Performance is one of the motivations to use NoSQL databases instead of traditional SQL databases. With data and queries suitable for the data model, NoSQL databases might offer significant performance benefits. In tResearche present study, we rs compared database systems of the traditional relational model and of the NoSQL graph model. In the graph model, which is one of the four major NoSQL types, the data consist of nodes and edges, and it has its own benefits when handling relationship rich data. While in SQL databases multiple tables may need to be joined for a relational query, in graph databases relational information can be queried by navigating through the graph.

Previous studies [4,5,6,7,8]^{[1][2][3][4][5]} where the performance of graph databases, especially Neo4j, was compared with the traditional SQL databases, indicate that graph databases possess better performance than relational databases. However, those studies mainly focused on quite simple queries. In contrast to the earlier studies, wresearchers investigated the performance of database systems in situations where the query complexity increased. In a complex query, the necessary data must be collected from several tables in an SQL database, or by traversing a path of different types of nodes, potentially using recursion, in a graph database. Using a complex query, an aggregated value (e.g., a count or an average) from a large data set can be calculated.

MariaDB and two versions of MySQL were selected as relational database systems. MariaDB was selected because it is a modern database system and, to the best of ourthe knowledge, it has not been compared to graph database systems before. MySQL 8 was included in the present investigation in order to compare how the query performance of complex queries differs between MySQL and MariaDB. Old MySQL 5.1 was included to observe the efficiency development of relational database systems. MariaDB and MySQL 8 were initially based on MySQL 5.1. Thus, they belong to the same database system family, and their comparison is an indication of how relational database query efficiency has developed during the last decade. In addition to using complex queries, weresearchers also paid attention to factors related to efficiency. Indexing is a traditional method to improve performance and it can be applied with both relational and graph databases. Additionally, for the selected graph database system (Neo4J), a more efficient query execution type (call-function) was developed. Recursive queries can be optimized in modern versions of Neo4J. WeResearchers considered all these optimizations in the efficiency evaluations.

In order to benchmark the database systems using complex queries, wresearchers designed and implemented a new test bench that also supports complex queries. Our The test bench was designed for testing queries using MariaDB, MySQL, and Neo4J. The test database relates to enterprise information systems, but it is worth noting that the query types are general, and processing of the data is similar in many other domains. The test bench is called Invoicing Database Test Bench and its source code is available in GitHub [15]^[6]. The program generated a selected amount of data for ourthe test invoicing database schema and performed various query tests. OurThe dataset is public. The source code for generating the data is available in GitHub, and, thus, it is possible for anyone to repeat these tests by installing the same test bench and generating the same data.

