Presentation of the paper "Stream Processing: The Matrix Revolutions" at the 12th SSWS Workshop colocated with ISWC 2018 in Monterey, CA, USA.

1. Department of Informatics R. Pernischová, F. Ruosch, D. Dell’Aglio, A. Bernstein
9th Oktober 2018, SSWS 2018, Monterey, USA
STREAM PROCESSING:
THE MATRIX REVOLUTIONS

3. MOTIVATING EXAMPLE
• User behavior and context
• Information about the advertised product
• Content of the advertisement

4. CHALLENGES
• Real-time processing
• Data heterogeneity
• Data arriving at different speed

5. Stream processing
engines
(Flink et al.)
Linear and relational
algebra in one language
Graph stream
processing
COMBINED IN
ONE SYSTEM
STATE OF THE ART

6. How can we build a distributed stream processing
engine that processes streams containing matrices,
tables, and graphs?
RESEARCH QUESTION

7. How to represent the
different streams in a
common formalism/model?
How to encode
the query?
How to execute
the query?3
1
THE MODEL
?
2

8. STREAM INTEGRATION - THE MODEL
RDF
– Relational data and graphs can be easily converted
– Matrices are added as nodes in a custom data format

9. STREAM INTEGRATION
MATRICES IN RDF
rlg:Matrix
rlg:matrix
rdfs:Class
rlg:colums
rlg:rows
rlg:data
xsd:int
rdfs:Datatype
rdf:type
rdf:type
rdfs:domain rdfs:range
rdfs:range

10. STREAM INTEGRATION
MATRICES IN RDF
:m1 rdf:type rlg:Matrix ;
rlg:data "[3 4 8][8 7 2][1 8 2]"^^rlg:matrix ;
rlg:columns 3 ;
rlg:rows 3 .

11. STREAM INTEGRATION
RDF STREAMS
:streamItem1 {
:m1 rlg:data "[3 4 8][8 7 2][1 8 2]"^^rlg:matrix ;
rlg:columns 3 ; rlg:rows 3 .
} { :streamItem1 prov:generatedAt 5 .}
:streamItem2 {
:m2 rlg:data "[1 0 2][9 6 2][6 4 0]"^^rlg:matrix ;
rlg:columns 3 ; rlg:rows 3 .
} { :streamItem1 prov:generatedAt 8 .}
5
8

12. REGISTER STREAM :outStr AS
CONSTRUCT RSTREAM {
?m :hasInverse [ rlg:data ?inverse ] .
}
FROM NAMED WINDOW :win ON :inStr [RANGE 1 STEP 1]
WHERE {
...
}
?QUERY MODELING
QUERYING RDF STREAMS

13. QUERY MODELING
MATRIX PROCESSING
[...]
WHERE {
?m rdf:type rlg:Matrix ;
rlg:data ?data .
BIND (afn:inverse(?data) AS ?inverse)
}
?

14. QUERY EXECUTION
SPARQL
SELECT ?add
WHERE {
?m1 a rlg:Matrix ;
rlg:data ?d1 .
?m2 a rlg:Matrix ;
rlg:data ?d2 .
BIND (add(?d2,?d1)
as ?add)
}
Operator
Tree
Topology
Sink
Sources

15. IMPLEMENTATION
?

16. Apache Flink
Streaming
operators
QUERY EVALUATION IMPLEMENTATION
Apache Jena
Basic Graph
Pattern evaluationJAMA
Matrix data
structure and
functions
?

17. LIMITATIONS
Prototype
• Stream Integration not implemented
• The query parser does not manage the stream operators (passed as extra
arguments)
• Not all SPARQL operators are implemented in Flink
Model
• Matrices are encoded with Strings à Limited in size
Evaluation
• Lack of baselines and benchmarks
?

18. CONCLUSIONS
It is possible to build a system that processes an RDF stream enriched with
matrices.
Data model is compliant with RDF.
Query model is almost compliant with SPARQL.
• SPARQL functions integrate operations over matrices.
• Continuous extensions are needed.
Implementation is done over a distributed stream processing engine.
• Processing can be distributed.
• Data parallelism is future work.

19. THANK YOU FOR YOUR ATTENTION!
QUESTIONS?
Romana Pernischova
pernischova@ifi.uzh.ch

21. REFERENCES
Images:
- https://www.mirror.co.uk/science/sitting-watching-tv-day-one-11045526
- https://signalprocessingsociety.org/sites/default/files/Anna_Suarez_Article_2.jpg
Icons:
- http://www.trafficsign.us
Paper:
R. Pernischova, F. Ruosch, D. Dell’ Aglio, A. Bernstein (2018). Stream Processing: The Matrix
Revolutions. 12th International Workshop on Scalable Semantic Web Knowledge Base Systems, CEUR
Workshop Proceedings, 15-27.