Adaptive Query Processing
for Data Aggregation:
Mining, Using and Maintaining Source Statistics
M.S Thesis Defense
by Jianchun Fan
Committee Members:
Dr. Subbarao Kambhampati (chair)
Dr. Huan Liu
Dr. Yi Chen
April 13, 2006
Introduction
• Data Aggregation: Vertical Integration
R (A1, A2, A3, A4, A5, A6)
Mediator
A1
A2
A3
A4
A5
S1
R1 (A1, A2, _, _, A5, A6)
A6
S2
S3
R2 (A1, _, A3, A4, A5, A6)
R1 (A1, A2, A3, A4, A5, _)
Introduction
• Query Processing in Data Aggregation
– Sending every query to all sources ?
• Increasing work load on sources
• Consuming a lot of network resources
• Keeping users waiting
– Primary processing task:
Selecting the most relevant sources regarding
difference user objectives, such as completeness
and quality of the answers and response time
– Need several types of sources statistics to guide
source selection
• Usually not directly available
Introduction
• Challenges
– Automatically gather various types of
source statistics to optimize individual goal
• Many answers (high coverage)
• Good answers (high density)
• Answered quickly (short latency)
– Combine different statistics to support
multi-objective query processing
– Maintain statistics dynamically
System Overview
System Overview
• Test beds:
– Bibfinder: Online bibliography mediator
system, integrating DBLP, IEEE xplore,
CSB, Network Bibligraph, ACM Digital
Library, etc.
– Synthetic test bed: 30 synthetic data
sources (based on Yahoo! Auto database)
with different coverage, density and latency
characteristics.
Outline
1.
2.
3.
4.
5.
6.
7.
Introduction & Overview
Coverage/Overlap Statistics
Learning Density Statistics
Learning Latency Statistics
Multi-Objective Query Processing
Other Contribution
Conclusion
Coverage/Overlap Statistics
• Coverage: how many answers a source
provides for a given query
• Overlap: how many common answers a set of
sources share for a given query
• Based on Nie & Kambkampati [ICDE 2004]
Density Statistics
• Coverage measures “vertical
completeness” of the answer set
• “horizontal completeness” is important
too – quality of the individual answers
Density statistics
measures the horizontal
completeness of the
individual answer tuples
Defining Density
• Density of a source w.r.t a given query:
Average of density of all answers
Projection Attribute set
A1
A2
A3
A4
Select A1, A2, A3, A4
From S
Where A1 > v1
Selection Predicates
Density = (1 + 0.5 + 0.5 + 0.75) / 4
= 0.675
• Learning density for every possible
source/query combination? – too costly
– The number of possible queries is exponential to
the number of attributes
Learning Density Statistics
• A more realistic solution: classify the
queries and learn density statistics only
w.r.t the classes
• Assumption: If a tuple t represents a
real world entity E, then whether or not t
has missing value on attribute A is
independent to E’s actual value of A.
Projection Attribute set
Select A1, A2, A3, A4
From S
Where A1 > v1
Selection Predicates
Learning Density Statistics
• Query class for density statistics: projection
attribute set
• For queries whose projection attribute set is
(A1, A2, …, Am), 2m different types of answers
A1
A2
22 different density patterns:
dp1 = (A1, A2)
dp2 = (A1, ~A2)
dp3 = (~A1, A2)
dp4 = (~A1, ~A2)
Density([A1, A2] | S) = P(dp1 | S) * 1.0 + P(dp2 | S) * 0.5
+ P(dp3 | S) * 0.5 + P(dp4 | S) * 0.0
Learning Density Statistics
R(A1, A2, …, An)
2n possible projection attribute set
(A1)
(A1, A2)
(A1, A3)
…
(A1, A2, …, Am)
…
2m possible density patterns
(A1, A2, …, Am)
(~A1, A2, …, Am)
(~A1, ~A2, …, Am)
…
(~A1, ~A2, …, ~Am)
For each data source S, the mediator needs to estimate
joint probabilities!
