Trio: A System for Data, Uncertainty, and Lineage

Trio: A System for Data, Uncertainty, and Lineage

Martin Theobald

Jennifer Widom

Stanford University

2

Uncertainty in Databases

• Not a new idea — proposed 20+ years ago• “Classic” probabilistic database systems

• Uncertainty about attribute values

• Modeling of missing values/incompleteness

[Cavallo ‘87, Abiteboul ‘91, Garcia-Molina ‘92, Fuhr/Rölleke ‘97]

• Most initial (18+ years) work largely theoretical; not much systems-building until recently:URank – U Waterloo, Orion – Purdue U, MayBMS -

Saarland U

• But applications weren’t ready anyway

Are they now?

3

The “Trio” in Trio

1. Data Student #123 is majoring in Econ: (123,Econ)

Major

2. Uncertainty Student #123 is majoring in Econ or CS:

(123, Econ ∥ CS) Major With confidence 60% student #456 is a CS major:

(456, CS 0.6) Major

3. Lineage 456 HardWorker derived from:

(456, CS) Major and“CS is hard” some web page

4

Original Motivation for Trio

New Application Domains

• Many involve data that is uncertain (probabilistic, imprecise, fuzzy, incomplete...)

• Many of the same ones need to track the lineage (provenance) of their data

Coincidence or Fate?Coincidence or Fate?

5

Original Motivation for Trio

New Application Domains

• Many involve data that is uncertain (probabilistic, imprecise, fuzzy, incomplete...)

• Many of the same ones need to track the lineage (provenance) of their data

Trio combines uncertainty & lineage for a closed & complete representation modelNeither uncertainty nor lineage is supported in current commercial database systems

Neither uncertainty nor lineage is supported in current commercial database systems

6

Sample Applications

Information extraction• Find & label entities in unstructured text• Often probabilistic

Information integration• Combine data from multiple sources• Inconsistencies, probabilistic schema mappings

“Human input” data / Scientific experiments• Inexact/incomplete data• Many levels of “derived data products”

7

Sample Applications

Sensor data management• Approximate readings• Missing readings• Levels of data aggregation

Deduplication (“data cleaning”)• Object linkage, entity resolution• Often heuristic/probabilistic

But *NOT* approximate query processing• Fast but inexact answers

8

Our Claim

The connection between uncertainty and lineage goes deeper than just a shared need by several applications

Coincidence or Fate?Coincidence or Fate?

9

Substantiation of Claim

[technical] Lineage:1. Enables simple and consistent

representation of uncertain data

2. Correlates uncertainty in query results with uncertainty in the input data

3. Can make computations over uncertain data more efficient

[fluffy] “Applications may use lineage to reduce or resolve uncertainty”

10

Our Goal

Develop a new kind of database management system (DBMS) in which:

1. Data2. Uncertainty3. Lineage

are all first-class interrelated concepts

With all the “usual” DBMS features

11

The “Usual” DBMS Features

1. Efficient,

2. Convenient,

3. Safe,

4. Multi-User storage of and access to

5. Massive amounts of

6. Persistent data

12

Standard Relational DBMSs

Persistent; Convenient• All data stored in simple tables (relations)

• Queries and updates via simple but powerful declarative language (SQL)

Multi-User; Safe• Transactions, Recovery

Massive; Efficient• Storage and indexing structures

• Query optimization

13

Trio: What Changes

Persistent; Convenient• All data stored in “uncertain” relations

• Queries and updates via simple but powerful declarative language

Multi-User; Safe• Transactions, Recovery

Massive; Efficient• Storage and indexing structures


14

Trio: What Changes

Persistent; Convenient• All data stored in “uncertain” relations

• Queries and updates via simple but powerful declarative language

Multi-User; Safe – standard DBMS underneath• Transactions, Recovery

Massive; Efficient – standard DBMS underneath• Storage and indexing structures


15

Another “Trio” in Trio

1. Data Model Simplest extension to relational model that’s

sufficiently expressive

2. Query Language Simple extension to SQL with well-defined

semantics and intuitive behavior

3. System A complete open-source DBMS that people

want to use

16

Another “Trio” in Trio

1. Data Model Uncertainty-Lineage Databases (ULDBs)

2. Query Language TriQL

3. System Trio-One — built on top of standard DBMS

17

Remainder of Talk

1. Data Model Uncertainty-Lineage Databases (ULDBs)

2. Query Language TriQL

3. SystemTrio-One

4. Real-World Applications PharmGKB.org

SERF

18

Running Example: Crime-Solving

Saw (witness, color, car) // may be uncertain

Drives (person, color, car) // may be uncertain

Suspects (person) = πperson(Saw ⋈ Drives)

