Ontology-based GraphQL server generation for data access and data integration

2.1. Ontologies and description logics

The term ontology originates in philosophy, in which it is the science of what is, of the kinds and structures of objects, properties, and relationships in every area of reality [57,64]. Ontologies can be viewed, intuitively, as defining the terms, relations, and rules that combine these terms and relations in a domain of interest [59]. Through ontologies, people and organizations are able to communicate by establishing a common terminology. They provide the basis for interoperability between systems and are applicable as an index to a repository of information as well as a query model and a navigation model for data sources. Moreover, they are often used as a foundation for integrating data sources, thereby alleviating the heterogeneity issue. The benefits of using ontologies are their improved reusability, share-ability and portability across platforms, as well as their increased maintainability and reliability. On the whole, ontologies allow a field to be better understood and allow information in that field to be handled much more effectively and efficiently (e.g., knowledge representation for bioinformatics discussed in [58]).

OPEN IN VIEWER

Fig. 1.

Example of an ontology of the university domain including 4 concepts and relationships among them, as well as relationships between concepts and datatypes.

From a knowledge representation point of view, ontologies usually contain four components: (i) concepts that represent sets or classes of entities in a domain, (ii) instances that represent the actual entities, (iii) relations, and (iv) axioms that represent facts that are always true in the topic area of the ontology. Relations represent relationships among concepts. Axioms represent domain restrictions, cardinality restrictions, or disjointness restrictions. Depending on the components and information related to the components they contain, ontologies can be classified. Figure 1 represents an example ontology for the university domain. The open-headed arrows represent axioms that represent is-a relationships that is, if A is a B, then all entities belonging to concept A also belong to concept B. We say that A is a sub-concept of B. In this example Professor is a sub-concept of Author. Therefore, all Professor entities are Author entities. The closed-headed arrows represent general relations among concepts other than is-a relations. For instance, the Professor concept has a connection to the University concept represented by the doctoralDegreeFrom relation. Additionally, a relation can exist between a concept and a data type reference. For instance, University has a connection to the data type reference xsd:string represented by the UniversityID relation. This means that each entity of the University concept can be associated with a string type value by having a UniversityID connection.

To formally define the above concepts and relationships, we need representation languages. Description logics (DL) are a family of knowledge representation languages. There are three basic building blocks of such a language, namely: (i) atomic concepts (unary predicates) such as University and Department, (ii) atomic roles (binary predicates) such as departments, and (iii) individuals (constants) [7]. Complex concepts and roles can be built by using atomic concepts and logical constructors (e.g., conjunction (⊓), disjunction (⊔), universal restriction (∀) and existential restriction (∃)). Axioms can be defined using general concept inclusions (⊑). For instance, the general concept inclusion (GCI) University ⊑ ∀ departments . Department represents the fact that University is a sub-concept of the concept ∀ departments . Department that represents all entities that may have departments relations to entities and if so, these latter entities must belong to the Department concept; an assertion University(Linköping University) means that the instance Linköping University belongs to the University concept. A DL TBox is a finite set of GCIs and a DL ABox is a finite set of assertions.

Data integration deals with combining data that resides at multiple different sources [14,26,43]. Ideally, a data integration system should enable unified access to a number of data sources [26,43]. Formally, according to [43], a data integration system can be formalized as a triple ⟨ G , S , M ⟩ , where,

Ontology-based data integration (OBDI) is a form of data integration in which an ontology plays the role of a global schema that captures domain knowledge [15]. Usually, in an information system with only one single data source, the formal treatment of OBDI is identical to that of ontology-based data access (OBDA) [15,67]. In this article, we generally refer to both OBDI and OBDA as OBDA. OBDA, as a semantic technology, aims to facilitate access to different underlying data sources [66]. Traditionally, these underlying data sources are considered to be relational databases. Ontologies play the role of global views over multiple data sources. There are different ways to implement an OBDA system. Generally, these systems can be categorized into two types, namely, data warehouse-based approaches and virtual approaches. These two categories of methods both make use of semantic mappings in order to overcome the differences between ontologies and local schemas, but in different ways [21,65]. In a data warehouse-based approach, data from multiple sources are usually loaded or stored in a centralized storage, which is the warehouse [26,63], based on semantic mappings. We refer to the data in such warehouses as materialized data. Depending on the aims or functionalities of a system, the materialized data could be stored in local databases or transformed into RDF graphs. Therefore, queries are evaluated against the materialized data. In a virtual approach, data is retained at the original sources and mediators are used to translate queries defined in terms of a global or mediated schema into queries defined in terms of each data source’s local schema, based on semantic mappings. Therefore, queries are evaluated and executed against each data source. SPARQL queries are widely supported by data integration systems that use ontologies as global schemas.

A number of semantic mapping definition languages have been proposed over the years. One such language is R2RML (RDB to RDF Mapping Language),5

GraphQL schemas and GraphQL resolver functions are basic building blocks in the implementations of GraphQL servers. The former describe how users can retrieve data using GraphQL APIs. The latter contain program code including how to access data sources and structure the obtained data according to the schema. We introduce GraphQL schemas and GraphQL resolver functions in Section 2.3.1 and Section 2.3.2, respectively.

2.3.1. GraphQL schemas

In a GraphQL API, the GraphQL schema defines types, their fields, and the value types of the fields. Such a schema represents a form of vocabulary supported by a GraphQL API rather than specifying what the data instances of an underlying data source may look like and what constraints have to be guaranteed [33]. There are six different (named) type definitions in GraphQL, which are scalar type, object type, interface type, union type, enum type and input object type. Listing 1 depicts a GraphQL schema example.

OPEN IN VIEWER

Listing 1.

Example of a GraphQL schema of the university domain

An object type represents a list of fields and each field has a value of a specific type such as object type or scalar type. A scalar is used to represent a value such as a string. In Listing 1, there are three basic object type definitions, which are University (line 1 to line 4), Department (line 5 to line 8), and Professor (line 12 to line 15). They all have field definitions which represent the relationships to scalar types or to other object types. For instance, the University type has a field definition UniversityID of which the value type is String (line 2), and a field definition departments of which the value type is a list of Departments (line 3). GraphQL allows defining abstract types by supporting the interface type and the union type. An interface type defines a list of fields and allows object types to implement. An object type can then implement an interface type with the requirement that the object type includes all fields defined by the interface type. The schema in Listing 1 contains an interface type, Author with an AuthorID field of which the value type is String (line 9 to line 11). The object type Professor implements Author and must have the same definition for AuthorID field (line 13) as that in Author. A union type defines a list of possible types. An enum type describes the set of possible values that are in scalars. For more details of union and enum types, we refer the reader to the latest GraphQL specification in [29].

