on unified stream reasoning

Daniele Dell’Aglio

On Unified Stream Reasoning


Advisor: Prof. Emanuele Della Valle

Milano, 15/07/2016. Ph.D. Viva


Problem setting

Real time integration of dynamic data from heterogeneous

sources

– Traffic Prediction

2/33


Problem setting


sources


– Social media analytics

2/33


Problem setting


sources


– Social media analytics

– Personalised services

2/33


Information Flow Processing

In literature there are two different main approaches to

process streams

3/33




process streams

Data Stream Management Systems (DSMSs)

• Roots in DBMS research

• Aggregations, filters and joins

3/33




process streams




Complex Event Processors (CEPs)

• Roots in Distributed Systems (Publish/Subscribe)

• Search of relevant patterns in the stream

• Non-equi-join on the timestamps (after, before, etc.)

3/33




process streams




Complex Event Processors (CEPs)

• Roots in Distributed Systems (Publish/Subscribe)

• Search of relevant patterns in the stream

• Non-equi-join on the timestamps (after, before, etc.)

Current systems implements feature of both of them

• EPL (e.g. Esper, ORACLE CEP)

3/33


Stream Reasoning

Information Flow

Processing

Semantic Technologies

Stream Reasoning

Inference over streams of data

Real-time processing of

highly dynamic data

Integration of heterogeneous data and access through

query answering

4/33


The SR research so far

Euro

pe

Map

fro

m W

ikip

edia

Key

5/33


The SR research so far

IMaRS

STARQL

DynamiTE

TROWL

Euro

pe

Map

fro

m W

ikip

edia

StreamRuleStarCITY

SPARKWAVE

LARS

Key

5/33

StreamQR


The problem (1)

t3 6 91

{:alice :isIn :hall}

{:bob :isIn :hall}

{:alice :isIn :kitchen}

{:bob :isIn :kitchen}

e1 e2 e3 e4S

6/33


The problem (1)

Where are Alice and Bob,

when they are together?

t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

6/33


The problem (1)



Let’s consider a tumbling

window W of size 5

Let’s execute the

experiment 4 times

t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

6/33


The problem (1)

Execution 1° answer 2° answer

1 :hall [6] :kitchen [11]



4 - [7] - [12]



Let’s consider a tumbling

window W of size 5

Let’s execute the

experiment 4 times

t3 6 91


{:bob :isIn :hall}



Which is the correct answer?

e1 e2 e3 e4S

6/33


The problem (2)





4 - [7] - [12]

CSPARQL

t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

7/33


The problem (2)





4 - [7] - [12]



2 No answers


4 No answers

CSPARQL CQELS

Which system behaves in the correct way?

t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

7/33


Research Question 1

In the context of continuous query answering over RDF

streams, how can the behaviour of existing systems be

captured, compared and contrast when deductive reasoning

processes are not involved?

8/33


Research Question 1

In the context of continuous query answering over RDF

streams, how can the behaviour of existing systems be

captured, compared and contrast when deductive reasoning

processes are not involved?

Why do we need it?

• Comparison and contrast

• Study RDF Stream Processing related problems

• Standard RSP query language

8/33


Background data

Streams

Applications

RDF

SPARQL

RSEP-QL

A reference model that formally defines the semantics of

RDF Stream Processing engines

9/33


Background data

Streams

Applications

RSP-QL

RDF

SPARQL

RSEP-QL



9/33


Background data

Streams

Applications

RSP-QL BGP evaluation over background

data

BGP evaluation over streams

Model to express continous queries

RDF

SPARQL

RSEP-QL



9/33


Background data

Streams

RSEP-QL

Applications

RSP-QL BGP evaluation over background

data

BGP evaluation over streams

Event Pattern detection operators

Model to express continous queries

RDF

SPARQL

RSEP-QL



9/33


Q (E, DS, QF)

RSEP-QLFrom SPARQL to RSEP-QL

Query Interface

10/33


Q (E, DS, QF)


Data layer

RDF graphs

DS

Query Interface

10/33


Q (E, DS, QF)