19

In Standard Relational DBMS

Drives (person, color, car)

Jimmy red Toyota

Frank red Mazda

Billy blue Honda

Frank green

Mazda

Saw (witness, color, car)

Amy red Mazda

Betty blue Honda

Carol green Toyota

Suspects

Frank

Billy

Create Table Suspects asSelect personFrom Saw, DrivesWhere Saw.color = Drives.color And Saw.car = Drives.car

Create Table Suspects asSelect personFrom Saw, DrivesWhere Saw.color = Drives.color And Saw.car = Drives.car

20

Data Model: Uncertainty

An uncertain database enodes a set of possible instances based on different kinds of uncertainty:

• Jimmy drives either a Toyota or a Mazda, but not both.

• Amy saw either a Honda or a Toyota, or nothing?

• Hank is a suspect with confidence 0.7

• Betty saw either an Acura with confidence 0.5 or a Toyota with confidence 0.3

or nothing with confidence 0.2 ….

21

Our Model for Uncertainty

1. Alternatives

2. ‘?’ (Maybe) Annotations

3. Confidences

22


1. Alternatives: uncertainty about value


3. ConfidencesSaw (witness, color, car)

Amy red, Honda ∥ red, Toyota ∥ orange, Mazda

Three possible

instances

23

Six possibleinstances


1. Alternatives

2. ‘?’ (Maybe): uncertainty about presence

3. Confidences

?



Betty blue, Acura

24


1. Alternatives

2. ‘?’ (Maybe): uncertainty about presence

3. ConfidencesSaw (witness, color, car)


Betty blue, Acura?

Betty blue, Acura ∥ NULL, NULL

absent unknownabsent unknown

25


1. Alternatives


3. Confidences: weighted uncertainty

Six possible instances, each with a probability

?


Amy red, Honda 0.5 ∥ red, Toyota 0.3 ∥ orange, Mazda 0.2

Betty blue, Acura 0.6

26

Our Model is Not Closed

Saw (witness, car)

Cathy

Honda ∥ Mazda

Drives (person, car)

Jimmy, Toyota ∥ Jimmy, Mazda

Billy, Honda ∥ Frank, Honda

Hank, Honda

Suspects

Jimmy

Billy ∥ Frank

Hank

Suspects = πperson(Saw ⋈ Drives)

???

Does not correctlycapture possibleinstances in theresult (not 12)!

CANNOT

27

to the Rescue

Lineage (provenance): “where data came from”• Internal lineage – from another Trio relation

• External lineage – from an external source

In Trio: A Boolean function λ from data elements to other data elements (or external sources)

Lineage

28

Example with Lineage

ID Saw (witness, car)

11

Cathy

Honda ∥ Mazda

ID Drives (person, car)

21


22


23

Hank, Honda

ID Suspects

31

Jimmy

32

Billy ∥ Frank

33

Hank


???

λ(31) = (11,2)˄(21,2)

λ(32,1) = (11,1)˄(22,1)

λ(33) = (11,1) ˄ 23

; λ(32,2) = (11,1)˄(22,2)

29

Example with Lineage

ID Saw (witness, car)

11

Cathy

Honda ∥ Mazda

ID Drives (person, car)

21


22


23

Hank, Honda

ID Suspects

31

Jimmy

32

Billy ∥ Frank

33

Hank


???