GraphQL allows fields to accept arguments to configure their behavior [29]. These arguments can be defined by input object types. An input object type defines an input object with a set of input fields; the input fields are either scalars, enums, or other input objects. This allows arguments to accept arbitrarily complex structs, which can capture notions of filtering conditions. For instance, according to the definitions of UniversityFilter (line 16 to line 22) and StringFilter (line 30 to line 36), we can define an input argument as UniversityID:{_eq:”u1”} to capture the meaning of “UniversityID is equal to ‘u1’”, where _eq represents the equal to operator. In our approach, _and, _or and _not are used to represent boolean expressions. For instance, _or:[{UniversityID:{_eq:”u1”}},{UniversityID:{_eq:”u2”}}] represents the expression “UniversityID is equal to ‘u1’ or ‘u2’”. In the example schema, we use the term filter to represent the name of an input argument (e.g., line 38). Such input arguments defined as input objects are not built-in constructs of GraphQL. Therefore, their meanings are essentially defined by the program code of the GraphQL server implementation, i.e., the resolver functions which manage requests to underlying data sources and structure the returned data according to the GraphQL schema. Our approach presented in Section 4 and Section 5 uses input arguments named as filter to represent filter conditions.

Additionally, a GraphQL schema supports defining types that represent operations such as query and mutation. The schema presumes the Query type as the query root operation type. As Listing 1 shows, in the Query type definition (line 37 to line 42), there are four field definitions, which are UniversityList, DepartmentList, AuthorList, and ProfessorList. For instance, the returned type of UniversityList is [University], a list of Universities. The UniversityList takes an argument defined as UniversityFilter as an input for capturing the notion of a filtering condition.

2.3.2. GraphQL resolver functions

In a GraphQL API, apart from the GraphQL schema defining types, their fields, and the value types of the fields, resolver functions are responsible for populating the data for fields of types in the GraphQL schema. For instance, for the schema example shown in Listing 1, there are four fields defined in the Query type. Therefore, in the GraphQL server implementation, we are supposed to define resolver functions to populate data for these fields, UniversityList, DepartmentList, AuthorList, and ProfessorList. In our approach, we assume that the GraphQL schema supports a query that retrieves all the instances for each interface type or object type. Therefore, we use the name of each interface type or object type concatenated with ‘List’ as the name of a field in the Query type, where the returned type is a list of the interface or object type. This is a way to state the behavior of a field in the Query type. We emphasize that what a GraphQL query can retrieve over the underlying data sources relies on how the resolver function is implemented. For instance, if the underlying data source is a relational database, the resolver function should contain code specifying the SQL query to be evaluated. Listing 2 illustrates an example resolver function (written in JavaScript syntax) for the UniversityList field. We assume that the underlying data source is a relational database that contains a table named university with a column named id. In line 4, given an input argument (university_id) representing the id of a university, a query is evaluated against the relational database. In line 7, the data is structured according to the University object defined in the JavaScript code which corresponds to the University type definition in the schema shown in Listing 1.

OPEN IN VIEWER

Listing 2.

Example of a resolver function for the UniversityList field

3.1. Overview of the framework

Figure 2 illustrates the framework for data access and integration based on GraphQL in which an ontology drives the generation of GraphQL server that provides integrated access to data from heterogeneous data sources. These data sources may be based on different schemas and formats and may be accessed in different ways (e.g., as tabular data accessed via SQL queries or as JSON-formatted data accessed via API requests). To address the heterogeneity, the framework relies on an ontology that provides an integrated view of the data from the different sources, and corresponding semantic mappings that define how the data from the underlying data sources is interpreted or annotated by the ontology (arrows  (a) and  (b) ). Furthermore, two processes are defined. The first process generates the GraphQL server. The second process deals with answering queries and is performed after the GraphQL server is set up. In accordance with these two processes, there are two types of intended users or developers in the framework. One type is users or developers of the GraphQL server generator, who should have prior knowledge of the ontology, semantic mappings and the domain. The other type is end users using a GraphQL server for data access and integration, who may or may not be familiar with the Semantic Web or ontologies. For the purpose of writing GraphQL queries, they need basic prior knowledge of GraphQL, which can be learned from the self-documenting API of the generated GraphQL server showing the schema. We introduce more details about these two processes in Section 3.2 and Section 3.3, respectively.

OPEN IN VIEWER

Fig. 2.

GraphQL-based framework for data access and integration.

This process includes generating both a GraphQL schema for the API provided by the server (arrow  (i) ) and a generic resolver function (arrow  (ii) ). Given an ontology as an integrated view of data from multiple data sources, we propose a method for generating a GraphQL schema based on this ontology, with the result that the schema becomes a view of the data to be integrated. Additionally, we propose a generic implementation of resolver functions that takes semantic mappings as inputs, so that the server is able to get data from underlying data sources. In Sections 4 and 5, we elaborate on the implementation of our approaches for generating a GraphQL schema and the generic resolver function, respectively. This GraphQL server generation process does not need to be repeated unless the ontology or the semantic mappings change. After this generation process, the GraphQL server can be set up.

This process requires users or developers who are familiar with the query mechanisms of underlying data sources, domain ontologies that can be used for data access or integration. Consequently, they can define the scope of the ontology that will be used for generating the GraphQL schema for the server, as well as the semantic mappings that will be used for generating the generic resolver function. This type of automatic generation of GraphQL servers based on ontologies and semantic mappings can also benefit general GraphQL application developers, since it can eliminate the need to build GraphQL servers from scratch.

3.3. GraphQL query answering process

During this process the query is validated against the GraphQL schema (arrow  (1) ); the underlying data sources are accessed via resolver functions, the retrieved data is combined, the data is structured according to the schema (arrows  (2) and  (3) ); and finally the query result is returned (arrow  (4) ).

OPEN IN VIEWER

Listing 3.

Example of a GraphQL query

OPEN IN VIEWER

Listing 4.

Example of a GraphQL query response

A GraphQL query example and corresponding query result are shown in Listing 3 and Listing 4, respectively. The example query is: “Get the university including the head of each department where the UniversityID is ‘u1’”. The query takes as an input an argument defined as filter:{UniversityID:{_eq:”u1”}}, which follows the syntax of the input object type UniversityFilter. As mentioned in Section 2.3.1, the meaning of an input argument defined as an input object type is essentially determined by the program code of the resolver functions. Thus the query example shown in Listing 3 illustrates one way that we make use of input objects to represent filtering conditions. In general, however, the input object types can be used in various ways for any field, depending on the implementation of the GraphQL server.