Evaluator

Data layer

RDF graphs

E

DS

Query Interface

10/33


Q (E, DS, QF)


Evaluator

Data layer

Result Formatter

Ans(Q)RDF graphs

E

DS

QF

Query Interface

10/33


Q (E, DS, QF)


Evaluator

Data layer

Result Formatter

Ans(Q)RDF graphs

E

DS

QF

RDF graphsRDF streams

Query Interface

SDS

Q (E, SDS, QF)

10/33


Q (E, DS, QF)


Evaluator

Data layer

Result Formatter

Ans(Q)RDF graphs

E

DS

QFContinuous EvaluatorET


Query Interface

SDS

Q (E, SDS, QF)

Q (E, SDS, ET, QF)

10/33

Ans(Q,t)


Q (E, DS, QF)


Evaluator

Data layer

Result Formatter

Ans(Q)RDF graphs

E

DS

QFContinuous EvaluatorET


Query Interface

SDS

Q (E, SDS, QF)

Q (E, SDS, ET, QF)

SE

10/33

Q (SE, SDS, ET, QF)

Ans(Q,t)


t

S3

S4 S5 S6

S7

S8 S9

S10

S11

SS1

S2

RSEP-QL: DatasetTime-based Sliding Window: 𝕎(ω,β,t0)

11/33


Sequence of timestamped graphs (stream items)

𝕎(2,1,1)

t

S3

S4 S5 S6

S7

S8 S9

S10

S11

SS1

S2


11/33



𝕎(2,1,1)

ω

t

width

S3

S4 S5 S6

S7

S8 S9

S10

S11

SS1

S2

t0


11/33



𝕎(2,1,1)

β

ω

t

widthslide

S3

S4 S5 S6

S7

S8 S9

S10

S11

SS1

S2

t0


11/33


𝕃(2)

t

S3

S4 S5 S6

S7

S8 S9

S10

S11

SS1

S2

t0


RSEP-QL: DatasetLandmark window: 𝕃(t0)

12/33


RSEP-QL: Dataset From SPARQL to RSEP-QL dataset

R={RDF graph}

SPARQL dataset

13/33



R={RDF graph}

SPARQL dataset

S3

S4 S5

S6

S7

S8

S9 S10

S11

S12

SS1

S2

13/33



R={RDF graph}

SPARQL dataset

S3

S4 S5

S6

S7

S8

S9 S10

S11

S12

SS1

S2

𝕎(S)

13/33



t1

G(t1)

T⊆ ℕ R={RDF graph}

SPARQL dataset

GInstantaneous GraphG(t1) RTime-Varying Graph

G: T R

S3

S4 S5

S6

S7

S8

S9 S10

S11

S12

SS1

S2

𝕎(S)

13/33



t1

G(t1)


SPARQL dataset

G

H

Instantaneous GraphG(t1) RTime-Varying Graph

G: T R

S3

S4 S5

S6

S7

S8

S9 S10

S11

S12

SS1

S2

𝕎(S)

13/33



t1

G(t1)


SPARQL dataset

G

H

Instantaneous GraphG(t1) RTime-Varying Graph

G: T R

RESP-QL dataset

S3

S4 S5

S6

S7

S8

S9 S10

S11

S12

SS1

S2

𝕎(S)

13/33


RSEP-QL: Operators

Stream Processing operators (RSP-QL)

• SPARQL operators

• WINDOW to specify that the active element is a window (similar to

GRAPH)

• RStream, IStream, DStream to create the output stream

14/33


RSEP-QL: Operators

Stream Processing operators (RSP-QL)

• SPARQL operators

• WINDOW to specify that the active element is a window (similar to

GRAPH)

• RStream, IStream, DStream to create the output stream

Event Processing operators (RSEP-QL)

• EVENT w P (Basic Event Pattern)

• E1 SEQ E2

• FIRST E1, LAST E1

• MATCH to describe an event pattern

• ...