λ(31) = (11,2)˄(21,2)λ(32,1) = (11,1)˄(22,1); λ(32,2) = (11,1)˄(22,2)

λ(33) = (11,1) ˄ 23

Correctly captures possible instances inthe result (7)

30

Trio Data Model

1. Alternatives


3. Confidences

4. Lineage

ULDBs are closed and complete

Uncertainty-Lineage Databases (ULDBs)Uncertainty-Lineage Databases (ULDBs)

[Databases with Uncertainty and Lineage, VLDB J. Special Issue 2/08]

31

ULDBs: Lineage

Conjunctive lineage sufficient for most operations (,,⋈)

• Disjunctive lineage for duplicate-elimination ()

• Negative lineage for set difference ()

General case after several queries: • Boolean formula

• Can represent any finite (sub)set of possible instances

• Can capture arbitrary Boolean constraints among tuples

• Can represent any finite (sub)set of possible instances

• Can capture arbitrary Boolean constraints among tuples

32

ULDBs: Minimality

A ULDB relation R represents a set of possible

instances

• Does every tuple in R appear in some possible instance? (no extraneous tuples)

• Does every maybe-tuple in R not appear in some possible instance? (no extraneous ‘?’s)

• Also

Data-minimalityData-minimality

Lineage-minimalityLineage-minimality

33

Data Minimality Examples

Extraneous ‘?’

. . .

10

Billy, Honda ∥ Frank, Mazda

. . .

. . .

40

Billy ∥ Frank

. . .

?λ(40,1)=(10,1)λ(40,2)=(10,2)

extraneous

. . .

20 Billy, Honda

. . .

. . .

30 Frank, Mazda

. . .

? ?

34

Data Minimality Examples

Extraneous tuple

extraneous

. . . . . .

10

Diane Acura ∥ Mazda

. . . . . .

. . .

40

Diane

. . .

? λ(40,1)=(10,1)˄(10,2)

. . . . . .

20 Diane Acura

. . . . . .

. . . . . .

30 Diane Mazda

. . . . . .

? ?

• Confidence computation over arbitrary Boolean expressions:

Efficient approximation algorithms (E.g. Monte Carlo simulations)

35

ULDBs: Complexity Questions• Does a given tuple t appear in some (all) possible

instance(s) of an uncertain relation R ?

• Is a given table T one of (all of) the possible instances of R ?

Polynomial algorithms based on data-minimizationPolynomial algorithms based on data-minimization

NP-hardNP-hard

#P-complete#P-complete

36

Querying ULDBs

Simple extension to SQL

Formal semantics, intuitive meaning

Query uncertainty, confidences, and lineage

TriQLTriQL

37

Formal Semantics

Relational (SQL) query Q on ULDB D

DD

D1, D2, …, DnD1, D2, …, Dn

possibleinstances

Q on eachinstance

representationof instances

Q(D1), Q(D2), …, Q(Dn)Q(D1), Q(D2), …, Q(Dn)

implementation of Q

operational semanticsD + ResultD + Result

38

Operational Semantics: Example

SELECT personFROM Saw, DrivesWHERE Saw.car = Drives.car

Saw (witness, car)

Cathy Honda ∥ Mazda


Jim ∥ Bill Mazda

Hank Honda

39



(Cathy,Honda,Jim,Mazda)∥(Cathy,Honda,Bill,Mazda)∥(Cathy,Mazda,Jim,Mazda)∥(Cathy,Mazda,Bill,Mazda)

Saw (witness, car)



Jim ∥ Bill Mazda

Hank Honda

40




Saw (witness, car)



Jim ∥ Bill Mazda

Hank Honda

41




Saw (witness, car)



Jim ∥ Bill Mazda

Hank Honda

?

42




(Cathy,Honda,Hank,Honda) ∥ (Cathy,Mazda,Hank,Honda)

Saw (witness, car)



Jim ∥ Bill Mazda

Hank Honda

?

43





Saw (witness, car)



Jim ∥ Bill Mazda

Hank Honda

?

44





Saw (witness, car)



Jim ∥ Bill Mazda

Hank Honda

??

45


CREATE TABLE Suspects ASSELECT Drives.personFROM Saw, DrivesWHERE Saw.car = Drives.car

Suspects

Jim ∥ Bill

Hank

??λ( ) = ...λ( ) = ...

