1. RDF Storage methods and Systems Nikolaou Charalampos (A.M.: M953)‏ Kotsifakos Alexios (A.M.: M964)‏ Department of Informatics and Telecommunications

2. 15/03/082 Presentation Outline RDF Triple Table Means of Storage Storage Methods RDF Storage Systems Storage Systems’ Evaluation

3. 15/03/083 Storing RDF data RDF triple:, e.g. : RDF database: a single table with three columns, subject, property, object

4. 15/03/084 Why studying different RDF storage methods? (1/2)‏ RDF triple store has the following disadvantages: –All interesting queries require many self-joins over the rdf table: SQL Query SELECT p5.obj FROM rdf AS p1, rdf AS p2, rdf AS p3, rdf AS p4, rdf AS p5 WHERE p1.prop = ’title’ AND p1.obj ˜= ’Transaction’ AND p1.subj = p2.subj AND p2.prop = ’type’ AND p2.obj = ’book’ AND p3.prop = ’type’ AND p3.obj = ’auth’ AND p4.prop = ’hasAuth’ AND p4.subj = p2.subj AND p4.obj = p3.subj AND p5.prop = ’isnamed’ AND p5.subj = p4.obj; RDF Triple Table person1 isNamed ‘‘Serge Abiteboul’’ person2 isNamed ‘‘Rick Hull’’ person3 isNamed ‘‘Victor Vianu’’ book1 hasAuthor person1 book1 hasAuthor person2 book1 hasAuthor person3 book1 isTitled ‘‘Foundations of Databases’’

5. 15/03/085 Why studying different RDF storage methods? (2/2)‏ -As the number of triples increases, the execution time of queries is also increased, due to memory consumption => more indexes required -Inference at assertion time is infeasible as the number of triples scales, since for each entailment rule many more triples are stored

6. 15/03/086 Triple store requirements – mapping considerations Triple Store Requirements Text searching URIs Datatypes RDF Containers (rdf:Seq, rdf:Bag, rdf:Alt, rdf:_n)‏ RDF Vocabulary Description Language support (aka RDF Schema)‏ Ontological support, Inferencing Triple provenance Mapping considerations The database schema The particular database implementation Non-relational databases such as ODBMS, XML Database tuning Database updates Exposing the database schema

7. 15/03/087 Means of storage Memory Hard Disk - Database - File System (Native)‏

8. 15/03/088 Storage Methods (1/2)‏ Schema-oblivious (also called generic or vertical): One ternary relation is used to store any RDF/S schema or resource description graph. (Figure 1)‏ Schema-aware: (also called specific or binary): One table per RDF/S schema property or class is used. (Figure 2)‏ Hybrid: One ternary relation for every different property range type and a binary relation for all class instances (as in schema-aware). Property (class) instances with range values of the same type are stored in the same relation, distinguished by the property (class) id (as in schema-oblivious). (Figure 3)‏

9. 15/03/089 Storage Methods (2/2)‏

10. 15/03/0810 Method Variations (1/2)‏ Schema-oblivious –URI: stores the URIs in the table holding the triples –ID: relies on integer identifiers to represent resources and properties in the triple table and stores them only once in a separate table

11. 15/03/0811 Method Variations (2/2)‏ Schema-aware –ISA: exploits the object-relational features of SQL99 for representing subsumption relationships using sub- table definitions –NOISA: stores RDF/S data using a standard relational representation as depicted in Figure 2 –Vertical Partitioning: NOISA + sort by subject + value column for each table can be optionally indexed

12. 15/03/0812 Extension to Schema-oblivious (1/2)‏ Property Tables –Clustered Property Table: Clusters of properties that tend to be defined together (Figure 4)‏ –Property-class Table: Clusters similar sets of subjects together in the same table (Figure 5)‏

13. 15/03/0813 Extension to Schema-oblivious (2/2)‏ Figure 4: Clustered Property Table Example Figure 5: Property-class Table