As mentioned earlier, domain users are the intended users of GraphQL servers, regardless of whether they have prior knowledge of the Semantic Web or ontologies. In order to write GraphQL queries, they only need to have a basic understanding of GraphQL, which can easily be explored via the GraphQL API provided by the server.

4.1. GraphQL schema formalization

According to [33,34], a GraphQL schema can be defined over five finite sets. These five sets are F ⊂ Fields , A ⊂ Arguments , T ⊂ Types , S ⊂ Scalars , and D ⊂ Directives where T is the disjoint union of O T (object types), I T (interface types), U T (union types), IO T (input object types) and S (scalar types). Fields , Arguments , Types , and Directives are pairwise disjoint, countably infinite sets representing field names, argument names, type names, and directive names, respectively. Scalars , which is a subset of Types , represents five built-in scalar types, which are Int , Float , String , Boolean , and ID . Moreover, the GraphQL schema definition language introduces non-null types and list types, called wrapping types, according to types in Types . Given a type t belonging to Types , the former is denoted as t ! , while the latter is denoted as [ t ] . W T is used to denote the set of all types that can be formed by wrapping the types in T , and W S denotes the set of all types that can be formed by wrapping the scalar types in S . In our current work, considering the knowledge representation language we use for the ontology (see next section), our GraphQL schema generator will neither produce union types nor directive definitions. Therefore, a GraphQL schema S is defined over ( F , A , T , S ) consisting of two assignments that are type S and implementation S :

OPEN IN VIEWER

Listing 5.

The formalization of the GraphQL schema shown in Listing 1

Listing 5 illustrates a formalized representation of the GraphQL schema shown in Listing 1. In the formalization, we have sets F , A , I T , O T , S and IO T , which contains all the field names, argument names, interface type names, object type names, scalar type names and input object type names, respectively. Additionally, the formalization contains field declarations in the set type S F ; argument declarations in type S AF ; object types implementing interface types declarations in implementation S . For instance, ( University , UniversityID ) ↦ String declares that the University type has a field UniversityID of which the returned type is String; ( ( Query , UniversityList ) , filter ) ↦ UniversityFilter declares that the UniversityList field accepts an input argument which is defined as the type UniversityFilter; Author ↦ { Professor } declares that the Professor type is one of the types that implement the interface Author.

In this work we assume that the ontology is represented by a TBox in a description logic which is an extension of FL 0 by adding qualified number restrictions and datatypes. FL 0 allows atomic concepts, the universal concept, intersection and value restriction [8]. This description logic can represent the semantics that can be reflected in a GraphQL schema for data access and integration where the schema follows the GraphQL schema definition language. Note that we do not aim to represent the full ontology in DL, but that, for our purposes, we only need the part of the ontology that is needed for data access. Therefore, we only use the DL constructors that cover the existing GraphQL features for data access.

The syntax and semantics of the description logic used in our approach are shown in Table 1. The introduction of datatypes is based on the work presented in [35] and [36]. Let N C , N R , N A , and D be disjoint finite sets of concept names, role names, attribute names, and datatype names respectively. An interpretation I consists of a non-empty set Δ I representing the domain of individuals, and an interpretation function · I . In addition, the interpretation includes an interpretation domain for data values Δ D I which is disjoint from the domain of individuals Δ I  [35]. A datatype, such as integer, is interpreted as a subset of Δ D I and a value such as the integer 5 is interpreted as an element of Δ D I . Thus, the interpretation function · I assigns to each atomic concept P ∈ N C a subset P I ⊆ Δ I , to each d ∈ D a set d I ⊆ Δ D I , to each role r ∈ N R a relation r I ⊆ Δ I × Δ I , to each attribute a ∈ N A a relation a I ⊆ Δ I × Δ D I , to each individual name i an element i I ∈ Δ I , and to each data value v an element v I ∈ Δ D I . (1) NF 1 : P ⊑ Q NF 2 : P ⊑ ∀ r . Q NF 3 : P ⊑ = 1 r . Q NF 4 : P ⊑ ∀ a . d NF 5 : P ⊑ = 1 a . d

A TBox over N C , N R , N A and D is a finite set of general concept inclusions (GCI) where each GCI is a statement in the form of B ⊑ C , where B and C are concepts. For generating GraphQL schemas, we use normalized TBoxes that contain only GCIs in the normal forms given in formula (1) where P , Q ∈ N C , r ∈ N R , a ∈ N A , and d ∈ D . There exist normalization rules to obtain such a TBox (e.g., an axiom P ⊑ Q ⊓ M is converted to two axioms, P ⊑ Q and P ⊑ M ) [9]. Baader et al. show that such normalization rules can preserve a conservative extension of a TBox in FL 0  [10]. A conservative extension guarantees that subsumptions according to the original TBox coincide with those with respect to the normalized TBox. An example TBox input is shown in Listing 6.

OPEN IN VIEWER

Listing 6.

TBox example representing the example ontology as shown in Fig. 1

Algorithm 1 shows the details to generate a GraphQL schema. The output for the example is the schema shown in Listing 1. First, the algorithm iterates over the concept names in N C (line 1 to line 5). For each concept, such as University in the TBox, the concept name (University) is used as the name of an object type to be generated (line 2); the term concatenated with ‘Filter’ is used as the name of an input type (UniversityFilter) to be generated (line 3); the term concatenated with ‘List’ is used as the name of a field (UniversityList) of the Query type (line 4). Additionally, each such field of the Query type is assigned an argument named ‘filter’, with a type that is the corresponding input type (e.g., filter:UniversityFilter to UniversityList) (line 5). Next, the algorithm iterates over GCIs in the TBox (line 6 to line 30). For a GCI in the form of NF 1 , the name of the super-concept is used as the name of an interface type to be generated; a field for the Query type named by concatenating the interface type name and ‘List’ is generated; the previously generated object type corresponding to the sub-concept implements the generated interface type.

From line 13 to line 21, the algorithm deals with GCIs containing roles, which can be of the form NF 2 or NF 3 (such as University ⊑ ∀ departments.Department). In both cases, a field definition (e.g., departments) of the object type (e.g., University) and a field definition (departments) of the input type (UniversityFilter) are generated. However, for NF 3 , the returned type of the field is defined as the original object type corresponding to the concept appearing on the right side of the GCI (line 20). For NF 2 , the returned type is defined as a wrapped type, which is a list type (line 17). For instance, the departments field declaration for the University type is departments:[Department]. The algorithm deals with GCIs containing attributes in a similar way (line 22 to line 30). For example, the University object type has a field declaration, which is UniversityID:String.