Saw (witness, car)



Jim ∥ Bill Mazda

Hank Honda

• Efficiently implementable on top of standard SQL

• Data computation overhead: O(n log(n))

• Efficiently implementable on top of standard SQL

• Data computation overhead: O(n log(n))

46

Confidences

Confidences supplied with base data

Trio computes confidences on query results• Default probabilistic interpretation (#P-complete)• Plug in different arithmetics, e.g., Min (linear)

Saw (witness, car)

Cathy

Honda 0.6 ∥ Mazda 0.4


Jim 0.3 ∥ Bill 0.6 Mazda

Hank 1.0 Honda

Suspects

Jim 0.12 ∥ Bill 0.24

Hank 0.6

0.3 0.4

0.6 ProbabilisticProbabilistic Min

47

TriQL: Querying Confidences Built-in function: Conf()

SELECT person FROM Saw, Drives WHERE Saw.car = Drives.car AND Conf(Saw) > 0.5 AND Conf(Drives) > 0.8

Similarly: Conf(*)

SELECT person FROM Saw, Drives WHERE Saw.car = Drives.car AND Conf(*) > 0.5

Naïve data-minimization:

WHERE Conf(*) > 0

Naïve data-minimization:

WHERE Conf(*) > 0

48

TriQL: Querying Lineage

Built-in join predicate: Lineage()

SELECT Suspects.person FROM Suspects, Saw WHERE Saw.witness = ‘Amy’ AND Lineage(Suspects,Saw)

X ==> Y shorthand for Lineage(X,Y)

49

More TriQL Constructs

• “Horizontal subqueries”• Refer to tuple alternatives as a relation

• Aggregations: • (sum, count, avg, min, max) (low, high,

expected)

• Fetch modifiers:• Merged (eliminate horizontal duplicates)• Flatten, GroupAlts, NoLineage, NoConf, NoMaybe

• Set operators: Union [All], Intersect, ExceptConjunctive, Disjunctive & Negative Lineage

50

Horizontal Subquery Example

PrimeSuspect (crime#, suspect, accuser)

1 Jimmy, Amy ∥ Billy, Betty ∥ Hank, Cathy

2 Frank, Cathy ∥ Freddy, Betty

Credibility (person,scor

e)

Amy 10

Betty 15

Cathy 5

Suspects

Jimmy 0.33 ∥ Billy 0.5 ∥ Hank 0.166

Frank 0.25 ∥ Freddy 0.75

List suspects with conf values based on accuser credibility

51






e)

Amy 10

Betty 15

Cathy 5

Suspects



SELECT suspect, score/[sum(score)] as confFROM PrimeSuspect, CredibilityWHERE accuser = person

SELECT suspect, score/[sum(score)] as confFROM PrimeSuspect, CredibilityWHERE accuser = person

SELECT suspect, FROM PrimeSuspect, CredibilityWHERE accuser = person

SELECT suspect, FROM PrimeSuspect, CredibilityWHERE accuser = person

52






e)

Amy 10

Betty 15

Cathy 5

Suspects



scaled as conf

Remove extraneous ‘?’ if:

[Sum(Conf(*))] = 1

Remove extraneous ‘?’ if:

[Sum(Conf(*))] = 1

53

Trio-Specific Additional Features

• Lineage tracing

• On-demand confidence computation

• Coexistence checks

• Extraneous data removal

Interrelated algorithms based on lineage

54

The Trio System

• Version 1 (“Trio-One”)• Layered on top of a standard DBMS

(PostgreSQL)

• Surprisingly easy and complete, reasonably efficient

• Open source as of summer 2007• Supports Linux, Windows Server/XP/Vista

& Mac-OSX

• Constantly updated & maintained

55

Trio-One Overview

Standard relational DBMS

Trio API and TriQL rewriter(Python)

Trio API and TriQL rewriter(Python)

TrioPlus(Command-line

Client)

TrioPlus(Command-line

Client)

TrioMetadat

a

TrioExplorer(GUI client)

TrioExplorer(GUI client)

Trio Stored

Procedures

EncodedData

TablesLineageTables

Standard SQL• Partition and “verticalize”• Shared IDs for alternatives• Columns for confidence,“?”• One per result table• Uses unique IDs• Encodes formulas

• Table types• Schema-level lineage structure• Conf()• Lineage()

• DDL commands• TriQL queries• Schema browsing• Table browsing• Explore lineage• On-demand confidence computation

Postgres SPI functions

Applications

Pharmaceutical / Medical Databases• PharmGKB.org

• Relationships between Drugs, Diseases, Genes

• Evidence from Literature References, Clinical Outcomes, etc.

