Virtuoso, The Prometheus of RDF -- Sematics 2014 Conference Keynote

Virtuoso:
The Prometheus of
RDF-based Relational Data
Management
By Orri Erling
Virtuoso Program Manager
OpenLink Software

Linked Data at Dawn
 The Promise and the Practice
 The Science of Speed
 The Structure which Is
 Ongoing Research
License CC-BY-SA 4.0 (International).

Linked Data Promises
 RDF is a generic, minimalistic
model for describing things
 RDF has global identifiers and
data is self-describing
 URIs may be dereferenceable
 RDF is flexible to query,
does not force a single hierarchical view like XML

Linked Data Scenarios
 RDF is used because of
 schema flexibility
 global identifiers
 Inference, if present, is usually trivial
 Subclass
 Sub-property

Where Triples Come From
 Relational extracts or web content is converted to
and stored as triples
 NLP extraction
 New applications with RDF as primary data model
 Doing SPARQL against data in RDBs is possible
but is rare and does not deliver the flexibility

Linked Data Verticals and Patterns
 Publishing: tagging & annotations, evolving vocabularies
 Archives: self description, long term identifiers, many
versions of schema
 Semantic search: structured, semi-structured, and
full text, all in one
 Business intelligence: many sources, ease of adding
sources, no 6 month DW schema change cycle
 E-science, often in life sciences: common interchange
format, nano-publications, NLP extracts, different users
cook their data differently, provenance

The Hopes and Perceptions
 The age of ad hoc
 Find insight in any data, when you need it,
from any source, any format
 No data warehouse planning cycles; make your own
from the pieces you need, when you need it
 Still, data integration remains hard work; quality and
coverage of sources vary
 Flexibility may be there, but is performance and
scalability on the level?

Yes, But ...
 Web and Big Data: Everybody reinvents the triple.
Self-description, long term identifiers, key-value pairs
in many non-RDF use cases
 SPARQL and RDF would be the natural, standards-compliant
choice if did beat SQL, information retrieval,
custom big data, key value, map reduce solutions
Is this intrinsic to linked data or is this lack of engineering?
 Linked data has unique advantages in breadth of
coverage and expressivity but performance must not lag
behind.

What is the RDF Tax?
 90% of bad performance comes from
non-optimal query plans
 Some comes from indexing too much
(e.g., SQL bulk load with no indices is 50x faster
than the equivalent in RDF with all indexed)
 Some comes from string ops on URIs, literals
 Some comes from having a join for every attribute.
Vectoring and right plans help, though

The Bane of the Triple
When data is stored as triples:
 There is structure still but it is harder to exploit.
Schema re-emerges as correlations
 More joins make more possible query plans,
bigger errors in plan cost estimation
 More joining reduces locality
 Lack of schema causes needless indexing;
data takes more space
 A URI for everything takes space and time
For the same workload, Virtuoso SQL can also be 2–20x
faster than Virtuoso SPARQL

The Question is Raised
 LOD2 FP7, now ending: RDF Performance parity with
relational?
 SQL is the senior science. Who ignores history is bound
to repeat it
 Integral mastery of RDB science is a prerequisite, but
do not forget the subtle twists of schema-less-ness

Virtuoso RDF Relational DBMS Leadership
 2000–2006, v1.x–4.x: SQL row store with SQL
federation and XML
 2007–2008, v5.x–6.x: SPARQL, adapted for RDF quads
with more compression, bitmap indices, special data
types, RDF awareness in query optimization
 2009, v6.x: Scale-out cluster-capable
 2010–2013, v7.x: Column store, vectored execution, 3x
more space efficient, 10+x more speed
 2013: Star Schema benchmark with SPARQL, 100x
MySQL SQL, 0.8x MonetDB SQL
 2014: Top of the line SQL analytics, 500 Gtriples,
Structure Awareness

Triples Done Right, so?
 Column-store techniques are a good fit; index-based
triple storage does not get much better
 RAM-only pointer-based techniques can be faster but
cost 10–100x more to scale up
 To take RDF to SQL parity, Virtuoso must first be on the
level with the best in SQL
 TPC-H is the checklist for mastery of DW and query
optimization; who survives shall not fear
 Parity is achieved when running with triples, just like
with tables