We define a function Φ for mapping a datatype that exists in the TBox to a scalar type in GraphQL. Due to the fact that current GraphQL supports five basic scalar types which are ID, Float, Int, Boolean, and String, our current implementation of function Φ focuses on mapping datatypes xsd:float, xsd:int, xsd:string and xsd:boolean to scalar types Float, Int, String and Boolean, respectively. However, GraphQL allows users to define custom scalar types, and the values of such custom types should be JSON serializable. Therefore, our Φ function can be extended in the future for mapping any datatype besides the above four types from a TBox into a custom scalar type in GraphQL.

By generating the GraphQL schema based on an ontology, the schema will contain object or interface types corresponding to concepts in the ontology, and field declarations corresponding to relationships in the ontology. When a GraphQL query is sent to the GraphQL server, a resolver function parses the query to determine which type in the schema is requested. It then parses the relevant definitions corresponding to such a type in the semantic mappings to retrieve data. For instance, if a query requests all the entities of the University object type, the resolver function parses the semantic mappings that are defined for the University concept to get information regarding how to access the underlying data sources.

4.4. The intended meaning of a generated GraphQL schema

In Section 4.3, we present the Schema Generator which takes a TBox representing an ontology as an input, to generate a GraphQL schema. Such a GraphQL schema can describe how to access underlying data sources in which the data can be annotated by the ontology. The underlying data can thus be viewed as an ABox for the TBox. Therefore, a GraphQL query that conforms to this GraphQL schema can be considered as a query over the ABox. To make this intention more formal we consider an ABox A as a finite set of assertions of the form P ( x ) , R ( x , y ) or A ( x , z ) , where P ∈ N c , R ∈ N R , A ∈ N A , x and y are instance names, z are literals. Listing 7 illustrates an example ABox based on the TBox in Listing 6.

OPEN IN VIEWER

Listing 7.

ABox example based on the example ontology as shown in Fig. 1

Let O be an ontology represented by a TBox  T ; let S be a GraphQL schema over ( F , A , T , S ) generated by Algorithm 1 based on T ; let Q be a GraphQL query over ( F , A , T , S ) such that Q conforms to S . Evaluating Q over underlying data sources that are instantiated in terms of O can be defined as retrieving an ABox A based on T :

For instance, given the query (as shown in Listing 3) and the above ABox, University(university_1), departments(university_1, d1), departments(university_1, d2), head(d1, “Harry, Potter”), head(d2, “Sheldon, Cooper”) are supposed to be retrieved. The above definition presents the meaning of the GraphQL schema generated based on a TBox for evaluating GraphQL queries. The definition relies on the Schema Generator where for each concept, the algorithm creates a corresponding type with the same name of the concept, same for roles and attributes. This guarantees to find the corresponding assertions from the ABox. However, in practice, as we presented in Section 2.3.2, how a GraphQL query retrieves data over the underlying data sources, depends on how the resolver function is implemented when we construct GraphQL servers. In the next section, we present how resolver functions can be implemented in a generic way based on semantic mappings.

5.1. GraphQL queries represented by abstract syntax trees

OPEN IN VIEWER

Fig. 3.

Example abstract syntax trees for the query shown in Listing 3.

In general, a GraphQL query can be represented using a single AST that contains nodes representing the fields requested in the query, and also contains additional nodes for the input arguments that may be used for each of these fields. In our approach, we assume that each query accepts an input argument which captures the notion of a filter condition. Therefore, we specify the query evaluation in two steps: (i) evaluating for a filter condition, which is represented via an input argument that is defined as an input object type in the schema, (ii) evaluating for those fields that are requested in the GraphQL query. For instance, in the query example shown in Listing 3, the field having a filtering condition is different from the requested fields (the former is UniversityID while the latter includes departments and head). In the evaluation step for the filter condition, the identifier information of the filtered out instances of the requested type (i.e., University) are obtained after accessing the underlying data sources. In the next step, the underlying data sources are accessed again to retrieve only the requested fields for the filtered instances. Therefore, to enable such two steps in the query evaluation, we use multiple ASTs to represent a GraphQL query (cf. Fig. 3, these two ASTs represent the query shown in Listing 3), one of which captures the input argument structure (Fig. 3a), and the other of which captures the structure of the query, including the requested fields and their types (Fig. 3b). More specifically, every node in such ASTs represents either a named type (i.e., object type, interface type, input type, or scalar type), a wrapping type, or a field. Additionally, ASTs that represent input arguments also contain nodes that represent the values of scalar-typed fields (e.g., “u1” in the AST shown in Fig. 3a). The types (i.e., UniversityFilter, StringFilter, String) or wrapping types (i.e., [University], [Department]) are drawn with rectangle nodes. The fields (i.e., UniversityID, _eq, departments, head) are drawn with rounded rectangle nodes.

In practice, a filter condition is converted into disjunctive normal form (DNF).7

OPEN IN VIEWER

Listing 8.

Example of RML mappings transforming university domain data, defined based on the example ontology as shown in Fig. 1

RML is a declarative mapping language for linking data to ontologies [51]. An RML document has one or more Triples Maps, which declare how input data is mapped into triples of the form (subject, predicate, object). An example of RML mappings is shown in Listing 8. A Triples Map contains the following three components (Logical Source, Subject Map and a set of Predicate-Object Maps). A logical source declares the source of input data to be mapped. It contains definitions of source that locate the input data source, reference formulation declaring how to refer to the input data, and logical iterator declaring the iteration loop used to map the input data. For instance, line 2 to line 6 in Listing 8 constitute the definition of a logical source. The definition declares that the data source is a JSON-formatted data source on the Web and also describes the way of iterating the JSON-formatted data (line 5). A subject map declares a rule for generating subjects when transforming underlying data into triples, including how to construct URIs of subjects (e.g., line 8) and specifying the concept to which subjects belong (e.g., line 9). A predicate-object map consists of one or more predicate maps declaring how to generate predicates of triples (e.g., line 12), and one or more object maps or referencing object maps defining how to generate objects of triples. An object map can be a reference-valued term map or a constant-valued term map. The former declares a valid reference to a column (relational data sources), or to an object (JSON data sources). The latter declares the value of the object as constant data. For instance, line 39 to line 41 make up a reference-valued term map. Line 19 to line 25 constitute a definition of a referencing object map including the join condition based on two triples maps. A referencing object map refers to another triples map (called a parent triples map) by using a rr:joinCondition property to state the join condition between the current triples map and the parent triples map. A join condition contains two properties, rr:child and rr:parent, of which the values must be logical references to logical sources of the current triples map and the parent triples map, respectively.