14. 15/03/0814 Advantages – Disadvantages (1/5)‏ Schema-oblivious (+) Straightforward Schema Evolution (-) Disregards Type Information (-) Significant storage overhead (URI)‏ (-) Additional join operation at the end of every query (ID)‏

15. 15/03/0815 Advantages – Disadvantages (2/5)‏ Schema-aware (-) Addition/deletion of a new property requires the addition/deletion of a table (-) Significant overhead when managing a potentially large number of tables (++) Subsumption is implicitly supported by the schema (ISA)‏ (+) Internal encoding of subsumption => efficient evaluation of taxonomic queries in secondary storage (NOISA)‏

16. 15/03/0816 Advantages – Disadvantages (3/5)‏ Schema-aware (NOISA)‏ (+) Support for multi-valued attributes (+) Support for heterogeneous records (+) Only those properties accessed by a query need to be read (+) No clustering algorithms are needed (+) Fewer unions for one property and fast joins (-) Join is not free (-) Slower inserts

17. 15/03/0817 Advantages – Disadvantages (4/5)‏ Hybrid (+) Schema evolution easily supported (+) Preserves type information (+) Internal encoding of subsumption => efficient evaluation of taxonomic queries in secondary storage (NOISA)‏

18. 15/03/0818 Advantages – Disadvantages (5/5)‏ Property Tables (+) Reduce subject-subject self-joins of the triples table (+) Speed up queries that can be answered from a single property table (-) Most queries require joins or unions to combine data from several tables (-) RDF data tends not to be very structured (NULL)‏ (-) Multi-valued attributes cause further complexity (-) Property clustering must be carefully done to avoid creating too wide tables (-) Problematic queries that do not select on class type (Property-class tables)‏ (-) Unspecified property (clustered property tables)‏

19. 15/03/0819 Inference Support Precompute inferring triples (compile time)‏ –Schema-oblivious (URI, ID)‏ Compute on demand (run time)‏ –Schema-aware (ISA, NOISA)‏ –Hybrid Computation in: –Main memory (ISA)‏ –Secondary memory (NOISA, Hybrid)‏

20. 15/03/0820 Column-Oriented DBMS Tuples are stored in column format instead of standard row format Query evaluation –Only columns relevant to a query read into memory and not entire rows before projection occurs (wasting bandwidth)‏ –Inserts might be slower in column-stores (especially if they are not done in batch)‏ Vertical partition uses column format for better performance

21. 15/03/0821 Performance Results (1/2)‏ Figure 6: Performance comparison of the triple-store schema with the property table and vertically partitioned schemas. Property tables contain only the columns necessary toexecute particular query.

22. 15/03/0822 Performance Results (2/2)‏ Figure 7: Performance as number of triples scale for a specific query

23. 15/03/0823 RDF Storage Systems Jena1, Jena2 ICS-FORTH RDFSuite Sesame 3store RStar Oracle

24. 15/03/0824 Jena1 (1/2)‏ Storage method: Schema-oblivious (ID)‏ Supports multiple graphs stored in the Statement Table (one more attribute specifying the graph id)‏ Supports reified statements by adding a statement id attribute to the Statement Table (does not allow multiple reified instances of any statement)‏

25. 15/03/0825 Jena1 (2/2)‏ Figure 8: Jena1 Schema (Normalized)‏

26. 15/03/0826 Jena2 (1/3)‏ Storage method: Schema-oblivious (URI)‏ Supports storage of similar graphs in multiple statement tables (one more attribute specifying the graph id)‏ Supports clustered property tables and property-class table Reified statements are stored in property-class table (allows multiple reified instances of any statement)‏

27. 15/03/0827 Jena2 (2/3)‏ Multi-valued properties may be clustered or may be stored in a separate table To face space consumption –Common prefixes in URIs are stored in a separate table and the prefix itself is replaced by a db reference (prefix in cache)‏ –Long values are stored only once (defined by a threshold)‏