Structure is Everywhere
CWI in LOD2:
 90% of triples in Common Crawl fall into 20 tables
 All relational extractions are 100% tables
 Even DBpedia is 90% covered by 500 tables, but is
unusually heterogeneous, albeit not very large

The Glorious Dawn:
Structure is the Servant, not the Tyrant
 A set of subjects with all the same single-valued
properties is in fact a table.
 So, store it as a table
 Allow exceptions, e.g., sometimes multiple values,
different values in different graphs, extra properties, etc.
 If it is big, it has repeating structure
 All RDF semantics are preserved; any triple is possible,
but the common ones are SQL compact and SQL fast
 With tables, query optimization returns to SQL
complexity and is much more reliable
 So, more tricks from the SQL analytics bag become
safe and applicable

Gains from Structure Awareness
 3+x Load Speed
 2x more space efficiency
 SPARQL queries against regular data within 10–20% of
SQL speeds
 Just declare which properties tend to occur together; no
strict schema-first like with SQL
 Later, self configuration

The Cycle of Adventure
 Rebels: SQL not cool, too rigid,
drop ACID, go key-value, map-reduce,
the triple is all there is,
semantic web
 Pioneers: Life on the frontier is
hard, infrastructure missing or bad
 Same everyday problems also in
Utopia
 Recognizing the objective values,
e.g., schema freedom and
identifiers, no AI. Do the job,
forget dogma
 Reconciliation: schema-first and
schema-last converge in structure
awareness

Present FP7 Research
 LDBC — Transparency and Relevance for
Graph DB, RDF performance
 GeoKnow — GeoData is everywhere,
how to carry the planet in your pocket
 LOD2 — Where no triple has gone before
(and come back)
 Open PHACTs — A Data Platform for
Drug Discovery

LDBC - Linked Data Benchmark Council
 Rebels: SQL not cool, too rigid, drop ACID,
go key-value, map-reduce, the triple is all
there is, semantic web
 Pioneers: Life on the frontier is hard,
infrastructure missing or bad
 Same everyday problems also in Utopia
 Recognizing the objective values, e.g.,
schema freedom and identifiers, no AI.
Do the job, forget dogma
 Reconciliation: Some of the rebel thinking
becomes mainstream, e.g., schema-first and
schema-last converge in structure awareness

LDBC, Independent Industry Forum for
Benchmarking
 The TPC for the frontiers of database
 Bootstrapped in the LDBC FP7, continues
as independent industry association
 OpenLink, Ontotext, Neo Technologies,
Sparsity as founding members
 IBM, Oracle Labs, Systap, SPARQL City
already joined
 DB superstars Peter Boncz and Thomas
Neumann as founders and scientific lead

LDBC Benchmarks
Social Network
 Online — Lookups, updates, analysis of
social environment
 Business Intelligence — Spotting trends,
key players, big query
 Graph analytics — Community detection,
Page rank, graph metrics
Semantic Publishing
 Modeled after the BBC linked data portal,
online lookups, drill downs and updates

GeoKnow - The Planet in your Pocket
Ms. Globe and Mr. Cube have a thing
going on:
 Mr. Cube: Desiloization ... integrated
metadata ... Explicit semantics .
 Ms. Globe: I can feel it ... but are you
man enough? ... you need to show me.

Planet Scale Roadmap
Jan 2014:
 Virtuoso SPARQL outperforms PostGIS in map lookups
with planet-wide Open Street Map
 Virtuoso SQL adds 5x more power

Next: Jan 2015
 Parity between SPARQL and SQL via structure
awareness
 Geospatial data clustering
 Graph analytics close to the data — Pregel, Giraph,
etc., in the DB itself
 Adding fine-grained geo dimension to LDBC social
network benchmark

The LOD2 scaling adventures
Experiments at CWI’s Scilens cluster
 Jan 2013: 150 Gtriples (8 x 256GB
RAM)
 Aug 2014: 500 Gtriples (12 x 256GB
RAM)
 Some trillion-triple claims exist, but
do not detail any query workload
BSBM explore and BI workloads
 10x speed gains for BI queries
between 2013 and 2014
Bulk load at 6M triples/s
 All done in triples, structure
awareness will go further still