OPEN IN VIEWER

Fig. 4.

Technical components in the generic resolver function.

We show the basic technical components of the generic resolver function including QueryParser and Evaluator in Fig. 4. In Algorithm 2, we show the generic resolver function. The inputs to the generic resolver function are a GraphQL schema, a GraphQL query and semantic mappings. The GraphQL query and schema are inputs of the QueryParser. The QueryParser parses a query including a filter expression given as an input argument, and outputs the corresponding ASTs (Fig. 3) for the input argument and the query structure, respectively (shown as arrows 1 ◯ and 2 ◯ in Fig. 4). As mentioned in Section 5.1, in our practical solution a filter condition is converted into disjunctive normal form. In Algorithm 2, the QueryParser parses the query, converts a filter expression into a union of conjunctive expressions, and generates an AST for each conjunctive expression and an AST for the query structure (line 2). Then, the filter expression (line 5 to line 7 in Algorithm 2, frame a ◯ in Fig. 4) and the query fields (line 9 and line 13 in Algorithm 2, frame b ◯ in Fig. 4) are evaluated. The Evaluator is responsible for sending requests to underlying data sources and fetching data according to an AST. During evaluation of the filter expression, for each AST representing a conjunctive (sub-)expression, an evaluator is called to request data that satisfies the conjunctive (sub-)expression (line 6). After a call to an evaluator based on an AST (filter_ast in line 6), data representing the requested type, which contains identifier information, is returned (identifier_info in line 6). Taking the query in Listing 3 represented by the ASTs shown in Fig. 3 as an example, the requested type is University and data that can identify university instances is supposed to be returned in identifier_info. Such identifier information is captured in semantic mappings, which are used to construct the URIs for subjects where such subjects represent instances of the University concept. For instance, in line 8 of the RML mappings example in Listing 8, the values of the uid attribute of the underlying data source are used to construct URIs of subjects representing instances of the University concept. The identifier information returned by evaluating each filter_ast is merged into filtered_identifiers (line 7). During evaluation of the query fields, such merged identifier information is taken into account in the call to the evaluator of the query fields (line 9 in Algorithm 2, arrow 3 ◯ in Fig. 4).

As mentioned in Section 4.3, by generating the GraphQL schema based on an ontology, we can therefore, for each object or interface type and each field declaration, find the corresponding concept and relationship in the ontology. Since such concepts and relationships are used to define semantic mappings, when a generic resolver function retrieves data from the underlying sources of a requested type and relevant fields, it can therefore understand the semantic mappings regarding how to access underlying data sources and structure the returned data according to the GraphQL schema. Taking the query in Listing 3 represented by the ASTs shown in Fig. 3 as an example, as the requested type is University, the generic resolver function can therefore make use of relevant triples maps (line 1 to line 26 in Listing 8) defined in semantic mappings which are used for transforming underlying data following the semantics related to the University concept in the ontology.

OPEN IN VIEWER

Algorithm 3:

Evaluator

We present the details of Evaluator in Algorithm 3 and show an example in Fig. 5 of how evaluators work for answering the query in Listing 3. An AST and a number of triples maps from the semantic mappings are essential inputs to the algorithm. For a given AST, we can obtain the object type and fields that are requested in the query based on the root node and child nodes, respectively (line 2). For instance, taking the ASTs in Fig. 3 as examples, the root type and the field for evaluating the filter expression are University and UniversityID, and the root type and the first level requested field for evaluating query fields are University and departments, respectively. After getting the relevant triples maps based on the root node type (line 3 in Algorithm 3, e.g., UniversityMapping in Listing 8) or from the argument (line 16, the parent triples map, DepartmentMapping, which is an argument in the recursive call of an evaluator), the algorithm iterates over triples maps and merges the data obtained based on each triples map (line 4 to line 17). Exploring this in more detail, the algorithm parses each triples map to get the logical source and relevant predicate-object maps (line 5 and line 5). As described in Section 5.2, there are three different types of predicate-object map depending on the different maps of object, which are a reference-valued term map, a constant-valued term map or a referencing-object map. The algorithm iterates over the predicate-object maps and parses each one (line 5 to line 13). For a reference-valued term map, the mapping between the predicate and the reference column or attribute is stored (line 7, e.g., {UniversityID: uid} is stored in pred_attr), which will be used for rewriting a filter expression according to the underlying data source (line 14, e.g., uid = ‘u1’), annotating the obtained underlying data (line 15, e.g., HEAD is annotated as head for Department data). For a constant-valued term map, the mapping between the predicate and the constant data value, type is stored (line 9). For a template-valued term map, the mapping between the predicate and the template format is stored (line 11). All the pred _ attr , pred _ const , and pred _ template will be used to annotate the data from underlying sources (line 15).

OPEN IN VIEWER

Fig. 5.

Example for answering the query in Listing 3, (1) – (3) indicate the requests to and responses from the data sources; (a) – (c) indicate the parameter passing between the calls to evaluators; (4) indicates a recursive call to evaluator for getting the data of departments; frame  (A) indicates a join operation.

In the phase of evaluating a filter expression, local_filter, which represents the rewritten filter expression, is a necessary argument when sending requests to underlying data sources (line 14). While in the phase of evaluating query fields, filter_ids, being a NULL value or having at least one element, is a necessary argument (line 14, arrow  (a) in Fig. 5). A NULL value represents the fact that the GraphQL query does not include an input argument. After obtaining the data from the underlying data sources, the data is serialized into JSON format (key/value pairs) in which the keys are predicates stated in the predicate-object map (line 15), where each predicate corresponds to a field in the GraphQL schema. In the next step, the algorithm iterates over predicate-object maps in which the object map refers to another triples map (called a parent triples map) (line 15 to line 16). An evaluator is called again to fetch data based on this parent triples map (line 16, arrow  (4) in Fig. 5). For the query example, the parent triples map refers to the DepartmentMapping. Since such a referencing-object map definition states the join condition between the current triples map (UniversityMapping) based on child _ field (uid) and the parent triples map (DepartmentMapping) based on parent _ field (university_id) (line 21 to line 23 of the mappings in Listing 8), we can pass referencing data ( ref ), which contains the data obtained according to the current triples map and parent _ field , to the call of an evaluator when we fetch data according to the parent triples map (line 16). Such referencing data is taken into account, in the recursive call to an evaluator, when the request is sent to the underlying data sources (line 14, arrow  (b) in Fig. 5). After the data is obtained according to the parent triples map (arrow  (c) in Fig. 5), it is joined with data obtained according to the current triples map (line 16, frame  (A) in Fig. 5).