14/33


RSEP-QL: Evaluation SemanticsStream Processing Evaluation Semantics

The SPARQL evaluation function is defined as

⟦𝑃⟧𝐷𝑆(𝐺)

15/33





The RSEP-QL evaluation function extends the SPARQL one

by introducing the evaluation time instant

⟪𝑃⟫𝑆𝐷𝑆(𝐴)𝑡

15/33





The RSEP-QL evaluation function extends the SPARQL one

by introducing the evaluation time instant

⟪𝑃⟫𝑆𝐷𝑆(𝐴)𝑡

SPARQL operators are straight extended to the new

evaluation function

Example: JOIN

⟦𝐽𝑂𝐼𝑁(𝑃1, 𝑃2)⟧𝐷𝑆(𝐺) = ⟦𝑃1⟧𝐷𝑆 𝐺 ⨝⟦𝑃2⟧𝐷𝑆 𝐺⟪𝐽𝑂𝐼𝑁(𝑃1, 𝑃2)⟫𝑆𝐷𝑆 𝐴

𝑡 = ⟪𝑃1⟫𝑆𝐷𝑆 𝐴𝑡 ⨝⟪𝑃2⟫𝑆𝐷𝑆 𝐴

𝑡

15/33



The main difference is on the BGP evaluation:

⟪𝐵𝐺𝑃⟫𝑆𝐷𝑆(𝐴)𝑡

=⟦𝐵𝐺𝑃⟧𝑆𝐷𝑆(𝐴,𝑡)

16/33






SDS(A,t) is:

SDS(G,t)= SDS(G(t)) if A is a time-varying graph G

16/33






SDS(A,t) is:

SDS(G,t)= SDS(G(t)) if A is a time-varying graph G

SDS(𝕎(S),t)=SDS(m(𝕎(S,t))) if A is from a sliding window 𝕎

SDS(𝕃(S),t)=SDS(m(𝕃(S,t))) if A is from a landmark window 𝕃

where m denotes a merge function

m(𝕎(S,t))= 𝑑𝑖,𝑡𝑖 ∈𝕎(S,t)

𝑑𝑖

• takes as input a window content i.e. a sequence of timestamped

RDF graphs

• produces an RDF graph

16/33


RSEP-QL: Evaluation SemanticsEvent Processing Evaluation semantics

A new evaluation function ⦅⋅⦆ 𝑜,𝑐𝑡

• t is the evaluation time instant (as in ⟪⋅⟫𝑆𝐷𝑆(𝐴)𝑡 ),

• 𝑜, 𝑐 is an additional window to identify the portion of the data

on which the event may happen

17/33







Event pattern evaluation produces event mappings 𝜇, 𝑡1, 𝑡2• 𝜇 is a solution mapping

• 𝑡1 and 𝑡2 denote the time inverval justifying 𝜇

17/33







Event pattern evaluation produces event mappings 𝜇, 𝑡1, 𝑡2• 𝜇 is a solution mapping

