of studies has increased significantly as this was the year when many countries started to contribute to the topic of interest. The leading countries in this topic are the USA and China. The USA have one of the most impactful studies.

The results from the SLR and from the Discussion also reveal some very interesting facts. For instance, more than 50% of the studies do not verify their proposals with real-world datasets. Big Data's velocity characteristic is not justified by 90% of the studies. ER is the most used model at the conceptual abstraction level, document-oriented is the most researched model at the logical abstraction level and MongoDB DBMS is the most frequent implementation at the physical level. Moreover, as the main gaps, we identified the lack of proposal evaluations and a few studies focused on query-driven methodology and hybrid database solutions.

The contribution of this SLR study, by clarifying the knowledge on the specific topic of Big Data modeling, can support researchers and practitioners in improving their Big Data projects. The relevant works collected can be a useful starting point to new studies into Big Data modeling.

Finally, we know that, due to the need to analyze large volumes of data with a variety of structures, which arrive in high frequency, database research became more focused towards NoSQL. However, NoSQL DBMSs have not been able to acquire some strengths that are already present in relational databases, such as the ability to support Consistency and Availability at the same time; for this reason, NewSQL has emerged and may pose a solution to the problems faced by both DBMSs. As a future work, we also expect to study this new database system.

We hope to keep the SLR up to date and present results for other concepts. Moreover, based on the gaps, we will drive our research in these aspects.