1 evaluating top-k queries over web-accessible databases paper by: amelie marian, nicolas bruno,...
TRANSCRIPT
1
Evaluating top-k Queries over Web-Accessible Databases
Paper By: Amelie Marian, Nicolas Bruno, Luis Gravano
Presented By
Bhushan Chaudhari
University of Texas at Arlington
2
Overview
More importance to top-k results Fagin’s algorithm talks about effective differentiation
between top-results by various ways e.g. FA, TA Here we discuss about more larger scenario in terms of web-
accessible databases Assumption: Mapping of keywords typed from search text
box to appropriate related modules (Web-accessible databases)
Larger query response times for probing web sources Tries to exploit the parallel access offered by web
3
Introduction
We never expect exact answers from search engine but the most nearest possible tuples
Difference between querying a general search engine and dedicated search engine e.g. Google vs Amazon
The paper tries to define the problem using example of restaurants
“ problem of finding nearest available restaurants given the current place, rating and price”
4
Approach
Thinking beyond relational databases Web accessible sources storing information about rating of
restaurants, map provider system etc.Rating => Zagat-Review websitePrice => New York Times’s NYT-Review websiteAddress => MapQuest website
Scenario where databases are geographically and functionally different but are related “in some way”
Assumption: 1. The interface required for accessing web sources is in place the dependency can be handled 2. The dependency constraints are handled
5
Approach (continued ..)
Can be compared with a similar scenario with several multimedia systems which are more closely connected
Here we try to use the intrinsic parallel nature of web We issue probes to various sources in parallel and try to
improve upon the final query processing time Assumption: Mapping of keywords typed in search text box
to routing it to appropriate related modules (Web-accessible databases)
Larger query response times for probing web sources Tries to exploit the parallel access offered by web
6
Data and Query models
The ordering is bases upon how closely the tuple matches with given query
Assignment of different weight to different attribute
Sources S-Source: Provides list of objects in order of their scores
e.g. Rating provider website Zagat-Review R-Source: Provides score of random object e.g. Map-Quest
for providing distance SR-Source: Source that provides both kind of access
U(t) : Upper bound score for t Uunseen : Score upper bound of any object not yet retrieved E(t) : Expected score for t
7
Query Model (continued ..)
Getting all k scores with S sources can be expensive Therefore availability of SR sources is important for this
approach Initially we assume that all object know about all other
object If any score is not possible to get then that can be replaced
with some default valuee.g. Opening of any new restaurant, it might not be ranked by other referencing websites
8
Sequential Query ProcessingStrategy
This strategy returns sorted unseen objects that might not be probed by other source
Or it can return already seen object with source that needs to be probed randomly for getting the corresponding score
9
TA strategy
Processes top-k queries over SR sources Algorithm retrieves the next “best” object via sorted access Probes all its unknown scores via random access Computes the final score for object At any given time keeps track of top-k tuples available When no unretrived object can have a score higher than
current top k tuples, the solution is reached
10
11
Improvements upon TA
The assumption for bounded buffer is removed and none of the object is discarded until algorithm returns
Because same objects might be referenced again by different SR source
For selection queries of nature,p1^p2^…^pn The calculation of each predicate pi can be expensive to
calculate Key idea is to order the evaluation to minimize expected
execution time The order is decided by,
Rank(pi) = 1-selectivity(pi)/cost-per-object(pi)
12
Improvements upon TA (Continued..)
Let w1, w2, …w2 be the weights of sources D1,D2,..,Dn Let e(Ri) be the expected score of randomly picked object Ri Then the expected decrease in U(t) after probing Ri for
object t is,di = wi * (1-e(Ri))
We sort the sources in decreasing order of their rank, where rank for a source Di is defined as,Rank(Ri) = di/tR(Ri)
Thus we favor fast sources that might have large impact on final score of object
13
14
Upper Strategy
Upper allows more flexible probes in which sorted and random accesses can be interleaved even when some objects have been partially probed
When a probe completes the Upper decides whether- to perform sorted-access probe on source to get new
objects to perform “most promising” random access probes on
some objects
15
Upper Strategy (Continued..)
16
Upper Strategy (Continued..)
Selection of further probes will again depend upon the weight for that source and our ranking function
17
Parallel Query Processing Strategy
The query processing is bound to take long processing time Web databases exhibit high and variable latency Attempt to maximize the source-access parallelism to
minimize query processing time
Source Access Constraints Possibility of access restrictions, variance in loads and
network capabilities The number of parallel probes for source Di can be
controlled
18
Parallel Query Processing Strategy
Adapting the TA strategy When a source Di becomes available pTA chooses which
object to probe for that source It can be optimized by not probing objects whose final
score cannot exceed that of the top-k objects already seen
The object is put on the “discarded” objects list pUpper Strategy
If t is expected to be one of the top-k objects all random accesses on sources for which t’s attribute score is missing will be considered
Otherwise only fastest probes expected to discard t are considered
19
Evaluation Settings
Local sources Real Web Accessible sources
Mix of SR and R sources
20
Evaluation Results
Sequential Algorithms – Local Database
21
Evaluation Results
Sequential Algorithms –Web Database
22
Evaluation Results
Parallel algorithms - Local Database
23
Evaluation Results
Parallel algorithms - Web Database pUpper is faster than pTA pUpper carefully selects the probs for each
object It considers probing time and source
congestion to make probing choices per object-level
Results in better use of parallelism and faster query processing
24
Conclusion
Probe interleaving greatly improves query execution time
Upper is desirable when source shows moderate to high random access time
The approach in this paper exploits the source access constraint of web very well
Extension of this model to capture more expressive web interfaces is possible
25
References
Optimal Aggregation Algorithms for Middleware. PODS 2001
Ronald Fagin, Amnon Lotem, Moni Naor Evaluating Top-k Queries over Web-Accessible
Databases. ICDE 2002 (Compact Version) Nicolas Bruno, Luis Gravano, Amelie Marian
Evaluating Top-k Queries over Web-Accessible Databases. ACM 2004 (Full Version)
Nicolas Bruno, Luis Gravano, Amelie Marian