• 𝑡1 and 𝑡2 denote the time inverval justifying 𝜇

Selection and consumption policies to determine the data

involved in the evaluation

17/33


RSEP-QL: Evaluation SemanticsEvaluation semantics - Example

⦅FIRST EVENT 𝑤1 𝑃1 SEQ EVERY EVENT 𝑤2 𝑃2⦆ 9,16𝑡

S2

S3 S4

S1 S1

S6

S7

S8

S9 S10

S11

S12

S2

FIRST EVENT 𝑤1 𝑃1

SEQ

EVERY EVENT 𝑤2 𝑃2

t10 12 14 1611 13 15

18/33




S2

S3 S4

S1 S1

S6

S7

S8

S9 S10

S11

S12

S2

S1 S10


SEQ


t10 12 14 1611 13 15

11 13

18/33




S2

S3 S4

S1 S1

S6

S7

S8

S9 S10

S11

S12

S2

S1 S10

S1 S12


SEQ


t10 12 14 1611 13 15

11 13

11 15

18/33


RSEP-QL: Evaluation SemanticsMATCH graph pattern

Event patterns are enclosed in MATCH graph patterns

• Event mappings exist only in the context of event patterns

• The evaluation of a MATCH graph pattern produces a bag of solution

mappings

⟪𝑀𝐴𝑇𝐶𝐻 𝐸⟫𝑆𝐷𝑆 𝐴𝑡 = {𝜇| 𝜇, 𝑡1, 𝑡2 ∈ ⦅𝐸⦆ 0,𝑡

𝑡 }

It is possible to combine the MATCH graph pattern with

other SPARQL graph patterns

19/33


RSEP-QL: Evaluation SemanticsContinuous evaluation

For each evaluation time t ∈ ET: ⟪𝑆𝐸⟫𝑆𝐷𝑆(𝐴)𝑡

• The continuous evaluation is a sequence of instantaneous

evaluations

20/33





evaluations

It is not always possible to compute ET a priori

20/33





evaluations


• Can be infinite

• Can be data dependent

20/33





evaluations


• Can be infinite

• Can be data dependent

ET is expressed through a Report Policy

• A set of conditions to one or more window operators in SDS

• Initially defined in SECRET for Stream Processing engines

20/33



Report Policy examples:

• P Periodic: the window reports only at regular intervals

• WC Window Close: the window reports if the active window

closes

• CC Content Change: the window reports if the content changes.

21/33


Research Question 2

Is it possible to add deductive reasoning features to RSEP-QL?

22/33


Research Question 2


Applications

RDF

SPARQL

Ontology

22/33


Research Question 2


Applications

RDF

SPARQL

Ontology

SPARQL Entailment

Regimes

22/33


Research Question 2


Streams OntologyBackground

data

RSEP-QL

Applications

RSP-QL

RDF

SPARQL

Ontology

SPARQL Entailment

Regimes

22/33


Research Question 2


Streams OntologyBackground

data

RSEP-QL Entailment

Regimes

RSEP-QL

Applications

RSP-QL

RDF

SPARQL

Ontology

SPARQL Entailment

Regimes

22/33


RSEP-QL: Reasoning

Stream Processing

At which point of the continuous query answering process

should the inference process be taken into account?

Event ProcessingWindow merge

Window operator

Stream S34 5

67

89

S 12

67

89

6-9

t

RDF Graph

RDF Stream

Window

23/33


RSEP-QL: Reasoning

Stream Processing


should the inference process be taken into account?Direct application

on SPARQLentailment regimes


Window operator

Stream S

Graph-level entailment

34 5

67

89

S 12

67

89

6-9

t

RDF Graph

RDF Stream

Window

23/33


RSEP-QL: Reasoning

Stream Processing



After applying the window operator

Direct application on SPARQL

entailment regimes


Window operator

Stream S


Window-level entailment

34 5

67

89

S 12

67

89

6-9

t

RDF Graph

RDF Stream

Window

23/33


RSEP-QL: Reasoning

Stream Processing



A larger, landmark window to include

relevant information from the past

After applying the window operator

Direct application on SPARQL

entailment regimes


Window operator

Stream S



Stream-level entailment

34 5

67

89

S 12

67

89

6-9

t

67

89

34 5

RDF Graph

RDF Stream

Window

23/33


RSEP-QL: ReasoningWindow-level and stream-level entailment regimes

How can DL ontology be extended to capture the window

content?

24/33




content?

• Ontology stream – a sequence of time-stamped ABoxes 𝑆𝑜𝑐 𝑖

w.r.t a static TBox

24/33




content?

• Ontology stream – a sequence of time-stamped ABoxes 𝑆𝑜𝑐 𝑖

w.r.t a static TBox

• Introduction of so-far closure, as the closure of the i-th

snapshot w.r.t. the Tbox and the previus snapshots

• The so-far closure of 𝑆𝑜𝑐 𝑖 is done w.r.t. the TBox and the

snapshots 𝑆𝑜𝑐 𝑜 … 𝑆𝑜

𝑐 𝑖 − 1

• The so-far closure of the ontology stream is obtained by so-far

closing all the ABox snapshots

24/33


Event Processing


RSEP-QL: ReasoningResults

1. Theorem: in the context of

Stream Processing operations,

graph- and window-level

entailments have the same

behaviour as far as no

inconsistency is entailed

25/33

Stream Processing

Window merge

Window operator

Stream S

Graph-level ent.