Learning Density Statistics
• Independence Assumption: the probability
of tuple t having a missing value on
attribute A1 is independent of whether or
not t has a missing value on attribute A2.
• For queries whose projection attribute set
is (A1, A2, …, Am), only need to assess m
probability values for each source!
Joint distribution: P(A1, ~A2 | S) = P(A1 | S) * (1 - P(A2 | S))
Learned from a sample of the data source
Outline
1.
2.
3.
4.
5.
6.
7.
Introduction & Overview
Coverage/Overlap Statistics
Learning Density Statistics
Learning Latency Statistics
Multi-Objective Query Processing
Other Contribution
Conclusion
Latency Statistics
• Existing work: source specific
measurement of response time
– Variations on time, day of the week,
quantity of data, etc.
• However, latency is often query specific
– For example, some attributes are indexed
• How to classify queries to learn latency?
– Binding Pattern
Same
different
Latency Statistics
Using Latency Statistics
• Learning is straightforward: average on
a group of training queries for each
binding pattern
• Effectiveness of binding pattern based
latency statistics
Outline
1.
2.
3.
4.
5.
6.
7.
Introduction & Overview
Coverage/Overlap Statistics
Learning Density Statistics
Learning Latency Statistics
Multi-Objective Query Processing
Other Contribution
Conclusion
Multi-Objective Query Processing
• Users may not be easy to please…
– “give me some good answers fast”
– “I need many good answers”
–…
• These goals are often conflicting!
– decoupled optimization strategy won’t work
– Example:
• S1(coverage = 0.60, density = 0.10)
• S2(coverage = 0.55, density = 0.15)
• S3(coverage = 0.50, density = 0.50)
Multi-Objective Query Processing
• The mediator needs to select sources
that are good in many dimensions
– “Overall optimality”
• Query selection plans can be viewed as
3-dimentional vectors
• Option1: Pareto Optimal Set
• Option2: aggregating multi-dimension
vectors into scalar utility values
Combining Density and Coverage
Combining Density and Coverage
Combining Density and Coverage
Multi-Objective Query Processing
• discount model
• weighted sum model
2D coverage
Multi-Objective Query Processing
Outline
1.
2.
3.
4.
5.
6.
7.
Introduction & Overview
Coverage/Overlap Statistics
Learning Density Statistics
Learning Latency Statistics
Multi-Objective Query Processing
Other Contribution
Conclusion
Other Contribution
• Incremental Statistics Maintenance (In
Thesis)
Other Contribution
• A snapshot of public web services (not
in Thesis) [Sigmod Record Mar. 2005]
Implications and Lessons learned:
•Most publicly available web services
support simple data sensing and
conversion, and can be viewed as
distributed data sources
•Discovery/Retrival of public web services
are not beyond what the commercial
search engines do.
•Composition:
•Very few services available – little
correlations among them
•Most composition problems can be
solved with existing data integration
techniques
Other Contribution
• Query Processing over Incomplete
Autonomous Database [with Hemal Khatri]
– Retrieving uncertain answers where constrained
attributes are missing
Select * from cars where model = “civic”
Make
Model
Year
Location
Mileage
Price
Body Style
Color
Honda
Civic
2000
Tempe
40,000
6,000
Coupe
Red
Honda
Civic
1998
Phoenix
60,500
5,000
Coupe
White
Honda
Civic
2001
Chandler
35,000
8,000
Sedan
NULL
Honda
NULL
2000
Tempe
50,000
6,800
Sedan
Beige
Honda
NULL
1999
Mesa
55,000
5,750
Coupe
Silver
– Learning Approximate Functional Dependency and
Classifiers to reformulate the original user queries
(Make, Body Style)
Model
Q1: select * from cars where make = Honda and BodyStyle = “sedan”
Q2: select * from cars where make = Honda and BodyStyle = “coupe”
Conclusion
• A comprehensive framework
– Automatically learns several types of
source statistics
– Uses statistics to support various query
processing goal
• Optimize in individual dimensions (coverage,
density & latency)
• Joint Optimization over multiple objectives
• Adaptive to different users’ own preferences
– Dynamically maintains source statistics