58

Applications – PharmGKB.org

59

GeneC

DiseaseB

DrugA

EffectiveA,BEffectiveA,B,C

• Relational Schema • Drugs(DrugID, Name, AltName)

• Diseases(DiseaseID, Name, AltName)

• Genes(GeneID, Name, Symbol, AltSymbol)

• Relationships(RelID, RelatedTo, Evidence)

• Independent Base Data

P[EffectiveA,B,C]

= P[A] P[B] P[C]

Applications – PharmGKB.org

60

GeneC

DiseaseB

DrugA

EffectiveA,BEffectiveA,B,C

• Relational Schema • Drugs(DrugID, Name, AltName)

• Diseases(DiseaseID, Name, AltName)

• Genes(GeneID, Name, Symbol, AltSymbol)

• Relationships(RelID, RelatedTo, Evidence)

• Non-Independent Base Data

P[EffectiveA,B,C]

= P[E|A,B,C]

≠ P[A] P[B] P[C]

ULDBs Vs. Bayesian Nets Factorized representation w/conditional probability tables (CPTs)

E.g.: C = Heart DiseaseB = High Blood PressureA = Warfarin

Applications – SERF

Stanford Entity Resolution Framework

61

Swoosh: A Generic Approach to Entity ResolutionBenjelloun, Garcia-Molina, Whang et al.

<id>r2<name>John Doe<phone>357-2635<email>[email protected]

<id>r1<name>J. Doe<phone>351-2535

<id>r3<name>John D.<phone>351-2535<email>[email protected]

<id>r4<name>{John Doe||John D.}<phone>{357-2635||351-2535}<email>[email protected]

<id>r5<name>{John Doe||John D.||J. Doe}<phone>{357-2635||351-2535}<email>[email protected]

Iteration 2Iteration 2


Stanford Entity Resolution Framework

62

Swoosh: A Generic Approach to Entity ResolutionBenjelloun, Garcia-Molina, Whang et al.

<id>r2<name>John Doe<phone>357-2635<email>[email protected]

<id>r1<name>J. Doe<phone>351-2535

<id>r3<name>John D.<phone>351-2535<email>[email protected]

<id>r4<name>{John Doe : 0.4 ||John D. : 0.3}<phone>{357-2635 : 0.5 ||351-2535 : 0.5}<email>[email protected]

<id>r5<name>{John Doe : 0.4 ||John D. : 0.3||J. Doe : 0.3}<phone>{357-2635 : 0.3 ||351-2535 : 0.6}<email>[email protected]



Applications – SERF

63

Future Topics (sample)

More forms of uncertainty• Continuous uncertainty (intervals, Gaussians)• Correlated uncertainty• Incomplete relations

More efficient confidence-based queries• Top-k by confidence• Confidence bounds & approximation

More forms of lineage• Update propagation w/lineage (versioning)• External lineage

64

Future Topics (sample)

Theory• AI models/Bayesian Nets & inferencing

• Design theory

System• Full TriQL query language• Version 2: Go native beyond Postgres SPI?

– Storage and indexing structures– Statistics– Query optimization

Search “stanford trio”

Trio contributors, past and present Parag Agrawal, Omar Benjelloun, Julien Chaumond, Ashok Chandra, Anish Das Sarma, Alon Halevy, Chris

Hayworth, Ander de Keijzer, Raghotham Murthy, Michi Mutsuzaki, Shubha Nabar, Tomoe Sugihara,

Martin Theobald, Jeff Ullman, Jennifer Widom

Trio: A System for Data, Uncertainty, and Lineage

Documents

Transcript of Trio: A System for Data, Uncertainty, and Lineage