Event Processing


RSEP-QL: ReasoningResults

1. Theorem: in the context of

Stream Processing operations,

graph- and window-level

entailments have the same

behaviour as far as no

inconsistency is entailed

2. Theorem: The stream-level

entailment brings to a larger

set of answers w.r.t. graph-

and window- entailment ones.

25/33

Stream Processing


Window operator

Stream S

Graph-level ent.








4 - [7] - [12]

RSEP-QL in actionCorrectness assessment

t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

26/33






4 - [7] - [12]

t0=0


t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

26/33






4 - [7] - [12]

t0=0

Window 1° answer 2° answer

t0=0 :hall [5] :kitchen [10]


t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

26/33






4 - [7] - [12]

t0=0



t0=1


t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

26/33






4 - [7] - [12]

t0=0




t0=1


t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

26/33






4 - [7] - [12]

t0=0




t0=1

t0=2


t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

26/33






4 - [7] - [12]

t0=0




t0=2 - [7] - [12]

t0=1

t0=2


t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S

26/33






4 - [7] - [12]



2 No answers


4 No answers

CSPARQL CQELS


27/33

t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S






4 - [7] - [12]



2 No answers


4 No answers

CSPARQL CQELS

Window-close vs content-change

Empty relation notification (yes|no)


27/33

t3 6 91


{:bob :isIn :hall}



e1 e2 e3 e4S


CSR Bench

CSR Bench is an extension of the SRbench benchmark

that focuses on correctness

A test suite

• LinkedSensorData as dataset

• 8 parametric queries to tests the RSP engines in different

conditions

An oracle (an automatic correctness validator)

• Based on RSP-QL

28/33


CSR BenchExperimental Results

All the three systems that we

considered in our experiments showed

wrong behaviours

The defects we identified are related to:

•The window operator

• Initialization

• Slide parameter

• Window contents

•Timestamps of the stream items

• Internal timestamp management

CQ

ELS

C-SPA

RQ

L

SPAR

QL

stream

Q1

Q2

Q3

Q4

Q5

Q6

Q7

29/33


What’s next?

An RSEP-QL query language

• W3C RSP CG ongoing activities

30/33


What’s next?



Implementations

30/33


What’s next?



Implementations

• Yet another RSP engine

30/33


What’s next?



Implementations


• Stream Reasoner composition

• RSEP-QL as model to capture the behaviour of SRs

• RSEP-QL queries to define tasks to be executed over one or

more SR engines

30/33


What’s next?



Implementations


• Stream Reasoner composition

• RSEP-QL as model to capture the behaviour of SRs

• RSEP-QL queries to define tasks to be executed over one or

more SR engines

Stream-level entailment scalable techniques

• Initial work on stream without inference

• Permanent storage of portions of data (raw or inferred)

30/33


What’s next?

Reasoning over streams and continuous query answering

over streams are still two different worlds

• Key to show why RDF Stream Processing is useful w.r.t.

DSMS and CEP

• Few initial attempts to put them together

31/33


What’s next?

Reasoning over streams and continuous query answering

over streams are still two different worlds

• Key to show why RDF Stream Processing is useful w.r.t.

DSMS and CEP

• Few initial attempts to put them together

Streams in the Web are getting popular – applications want

more and more sophisticated features

• Different timestamps, out-of-orders

• Noise (inductive reasoning)

31/33


Conclusions

RSEP-QL captures the evaluation semantics of existing

RSP engines

32/33


Conclusions


RSP engines

RSEP-QL can determine which are the correct answers of

an RSP engine, given the input data and query

• At the basis of CSR Bench

32/33


Conclusions


RSP engines

RSEP-QL can determine which are the correct answers of

an RSP engine, given the input data and query

• At the basis of CSR Bench

The dynamics introduced in the continuous query evaluation

process have not been totally understood

• Not fully captured by existing models

• RSEP-QL captures those dynamics

32/33


Thank you! Questions?

On Unified Stream Reasoning


Advisor: Prof. Emanuele Della Valle

33/33

on unified stream reasoning

Science