In the real case evaluation, we focus on a use case in the materials design domain where the task is data integration over two data sources, Materials Project [37] and OQMD (The Open Quantum Materials Database) [54].

OPEN IN VIEWER

Fig. 6.

Example of searching materials from materials project, OQMD and NOMAD.

Motivation The materials science domain, like many other domains, is at an early stage when it comes to introducing Semantic Web-based technologies into its data-driven workflows. A large number of research groups and communities have thus developed a variety of data-driven workflows, including data repositories [40,41] and data analytics tools. As data-driven techniques become more prevalent, more data is produced by computer programs and is available from various sources, which leads to challenges associated with reproducing, sharing, exchanging, and integrating data among these sources [1,38,39,53,61]. Figure 6 illustrates an example of searching for gallium nitride materials with the reduced chemical formula of Ga N in three databases of the materials design domain, Materials Project [37], OQMD [54] and NOMAD (Novel Materials Discovery) [27]. As shown in the results, each of them contains a column that represents chemical composition, but with different column names or different insights (i.e., ‘Formula’ for Materials Project and NOMAD, ‘Composition’ for OQMD). The ‘Formula’ column for Materials Project actually represents the reduced chemical formula. More detailed information regarding the chemical composition can be found based on the value of the ‘Nsites’ column. For instance, for the second row of the result from Materials Project, we can derive that the unit cell formula is Ga 2 N 2 based on the values of the ‘Formula’ and ‘Nsites’ columns. Meanwhile, the ‘Formula’ column for NOMAD represents the unit cell formula rather than the reduced chemical formula. Unlike the other two databases, OQMD contains a column for reduced chemical formulas, but with a different column name (‘Composition’). Such differences have to be addressed in order to integrate or exchange data from these data sources. Apart from such differences in terminology, the data that needs to be accessed or integrated from multiple data sources is typically heterogeneous in different models (i.e., relational data stored in relational databases, and hierarchical data stored in JSON data stores). In our previous work, we developed the Materials Design Ontology (MDO) [45] to enable ontology-driven data access and integration. Furthermore, we have applied our MDO and OBG-gen in the OPTIMADE (Open Database Integrations for Materials Design) [3] consortium which is a domain effort to make materials databases inter operable. As the work in the consortium is under development, our application is at the level of a proof of concept. Details are presented in [44, p. 141–148].

Data We collect data from the Materials Project and OQMD representing five different types of real-world entities (Calculation, Structure, Composition, Band Gap and Formation Energy). We define semantic mappings (for all the systems, see the next paragraph) based on MDO to interpret such data. We collect data in the sizes of 1K, 2K, 4K, 8K, 16K and 32K from each database for populating the five entities. The size 1K means 1000 entities of each entity type. We represent this data in different formats such as tabular data for relational databases and for CSV files, and JSON-formatted data for JSON files. Additionally, for HyperGraphQL and UltraGraphQL in our evaluation, we create an RDF file based on RML mappings and MDO for each dataset setting. We have six dataset settings for the experiments, which are 1K–1K, 2K–2K, 4K–4K, 8K–8K, 16K–16K and 32K–32K. Taking 2K–2K as an example, for each entity type, the test data contains data in the size of 2K from Materials Project and 2K from OQMD, respectively.

OPEN IN VIEWER

Fig. 7.

Outline of the real case evaluation.

Systems We compare our tool, OBG-gen in two versions (OBG-gen-rdb and OBG-gen-mix) with four systems: Morph-RDB [49], Ontop [50], HyperGraphQL [55], and UltraGraphQL [56]. OBG-gen-rdb represents the case where the generated GraphQL server handles data in relational databases, and OBG-gen-mix represents the case where the generated GraphQL server handles data not only in relational databases but also data in JSON and CSV formats. They take different RML mappings as inputs. Morph-RDB and Ontop are representatives from the group of OBDA-based tools. They can access relational databases as data sources by translating SPARQL queries into SQL queries based on semantic mappings, written in R2RML. As for the group of GraphQL-related tools, we intended to include Morph-GraphQL and Ontology2GraphQL in our evaluation. However, Morph-GraphQL fails to parse mappings; Ontology2GraphQL cannot be run due to a lack of detailed instructions regarding its setup. In the case of GraphQL-LD, since it focuses on querying Linked Data via GraphQL queries and a JSON-LD context using a SPARQL engine instead of a GraphQL interface, we did not consider it in our evaluation. Therefore, HyperGraphQL and its extension UltraGraphQL are the GraphQL engines that are included in our evaluation. They can query Linked Data that may be provided by local RDF files and remote SPARQL endpoints. The semantic mappings for all the systems in the evaluation are based on MDO. OBG-gen generates the GraphQL schema based on MDO. UltraGraphQL and HyperGraphQL use a modified version of the generated schema since they require directive definitions to specify the correspondences between query entries and the data. Figure 7 shows how the systems are configured in the evaluation. HyperGraphQL and UltraGraphQL are provided with the same RDF data for each dataset setting. OBG-gen-rdb is provided with two MySQL database instances hosting data from the Materials Project and OQMD respectively. Morph-RDB and Ontop are provided with one single MySQL database instance hosting data from the two sources. Conceptually, OBG-gen-mix is also provided with two database instances. However, each instance contains different formats of data such as data in a MySQL database, or in CSV or JSON files. More detailed, the instance for Materials Project has Composition data in JSON format and Band Gap data in CSV format. The instance for OQMD has Structure and Band Gap data in JSON format and Formation Energy data in CSV format. The data representing other entities for each instance is stored in MySQL database instances.

Queries We create queries that cover different features, aiming to evaluate our system based on qualitative aspects regarding what functionalities the system can satisfy and quantitative aspects regarding how the system performs over different data sizes. Additionally, we use competency questions stated in the requirements analysis of MDO to create queries with domain interests. The features of queries without and with filter expressions are shown in Table 3 and Table 4, respectively. From the perspective of GraphQL, we consider which choke point a query covers. The details of choke points are introduced in LinGBM.13

Table 5 shows more details of meanings of different filter expressions for Q6–Q12. The filter expressions for Q6 and Q12 are simpler than those for Q7–Q11 where the filter expressions have sub-expressions connected by boolean operators. Query features in terms of DI, and the filter expression form can help us understand systems qualitatively; Diffs and RS help in understanding systems quantitatively in the scaling analysis over different data sizes. We show Q1 in Listing 10 and Q7 in Listing 12. The results of these two queries are given in Listing 11 and Listing 13, respectively. Q1 requests all the structures containing the reduced chemical formula of each structure composition. Q7 requests all the calculations where the ID is in a given list of values, and the reduced chemical formula is in a given list of values.