Open PHACTs
Partners:

Virtuoso Now
Snapshot of RDF Linked Data customers in the Enterprise:
 Data.Gov (U.S. Govt. Open
Linked Data initiative)
 Bank of America
 Booz Allen Hamilton
 Northrop Grumman
 Elsevier
 French National Library
 Samsung
 Globo
 Daimler Benz
 Johnson & Johnson
 Bayer
 St Jude's Medical
 Fuijitsu
 Syngenta
 and many more

Virtuoso Availability
 Most capabilities as open source
 Commercial adds
 Cluster scale-out
 SQL Federation
 Replication (SQL & RDF)
 Advanced RDF security; ABAC & RBAC (ACLs)
 Wide tables
 and more
 Up to the minute tech previews via v7fasttrack on
github, e.g., superfast TPC-H implementation

Virtuoso Future
 Preview of structure-aware RDF store
in fall 2014 via v7fasttrack
Integrated graph analytics framework
 Embed complex graph algorithms, e.g., community
detection, shortest path inside SPARQL/SQL
 Comparison of SQL and SPARQL for big data analytics

Linked Data Now
 Adoption across major industries
 Superior flexibility and time to solution
 Dramatic performance gains in the last 5 years
 Benchmarking will continue to drive progress, to the
benefit of users and vendors alike
 Run circles around most open source SQL in SPARQL:
Virtuoso SPARQL beats MySQL in SSB by 100x
 With structure awareness, SPARQL to match the best in
SQL for data warehousing, OLTP
 Linked Data no longer a long shot but a technology that
makes sense

About OpenLink Software
OpenLink Software is a privately-held company founded in 1992 by its
President & CEO, Kingsley Idehen. The company is an industry acclaimed
technology innovator in the following areas:
 ODBC, JDBC, ADO.NET, and
OLE DB compliant Data Access
Drivers for Oracle, Microsoft SQL
Server, Informix, Ingres, Sybase,
Progress, MySQL, and PostgreSQL
 High-Performance & Scalable Multi-
Model (Relational & Graph)
Database Technology
 Data Integration Middleware (Data
Virtualization Technology across a
wide variety of Protocols & Formats)
 Socially-enhanced Distributed
Collaborative Applications Platforms
(Weblogs, Wikis, Feed Aggregation
and Syndication, Web File Systems,
Discussion Forums, etc.)
 Web Application Server Technology
 Linked Data Deployment &
Management
 Identity Management

Office Locations
USA
OpenLink Software, Inc
10 Burlington Mall Road
Suite 265
Burlington, MA 01803
Tel.: +1 781 273 0900
Fax: +1 781 229 8030
UK
OpenLink Software Ltd.
Airport House
Purley Way
Croydon, Surrey CR0 0XZ
Tel.: +44 (0)20 8681 7701
Fax: +44 (0)20 8681 7702

Additional Information
Web Sites
OpenLink Software
YouID – Digital Identity Card (Certificate) Generator
OpenLink Data Spaces – Semantically enhanced Personal & Enterprise Data Spaces &
Collaboration Platform
OpenLink Virtuoso - Hybrid Data Management, Integration, Application, and Identity Server
Universal Data Access Drivers - High-Performance ODBC, JDBC, ADO.NET, and OLE-DB Drivers
LDAP and NetID-TLS – How to use LDAP scheme URIs with NetID-TLS Authentication
Social Media Data spaces
http://www.openlinksw.com/weblog/oerling/ (Orri Erling weblog)
http://kidehen.blogspot.com (Kingsley Idehen weblog)
http://www.openlinksw.com/blog/~kidehen/ (Kingsley Idehen weblog)
https://twitter.com/OpenLink (Twitter)
Hashtags: #LinkedData #SemanticWeb #BigData #RDF (Anywhere).

Virtuoso, The Prometheus of RDF -- Sematics 2014 Conference Keynote

Recommended

More Related Content

What's hot (20)

Similar to Virtuoso, The Prometheus of RDF -- Sematics 2014 Conference Keynote (20)

More from Kingsley Uyi Idehen (6)

Recently uploaded (20)

Virtuoso, The Prometheus of RDF -- Sematics 2014 Conference Keynote