Early%20Profile%20Pruning%20on%20XML-aware%20Publish-Subscribe%20Systems

1
Early Profile Pruning on XML-aware
Publish-Subscribe Systems

• Mirella M. Moro, Petko Bakalov, Vassilis J.
  Tsotras
• University of California, Riverside

2
Overview

• Motivation
• Bottom-up Filtering FSM (BUFF)
• Bounding-based XML Filtering (BoxFilter)
• Core Modules
• Filtering algorithms
• Experimental results

3
Motivation

• Publish-subscribe systems The message
  transmission is defined by the message content
• Examples notification websites hotwire.com or
  ticketmaster.com

Publisher
Publisher
Publisher
Publisher
Docu ments
Docu ments
Docu ments
Docu ments
Matching algorithm
Re su l t
Re su l t
Re su l t
Re su l t
Prof ile
Prof ile
Prof ile
Prof ile
Submit, Update, Delete
Submit, Update, Delete
Submit, Update, Delete
Submit, Update, Delete
Subscriber
Subscriber
Subscriber
Subscriber
4
Publish-subscribe systems

• The data is exchanged in XML format.
• Nodes - correspond to elements, attributes or
  text values
• Edges represent immediate element-subelement or
  element-value relationships

ltBibgt ltarticle vol7 no11gt lttitlegtt1lt/titl
egt ltauthorgt ltlastgtDeWittlt/lastgt ltmigtJlt/migt
ltfirstgtDavidlt/firstgt lt/authorgt ltjournalgtTP
DSlt/journalgt ltyeargt1996lt/yeargt lt/articlegt ltart
iclegt lttitlegtt2lt/titlegt ltauthorgt ltlastgtFlor
escult/lastgt ltfirstgtDanielalt/firstgt lt/authorgt
ltproceedingsgtSIGMOD lt/proceedingsgt lty
eargt2006lt/yeargt lt/articlegt lt/Bibgt
(a) Document
(b) Tree representation
5
Publish-subscribe systems (cont.)

• The user profiles are expressed in XML query
  language (XPath, XQuery)
• XML query contains
• structural constraints
• value-based constraints

Structural constraints ////article/author_at_last
Smith''//procs_at_confVLDB''
Tree pattern
article
proceedings
author
conf
last
6
Related Work/Our Contribution

• Current work
• Construction of overlay network
• Dissemination/indexing of profiles (queries)
• Processing of stream of messages
• We focus on the matching process that takes place
  within a broker
• Improves the performance of regular FSM by using
  a bottom-up evaluation of the document
• Develop index-based filtering technique that
  performs early pruning of the query profile

7
Overview

• Motivation
• Bottom-up Filtering FSM (BUFF)
• Bounding-based XML Filtering (BoxFilter)
• Core Modules
• Filtering algorithms
• Experimental results

8
Bottom-up vs. Top-down filtering

• State machines are among the most common methods
  for the XML matching process
• Top-down approach (i.e. in-order traversal or
  depth first order) advancing the state machine
  for each XML element (or attribute) read.
• Do not consider any form of early pruning
• Bottom-up approach This approach takes into
  consideration the (usual) fact that an XML
  document has its more selective elements located
  at its leaves

9
Example

• Top-down approach groups the queries according to
  their common prefixes
• Bottom up groups them according to their common
  suffixes.

root
Q2 a
Q5 e
Q3 a
a
a
a
a
a
a
a
a
a
a
a
c
f
e
b
b
b
b
b
b
b
b
b
b
b
d
h
f
c
c
c
c
c
c
c
c
c
c
c
d
(b) Queries
(a) Document
c
d
a
3
2
3
4
4
b
b
c
Q1
1
2
Q1
d
a
c
d
1
5
6
5
e
Q2
Q2
a
f
h
f
e
a
7
8
9
0
6
7
8
0
Q4
Q3
Q3
h
e
h
e
a
f
f
11
12
10
11
12
Q5
10
9
Q4
Q5
g
g
h
e
13
14
13
14
Q6
Q6
(d) Bottom up
(c) Top-down
10
BUFF

• FSM-based Bottom-up approach for XML filtering.
• BUFF avoids translating documents and queries to
  Prüfer sequences (as the other algorithms do),
  and employs a more direct evaluation algorithm.
• The document is parsed through a SAX parser,
  which triggers events for specific marks (tags)
  in the XML document
• The machine keeps a runtime stack that stores the
  current document path being processed.

11
BUFF Example
e
ltegt
d
ltdgt
c
d
b
1
2
3
4
e
c
ltcgt
Q1
0
f
b
ltbgt
c
a
b
5
6
7
8
a
ltagt
Q2
(a) Document and BUFF
(b)?
(c)?
(d)?
(e)?
(f)?
(g)?
12
Overview

• Motivation
• Bottom-up Filtering FSM (BUFF)
• Bounding-based XML Filtering (BoxFilter)
• Core Modules
• Filtering algorithms
• Experimental results

13
Bounding-based XML Filtering

• Two major processes working asynchronously
• Profile Management
• Profile Matching

Prüfer Sequence
Profile Manager
Matching Algorithm
Matching Module
Input Documents
Profiles (queries)?
Matched Documents
14
Prüfer Sequence

• A unique sequential encoding of a labeled tree
• Algorithm
• Iteratively removes nodes from the tree until
  all nodes but the last two have been removed.
• At each iteration, the algorithm finds and
  removes the leaf with the smallest label and adds
  to the Prüfer sequence the label of that leaf's
  parent.
• Theorem If a query tree Q is a subgraph of a
  document tree D then the Prüfer sequence of Q is
  a subsequence of the Prüfer sequence of D

15
Sequence Envelope

• Assume a set of k Prüfer sequences representing
  user profiles S1,..,Sk
• We can derive two new sequences
• Upper bound U for each position take largest
  element
• Lower bound L for each position take smallest
  element
• L and U form the smallest possible bounding
  envelope that encompasses all members of the set
  of sequences from above and below.

16
Example

• Assume 3 sequences with 11 symbols each
• abcabababcd
• cdcdecdcdec
• dedededebab

17
Sequence Envelope (Cont.)

• The sequence envelope structure is that it can be
  used as an aggregation of the sustaining set of
  sequences

18
BoXFilter Tree

• Sequence envelopes can be nested forming
  BoXFilter tree

19
Filtering algorithms

• The profiles in the system are organized in
  BoXFilter tree. Documents are traversed thought
  the tree
• There are two variations of the filtering
  algorithm
• Sequential documents are processed one by one
• Batch processing documents are organized in a
  tree like the queries and both trees are joined
• After the traversal of the BoXFilter tree, there
  is a verification step

20
Overview

• Motivation
• Bottom-up Filtering FSM (BUFF)
• Bounding-based XML Filtering (BoxFilter)
• Core Modules
• Filtering algorithms
• Experimental results

21
Experimental Results

• We have generated datasets with 1000, 10000 and
  100000 small documents (with up to 8KB)
• We generated up to 100000 queries with
  selectivity fixed to 50

(a)?
(b)?
(c)?
22
Experimental Results (cont.)

• In this set of experiments, we vary the number
  of documents that match any of the profile
  queries. (selectivity 1\ means that one percent
  of the documents satisfy \textitany of the
  queries.)

23
Thank You!