Experiments and measurements We evaluate the query execution time (QET) of the different systems over the six dataset settings. Separately for each query, we run the query four times and always consider the first run to be a warm-up, then take the averaged value of the remaining three runs. Figure 8 illustrates the measurements over the six data sizes per query (Q1–Q12). Figure 9 and Figure 10 illustrate the measurements of all systems per data size for queries without filtering conditions and with filtering conditions, respectively. The measures for all data sizes and all queries are available online.14

OPEN IN VIEWER

Fig. 8.

Query Execution Time (QET) per query on materials dataset.

Results and discussion By analyzing the obtained measurements, we summarize three observations. The first observation is that both GraphQL servers generated by OBG-gen-rdb and OBG-gen-mix can answer all 12 of the queries covering different features (such as choke points). Therefore, the framework presented in Section 3 is feasible for data access and integration; this answers RQ1. Particularly, the GraphQL schema generated based on the ontology can provide an (integrated) view of underlying (heterogeneous) data; the generic resolver function based on the semantic mappings is capable of accessing heterogeneous data sources, combining the retrieved data (which may be in different formats), and structuring the data according to the GraphQL schema.

The second observation is regarding queries without filtering conditions (Q1–Q5) (cf. Figs 8 and 9). All of the systems have increases of QETs as the size of the dataset increases. However, Morph-RDB is less sensitive to the data size increase compared with other systems. UltraGraphQL and HyperGraphQL outperform other systems for some smaller datasets (e.g., UltraGraphQL’s QETs of Q1 and Q2, HyperGraphQL’s QETs for Q1 from 1K–1K to 4K–4K). We explain this by the fact that these two systems have additional context information declaring URIs of classes to which instances in the RDF data belong, which is unlike the other systems which have to make use of semantic mappings to output queries to be evaluated against the underlying data sources. OBG-gen-rdb outperforms Morph-RDB for some queries in smaller datasets (e.g., Q1 in 1K–1K, Q5 in 1K–1K and 2K–2K). For some queries, OBG-gen-rdb and Morph-RDB have close QETs (e.g., Q2 in 1K–1K). Ontop outperforms the other two in smaller datasets (e.g., Q1 in 1K–1K to 8K–8K, Q5 in 1K–1K to 4K–4K), but is more sensitive to data size increase compared with Morph-RDB.

OPEN IN VIEWER

Fig. 9.

Query Execution Time (QET) per data size on materials dataset for queries without filtering conditions.

OPEN IN VIEWER

Fig. 10.

Query Execution Time (QET) per data size on materials dataset for queries with filtering conditions.

The third observation is regarding how OBG-gen-rdb, Ontop and Morph-RDB perform for queries with filter conditions (Q6–Q12) (cf. Figs 8 and 10). Ontop outperforms the other two engines in most cases but is more sensitive to increases in dataset size (e.g., Q9 from 1K–1K to 8K–8K). According to [13], Ontop has a mapping optimization step which is not included in the query execution period. This could be a reason why Ontop outperforms the other engines. OBG-gen-rdb and Morph-RDB behave similarly for Q6 with stable QETs and Q12 with slight increases, as the data size increases. As Table 4 shows, the result size of Q6 is a constant over all the datasets in different sizes. Additionally, the filter expressions for Q6 and Q12 are simpler compared with those of Q7–Q11. Therefore, the QETs for evaluating filtering expressions for Q6 and Q12 are less than those of Q7–Q11. For other queries (Q7–Q11) Morph-RDB outperforms OBG-gen-rdb, however the differences between the two systems are less than those for queries without filtering conditions (e.g., Q1–Q4). The filtering conditions in GraphQL queries for OBG-gen-rdb and in SPARQL queries for Morph-RDB are written within WHERE clauses in SQL queries, thus will be evaluated against the back-end databases. A similar observation is also found in [17] where the experiment metrics shows that Morph-RDB outperforms other systems (e.g., Morph-GraphQL) as the size of dataset increase due to the SPARQL to SQL optimizations [17].

Based on the second and the third observations, we can answer the research question RQ2. The GraphQL servers generated by OBG-gen perform similarly compared with other systems for queries without filtering conditions, but are more sensitive to the increase of datasets even they can outperform for some queries in smaller datasets. By comparing OBG-gen-rdb, Ontop and Morph-RDB, we summarize the reasons as follows. As shown in Section 5, the implementation of OBG-gen is based on representing a GraphQL query with Abstract Syntax Trees (Fig. 3). In this way, two basic requests are sent to underlying data sources to get the data with respect to the semantic mappings. While for Morph-RDB and Ontop, based on semantic mappings, a SPARQL query is translated into a single SQL query. For queries with filtering conditions, all the three engines (OBG-gen-rdb, Morph-RDB and Ontop) can take the advantages of rewriting filter conditions into SQL queries so that the QETs do not show a significant increase as the data size increases.

To show the generalizability of our system, we conduct an evaluation based on LinGBM. LinGBM provides tools for generating datasets (data generator)15

Data The dataset generated by the data generator is a scalable, synthetic dataset regarding the University domain, including several entity types (e.g., University and Department). We generate data in scale factors ( sf ) 4, 20 and 100 where a scale factor represents the number of universities [20]. We then create three MySQL database instances to store the data in these three scale factors, respectively. We use a modified version of the GraphQL schema provided by LinGBM for our GraphQL server, and define RML mappings according to the work in Morph-GraphQL.18

OPEN IN VIEWER

Listing 9.

A query according to QT5, such a query goes from a given department to its university, then retrieves all graduate students who get the bachelor’s degree from the university, then comes back to department. This cycle is repeated two times

Queries The experiments are performed over eight query sets, where each set contains 100 queries that are generated using the LinGBM query generator based on a query template (QT). A query template has placeholders for input arguments. The query generator can generate a set of actual queries (query instances) based on a query template in which the placeholder in the query template is replaced by an actual value. We select eight query templates (QT1–QT6, QT10 and QT11) for constructing eight query sets (QS1–QS8). We show an example query according to QT5 in Listing 9. The other six query templates from LinGBM require GraphQL servers to have implementations for functionalities such as ordering and paging which are not considered currently by OBG-gen. However, these functionalities are interesting for future extension of OBG-gen.