28. 15/03/0828 Jena2 (3/3)‏ Figure 9: Jena2 Schema (Denormalized)‏

29. 15/03/0829 Jena1 vs. Jena2 The denormalized schema of Jena2 is faster than the normalized schema of Jena1, twice as fast for many operations For queries with high selectivity Jena1 and Jena2 perform about the same, but Jena2 performs better on queries which join a large number of tuples

30. 15/03/0830 ICS-Forth RDFSuite (1/3)‏ Storage method: Schema-aware (ISA)‏ Core RDF/S model: Class, Property, SubClass and SubProperty tables to represent RDFS hierarchies Uses a Namespace table Uses a Type table to hold the built-in types of RDF/S (e.g. rdf:Property, rdfs:Class, etc), Literal types (e.g. integer, date), Container types (e.g. rdf:Bag)‏ For each class/property there is a table in which its instances are stored

31. 15/03/0831 ICS-Forth RDFSuite (2/3)‏ Figure 10: Relational Representation of RDF Description Bases

32. 15/03/0832 Variations of RDFSuite (3/3)‏ Construct a unified Instance table containing all instances of all classes Construct a unified Instance table containing all instance properties of all properties For every property with range a literal type, add an attribute to the class table of its domain (this is for single valued properties which are not specialized)‏

33. 15/03/0833 Sesame (1/5)‏ Storage method: Schema-oblivious (ID)‏ Memory, Database and File System Support Abandonment of object-relational backend –One table for each class/property –The semantics of subtables do not match the semantics of rdfs:subClassOf (multiple parents)‏

34. 15/03/0834 Sesame – Database (2/5)‏ Figure 11: Sesame RDBMS Schema

35. 15/03/0835 Sesame - Memory (3/5)‏ Figure 12: Bipartite graph representation of an RDF model

36. 15/03/0836 Sesame – File System (4/5)‏ Read/write with Java’s I/O classes Stores indices (B-trees, hash tables) for quick searching Employs selective caching of data in memory for increasing retrieval performance

37. 15/03/0837 Sesame - Performance (5/5)‏ Figure 13: Left: first 20 min., Right: complete upload

38. 15/03/0838 3store (1/2)‏ Figure 14: 3store Database Schema

39. 15/03/0839 3store (2/2)‏ Datatypes: All datatypes are stored as strings => determination of datatype is done by conversion at runtime Hybrid approach for the production of entailment rules: –Those generating fewer entailments are evaluated at assertion time with forward chaining rules –Those with greater storage cost and a lower evaluation cost are evaluated at query time with backward chaining rules

40. 15/03/0840 RStar (1/2)‏ Storage method: Schema-oblivious (ID)‏ Core RDF/S model: Class, Property, SubClass, SubProperty and Property-Class tables to represent RDFS hierarchies Uses a Triples table Uses a Namespace table Uses a Type table to hold Literal types (e.g. integer, float, date, string)‏ For efficient searching all property, class, resource and data names are hashed

41. 15/03/0841 RStar (2/2)‏ Figure 15: A database schema of RStar

42. 15/03/0842 Oracle Storage method: Schema-oblivious (ID)‏ One table for triple storage (IdTriples), one table for mapping URIs and Literals to IDs (UriMap)‏ Supports user views on selective portions of the RDF data Supports multiple representations for the same value (canonical literal ID)‏ Pre-defined datatypes are partitioned into families according to their value space (e.g. numeric family)‏ To speed up query answering “subject property matrices” are used as auxiliary tables

43. 15/03/0843 Storage Systems’ Evaluation (1/5)‏ Two storage systems based on memory (Sesame-Memory, Jena-Memory)‏ Two persistent storage systems with RDBMS (Sesame-DB, Jena-DB)‏ Three native RDF approach systems (Sesame-Native, Kowari, YARS)‏