2. Previous pPerformance rRelated sStudy

An older MySQL version was included in the comparison to make ourthe research compatible with earlier studies. From this perspective, MariaDB is a natural choice for a modern database, because MariaDB was initially a descendant of MySQL. Based on the popularity of databases, the DB-Engines site ranks MariaDB as 8th out of 138 of relational databases [16]^[7]. Neo4j ranks 1st out of 32 graph databases on the same sites. DB-Engines ranks the databases according to current popularity. Popularity is measured using six parameters. The first parameter is the number of mentions on the websites Google and Bing. Second is the general interest which is measured by frequency in Google Trends. Third is the frequency of technical questions in Stack Overflow or DBA Stack Exchange. Fourth is the number of job offers in Indeed and Simply Hired. Fifth is the number of profiles in LinkedIn in which the system is mentioned. Sixth is the relevance in social networks which is counted by the number of tweets on Twitter, in which the system is mentioned. As all of the databases wresearchers studied are quite popular and are often candidates for use in enterprises. One of the goals of this study was to identify differences in what use case the databases should be used. SQL databases and Neo4j have been compared in several studies [4,5,6,7,8]^{[1][2][3][4][5]}. Khan et al. compared tuned Oracle 11g and Neo4j 3.03 Community Edition [4]^[1]. They used healthcare data, including data of patients, medication, and medical staff. Performance of the databases was evaluated using ten different count(*) queries. Many of the queries included some table joins. A physical database tuning technique called tablespaces was used for Oracle. The same databases were compared without physical database tuning by Khan et al. [6]^[3]. The physical database tuning technique decreased the overall average query time of Oracle from 4.34 to 2.78 s. However, the overall average query time for Neo4j in query tests was only 0.67 s. Thus, Neo4j outperformed Oracle. Holzschuher et al. tested Neo4j version 1.8 performance with different backend solutions [5]^[2]. Neo4j was benchmarked as embedded with native object access, as a dedicated server through RESTful Web Services, with embedded Cypher queries, with Cypher optimized for remote execution with REST, and with Gremlin queries through REST. MySQL version 5.5.27 was also included with Java Persistence API based backend. Queries were written using Cypher, Gremlin and SQL query languages. The test data consisted of data of persons and their relationships. Tests included such queries as friends of friends. As the size of the database increased, the advantages of Neo4j over MySQL became more evident. Neo4j performance stayed nearly constant when MySQL performance dropped by factors of 5 and 7–9. Queries in Neo4j query languages Gremlin and Cypher executed faster than queries using MySQL with JPA. Vicknair et al. compared MySQL Community Server version 5.1.42 and Neo4j version 1.0-b11 in 2010 [7]^[4]. The graph database was transferred into a relational database as nodes and edges. Three types of structural and three types of data queries were made. The first structural query found all orphan nodes and the two other structural queries traversed the graph at depths of 4 and 128. The data queries were count(*) queries counting nodes with certain payloads. Neo4j performed better in structural queries. However, in data queries, MySQL was more efficient, partly due to the use of Lucene indexing in the tested Neo4j. The data contained integers, and Lucene treated the data as text by default, so conversions were necessary and thus impacted the performance. The work [7]^[4] by Vicknair et al. has been referenced in [4,5,6]^[1][2][3]. Batra et al. compared MySQL version 5.1.41 and Neo4j Community version 1.6 in 2012 [8]^[5]. They used a schema with tables user, friends, fav_movies, and actors for testing, and they tested the databases with three queries: “Find all friends of Esha”, “Find all favourite movies of Esha’s friends” and “Find the lead actors of Esha’s friends’ favourite movies”. Queries were executed on 100 and 500 objects. Neo4j had 2–5 times faster query execution times with a 100-objects data set and 15–30 times faster query execution with a 500-objects data set. The work by Batra et al. [8]^[5] was similar to that of the present study as the data were stored in an SQL database with a relational schema unlike in the work by Vicknair et al. [7]^[4]. The work [8]^[5] by Batra et al. is referenced in [5]^[2]. There also exist previous performance studies where MariaDB is involved. Tongkaw et al. compared the performance of MariaDB 10.0.21 and MySQL 5.6 [17]^[8]. They used the Sysbench and OLTP [14]^[9] software systems with OLTP-Simple and OLTP-Seats workloads. Both databases consumed the same number of resources. However, when increasing the number of threads in OLTP-Simple and the number of workers in OLTP-Seats, MySQL became clearly more efficient and outperformed MariaDB. Shalygina et al. studied the Common Table Expression capabilities of MariaDB by comparing it to Postgres [18]^[10]. The study showed that Postgres had better results, when only a few steps of recursion were needed. However, MariaDB was a better choice for longer-executing recursive queries on huge amounts of data. Stanescu [19]^[11] compared the performance of SQL Server 2009 and Neo4j 4.0. Four datasets were used consisting of 350,000, 700,000, 1,400,000 and 2,100,000 entries. A schema with multiple relations between entities was used. Five different query tests were performed with the different datasets. All the queries addressed relations between entities, so joins or matches between relationships were used. The results show that as query complexity and dataset complexity grew, Neo4j performed faster than the SQL Server. Sholichah et al. [20]^[12] compared MySQL and Neo4j. Four queries were used. Three of the queries were tested with datasets containing 10, 100, 500, 1000 and 10,000 records. The fourth query was used to infer the databases’ ability to handle unstructured data. In these tests, MySQL was in general faster and used less memory as the query complexity and number of records increased. Both databases were able to handle unstructured data. Cheng et al. [21]^[13] compared RocksDB 5.8, Hbase 2.2, Cassandra 3.11, Neo4j 3.4.6 and MySQL 5.7. Four relational datasets from TPC-H benchmark were used as well as four real graph datasets. Different types of query workloads were tested, including atomic relational queries such as projection, aggregation, join and order by, TPC-H workloads, and different graph query algorithms. The conclusion of the tests was that relational databases outperformed graph databases with workloads that mainly consisted of group by, sort, aggregation operations and their combinations. Graph databases outperformed relational databases with workloads that mainly consisted of multiple table joins, pattern matching, path identification and their combinations.

In our graph database schema, entities are represented as nodes, and relationships as directed edges. Two edges are used to represent many-to-many relationships. Customer, invoice, target, work, and worktype entities are represented as nodes. Relationships PAYS from the customer to invoice, and CUSTOMER_TARGET from the customer to the target, and PREVIOUS_INVOICE from an invoice to another are represented by directed edges, the last being a recursive relationship. WORK_TARGET, WORK_INVOICE, WORKHOURS and USED_ITEM are each represented by two edges. Below image represents the database structure in a graph format. The attributes of nodes and edges are not illustrated.

Figure 1. Database structure in relational format.

In the graph database schema, entities are represented as nodes, and relationships as directed edges. Two edges are used to represent many-to-many relationships. Customer, invoice, target, work, and worktype entities are represented as nodes. Relationships PAYS from the customer to invoice, and CUSTOMER_TARGET from the customer to the target, and PREVIOUS_INVOICE from an invoice to another are represented by directed edges, the last being a recursive relationship. WORK_TARGET, WORK_INVOICE, WORKHOURS and USED_ITEM are each represented by two edges. Below image represents the database structure in a graph format. The attributes of nodes and edges are not illustrated.