Experiments, results and discussion Same as the real case evaluation, we evaluate the query execution time (QET) of our system on the three datasets. Each query from a query set is evaluated once. We show the average query execution times for the different query sets in Table 6. Based on the obtained measurements, we observe that our system has slight increases for QS1, QS2, QS4, QS6 and QS7 in terms of the average QETs. For QS3, the average QET is stable for all the three datasets. For QT5, the increase from 0.51 seconds at data scale factor 20 to 13.85 seconds at data scale factor 100 is due to the dramatic increase in result size. More specifically, the queries in QS5 and QS8 need to access the ‘graduateStudent’ table which increases dramatically in size from 50,482 rows in the table ( sf = 20 ) to 252,562 ( sf = 100 ). This is the reason for the average QET of QS8 increasing in sf = 100 . Additionally, each query in QS5 repeats a cycle two times (‘university’ to ‘graduateStudent’ to ‘university’) and requests the students’ emails and addresses along the way. This causes the larger increase in average QET of QS5. The above synthetic experiments indicate that our system can work in another domain than the materials science domain.

We furthermore demonstrate the generalizability of our system by evaluating it against GTFS-Madrid-Bench, which is a benchmark for evaluating OBDI systems.

Data, queries and systems The dataset provided by GTFS-Madrid-Bench is a scalable dataset regarding the Transport domain (the metro system of Madrid), including several entity types (e.g., Route, Stop, Shape and Trip). We use the data generator provided by GTFS-Madrid-Bench to generate data in scale factors ( sf ) 1, 5, 10 and 50. For instance, the dataset in sf 1 contains 13 instances for the Route type, 1,262 instances for the Stop type, 58,540 instances for the Shape type and 130 instances for the Trip type. An increase in sf from 1 to 5 results in an increase in the dataset size of 5 times (e.g., 65 instances for the Route type and 6,310 instances for the Stop type). Each instance is represented by a row in the corresponding relational table. These scale factors are also used for the experiments in [18]. We then create four MySQL database instances to store the data in these four scale factors, respectively. A total of 18 queries are included in the GTFS-Madrid-Bench benchmark that cover the different features of SPARQL 1.1. For conducting the experiment based on GTFS-Madrid-Bench, we select seven queries (Q1–Q5, Q9, Q13) to create corresponding GraphQL queries. Among these four queries, Q1 retrieves all the shape entities where each shape entity is a polygon associated with a trip; Q2 retrieves all the stop entities where the latitude is greater than a specific value; Q3 retrieves accessibility information of all stop entities; Q4 retrieves all the route entities and their associated agency entities; Q5 retrieves services that have been added after a specific date; Q9 retrieves trips and associated shapes with latitude higher than a specific value; Q13 retrieves all the accesses of the stations [18]. The other queries contain SPARQL 1.1 features such as order by, group by and distinct. Currently, OBG-gen does not implement functionalities to cover these features. However, these functionalities are interesting for future extension of OBG-gen. In addition to OBG-gen-rdb, we conduct experiments based on Ontop to learn how two engines behave in this GTFS-Madrid-Bench benchmark scenario.

Experiments, results and discussion Same as the previous two evaluation scenarios, we evaluate the query execution time (QET) of systems on different datasets. We show the measurements in Table 7. According to the measurements, both OBG-gen-rdb and Ontop show increases in QETs for all four queries as the dataset increases. However, as with the observation in the real case evaluation, Ontop behaves less sensitively to the increase in dataset. In terms of how the two systems behave for different queries, both engines spend more time to answer Q1 (without any filter conditions) and Q9 (with several relationship retrievals). For answering Q1, OBG-gen-rdb spends more than 3,600 seconds for scale factors 10 and 50. Although Ontop is able to answer Q1 in less time than OBG-gen, it cannot finish the execution because it runs out of the reserved 4 GB memory for scale factor 50. More specifically, Q1 needs to access the ‘Shape’ table which increases dramatically in size from 58,540 rows in the table ( sf = 1 ) to 292,700 ( sf = 5 ) and furthermore to 585,400 ( sf = 10 ) and 2,927,000 ( sf = 50 ). For answering Q9, OBG-gen can only return query result for scale factor 1, while Ontop for just scale factors 1 and 5. Both engines have relatively stable QETs for Q4 that retrieves all the route entities without any filter conditions, and for Q5 that retrieves services with a filter condition. The corresponding ‘Route’ table is relatively small (e.g., 13 for sf 1 and 1,300 for sf 100). Q2 and Q3 retrieve all the station entities but with different filter conditions. The two engines spend more time to evaluate Q2 since the result size of Q2 is larger than that of Q3. OBG-gen has an obvious increase for answering Q13 from scale factor 10 to 50. This is because that the corresponding retrieved table, ‘Stop’ increases in size from scale factors 10 to 50 (12,620 rows to 63,100 rows).

For evaluating our approach, ontology-based GraphQL server generation, we conducted an experiment motivated by the materials design domain and experiments based on two synthetic benchmark scenarios (LinGBM and GTFS-Madrid-Bench). Based on the measurements of these experiments, we can answer the three research questions presented at the beginning of Section 7. Our approach can generate GraphQL servers for data access and data integration and can be used in various domains (RQ1 and RQ3). The other GraphQL interfaces, HyperGraphQL and UltraGraphQL, can be used for data integration to a limited extent due to the fact that they do not support various filter conditions. This means questions with filter conditions cannot be answered. By comparing our approach with other well-known systems (RQ2), we learn that our system can perform relatively similar to others (e.g., Morph-RDB) in terms of QETs for queries with filter conditions (as shown in Fig. 10), and for some queries without filter conditions in smaller datasets (as shown in Fig. 9). Morph-RDB and Ontop are both less sensitive to the data size increase. The reason for this can be explained by the fact that they have optimization techniques that enable queries to be executed in a shorter amount of time. For instance, Ontop has a mapping optimization step which is not included in the query execution period [13,62]; Morph-RDB has a query rewriting optimization step where projections and selections are pushed down for removing non-correlated subqueries [52,62]. However, our approach supports data integration where the underlying data is from different kinds of sources (i.e., OBG-gen-mix), in contrast to Morph-RDB that only support data integration where the underlying data is from relational databases. This is due to the implementation of the generic resolver function presented in Section 5, that combines data from multiple sources according to a GraphQL schema, as an integrated view of underlying data. Moreover, OBG-gen-mix can integrate data where underlying data is provided in different ways. In contrast, the other systems require the underlying data to be materialized as RDF triple stores (e.g., UltraGraphQL and HyperGraphQL) or to be stored in one single database (e.g., Morph-RDB and Ontop).