Much of the data found in practice resides in relational DBs. However, many contemporary analytical tasks are performed on graphs. Property graphs are currently one of the most prevalent data models for graph data management in industry. Therefore, a key challenge is to understand the fundamental relationships between relational databases and property graph databases. This paper reports our ongoing work towards understanding these relationships by proposing R2PG-DM, a direct mapping of relational databases to property graphs. Given a relational database schema and instance, a direct mapping generates a corresponding property graph instance. The semantics of our mapping is de ned using Datalog. Our work is inspired by existing approaches for direct mappings of relational databases into earlier graph data models. Future work is to study our mapping with respect to fundamental properties such as information and query preservation.
A major class of contemporary data analytics focuses on gaining insights from the
rich complex patterns found in graph-structured data collections such as social,
communication, nancial, biological, and mobility networks [
The O shore Leaks PG was constructed as a mapping from relational database (RDB) sources.4 Indeed, much of the data found in practice resides in RDBs. Therefore, a key challenge is to understand the fundamental relationships between RDBs and PGs. A crucial rst step in exploratory graph analytics on
RDBs is to transform the database into a PG. Only then can the basic graph structure be explored and new relationships discovered using contemporary graph database systems.
This motivates the study of direct mappings from RDBs to PGs. A direct mapping is a transformation from RDBs to PGs which (1) is domain and schemaindependent, i.e., works regardless of the source database schema and instance, and (2) transforms the content of the source instance into a target instance, i.e., given a RDB instance generates a corresponding PG instance.
State of the art. Most approaches to graph analytics over relational data
extract graphs as user-speci ed views on relational data, e.g., [
The study of direct mappings from relational to property graphs is not as
well-developed. Of the small handful of approaches, the focus has been on
optimized layout for e cient query processing or mappings which are lossy or
obfuscate the input RDB schema [
Let D be an enumerable set of data values containing the special value null, and let A be an enumerable set of attribute names containing the special attribute tid. An RDB schema is a triple S = (R; att; ), where R is a nite set of relation names, att is a function assigning to each r 2 R a nite set att(r) A n ftidg, and is a nite set of primary and foreign keys over R and att.5 For r 2 R, an r-tuple is a function t : att(r) [ ftidg ! D such that t(tid) 6= null. An instance I of S is an assignment to each r 2 R of a nite set I(r) of r-tuples satisfying , such that for distinct t; t0 2 Sr2R I(r) it holds that t(tid) 6= t0(tid). An example schema and instance is given in Figure 1 (top left).
A property graph is a structure (V; E; e; `; p) where V is a nite set of vertices, E is a nite set of edges, e is a function assigning an ordered pair of vertices to each edge (i.e., the source and target vertices of the edge, resp.), ` assigns a nite set of labels to each vertex and edge (from some domain of labels), and p assigns a nite set of key-value pairs to each vertex and edge. An example property graph is given in Figure 1 (bottom left). 5 https://en.wikipedia.org/wiki/Foreign_key
On Directly Mapping Relational Databases to Property Graphs
Let PG denote the set of all PGs and RDB denote the set of all pairs (S; I) where S is an RDB schema and I is an instance of S. A direct mapping is a total function from RDB to PG. 3
We next informally present R2PG-DM, a direct mapping which addresses the shortcomings of current solutions discussed in Section 1. With R2PG-DM, RDB relations are interpreted as classes of \things" (i.e., labeled vertices) and foreignkey relationships are interpreted as edges between these things. In particular: { each input r-tuple t is represented in the output graph by a vertex v (identi ed by t(tid)), which is labeled with r, the name of the relation to which t belongs; furthermore, v has key-value pair (a; t(a)), for each a 2 att(r). { for each foreign key relationship f 2 (from, say, relation r to relation s), for each pair of tuples t 2 I(r) and t0 2 I(s) participating in f , this relationship is represented by an edge with label r-s from vertex vt to vertex vt0 , representing t and t0, respectively.
Similarly, as there currently does not exist a standard PG schema language, the input schema is also fully represented in the output PG, e.g., each relation and each attribute of a relation is represented by a vertex, and foreign keys are represented by edges between the relevant attributes of relations, etc. Figure 1 illustrates an RDB database (S; I) and PG translation R2PG-DM(S; I). We omit here the translation of S to PG, and only illustrate the translation of I.
Clearly, each component of the R2PG-DM mapping can be speci ed by a
declarative non-recursive Datalog query [
Our ongoing work proceeds along three lines. First, we are proceeding with a
formal study of R2PG-DM, establishing basic properties such as information and
query preservation [