44. 15/03/0844 Storage Systems’ Evaluation Query response time (2/5)‏ Sesame-DB, Jena-DB slowest systems: –Query mechanism represented by a series of join instructions –Store most triples in a single table Sesame-Memory, Jena-Memory -> quick Native systems –Sesame-Native like Sesame-Memory (use same query evaluation processing)‏ –Kowari solid performance –YARS several times faster than Sesame, using index

45. 15/03/0845 Storage Systems’ Evaluation Data loading time, repository size (3/5)‏ Figure 16: Data loading time and repository size

46. 15/03/0846 Storage Systems’ Evaluation Data loading time, repository size (4/5)‏ Memory-based systems -> best performance DB-based systems -> bad scalability –Sesame-DB: a) performs duplication check, without caching mechanism, since it employs a unique ID to RDF node mapping system, b) re-invokes its inference engine on loading RDF file Native systems -> nice performances –Sesame-Native, YARS use B-trees. YARS uses additional indices => more space –Kowari even more spaces using 64-bit storage

47. 15/03/0847 Storage Systems’ Evaluation (5/5)‏ fixed IBM DB2Forward ChainingRSQL Rstar fixed Oracle, any RDBMS (Btrees, table functions)‏ Forward ChainingSPARQL-like, RDQL- like Oracle fixed MySQLHybridRDQL 3store fixed and non- fixed PostgreSql, any RDBMS Backward ChainingRQL RDF Suite non-fixed PostgreSql, MySQL, Oracle9i Forward chainingRDQL, RQL, SeRQL Sesame non-fixed PostgreSql, MySQL, Oracle Backward/Forward and hybrid chaining RQL, RDQL, SPARQL Jena SchemaDBMSInference support Query Language System

48. 15/03/0848 Thank you!!!

49. 15/03/0849 References (1/2)‏ Y. Theoharis, V. Christophides and G. Karvounarakis, Benchmarking Database Representations of RDF/S Stores. Fourth International Semantic Web Conference (ISWC'05), Galway, Ireland, November, 2005. Abadi, Daniel J., Marcus, Adam, Madden, Samuel R., and Hollenbach, Kate. Scalable Semantic Web Data Management Using Vertical Partitioning. In Proceedings of VLDB 2007. Stephen Harris, Nicholas Gibbins. 3store: Efficient Bulk RDF Storage, In Proceedings of the 1st International Workshop on Practical and Scalable Semantic Systems(PSSS'03). Jeen Broekstra, Arjohn Kampman, and Frank van Harmelen, Sesame: A Generic Architecture for Storing and Querying RDF and RDF Schema. In Proceedings of ISWC 2002. S. Alexaki, V. Christophides, G. Karvounarakis, D. Plexousakis, K. Tolle, The ICS-FORTH RDFSuite: Managing Voluminous RDF Description Bases, 2nd International Workshop on the Semantic Web (SemWeb'01), pp. 1-13, Hongkong, May 1, 2001. L. Ma, Z. Su, Y. Pan, L. Zhang, T. Liu: RStar: An RDF Storage and Query System for Enterprise Resource Management. In Proc. of the ACM CIKM 2004.

50. 15/03/0850 References (2/2)‏ K. Wilkinson, C. Sayers, H. A. Kuno, D. Raynolds: Efficient RDF Storage and Retrieval in Jena2. In Proc. of SWDB'03 (co-located with VLDB'03). E. I. Chong, S. Das, G. Eadon, and J. Srinivasan. An Efficient SQL-based RDF Querying Scheme. In VLDB, pages 1216–1227, 2005. D. Beckett, J. Grant, SWAD-Europe: Mapping Semantic Web Data with RDBMSes, http://www.w3.org/2001/sw/Europe/reports/scalable_rdbms_map ping_report/ J. Broekstra, “Storage, Querying and Inferencing for Semantic Web Languages”, Phd Thesis. Liu Baolin, H. Bo, An Evaluation of RDF Storage Systems for Large Data Applications, In Proc. of First International Conference on Semantics, Knowledge, and Grid (SKG 2005), 2005.