Selective Dissemination of Streaming XML

1
Selective Dissemination of Streaming XML

• By Hyun Jin Moon, Hetal Thakkar

2
Overview

• Introduction
• Background
• XFilter
• Architecture
• Implementation
• Optimizations
• Experiments/Analysis
• Conclusion
• Related Work XTrie

3
Introduction

• Information Dissemination
• Enormous Amount of Data
• Lots of Users
• User Profiles
• Bag of Keywords
• Selective Distribution of Data
• Applications
• Stocks, Sports, Traffic, Electronic Personalized
  Newspapers, Entertainment, etc.

4
Introduction (Contd)

• Emergence of XML as Standard of Information
  Exchange on Internet
• Utilize Structure of XML for Better Dissemination
• Use XPath(s) for User Profile
• Optimizations for Searching a Streaming XML
  Document for Many XPaths
• XFilter
• XTrie Structure

5
Background

• SDI Structure
• XPath
• XML Parsers
• DOM
• SAX

6
Background SDI Architecture
7
Background XPath

• Query Structure and Data
• Enough Complexity for Dissemination
• Constructs
• Relative Path
• //productprice/msrplt300/name

8
Background XML Parser

• DOM Document Object Model
• SAX Simple API for XML (SAX)
• Standard Interface for Event-Based XML Parsing
• Suitable for Streaming XML
• Example

9
XFilter

• Architecture
• Implementation
• Optimizations
• List Balancing
• Prefiltering
• Experiments/Analysis
• Conclusions

10
XFilter Architecture
11
XFilter Implementation

• Filter Engine
• Brute Force Approach
• Instead,
• Decompose Queries into Path Nodes
• Create a Query Index from Path Nodes
• Build a Finite State Machine on the Query Index
• As a Document Arrives Traverse the FSM for All
  Queries (In One Pass)

12
XFilter Implementation

• Path Nodes
• QueryId
• Position
• Sequence Number for Path Node in the Query
  (XPath)
• RelativePos
• Relative Distance in the Document
• Level (Can be Updated During Evaluation)
• Absolute Level in the XML Document, at Which the
  Path Node should be Checked

13
XFilter Implmentation

• Query Index
• Hash Table
• Key Element Names that Appear in XPath
  Expressions
• Data 2 Lists Containing Path Nodes
• Candidate List Current Node of Each Query
  Representing Current State of the Query
• Wait List Path Nodes Representing Future States

14
XFilter Implementation
15
XFilter Implementation

• Start Element Handler
• Inputs Name, Level, and Attribute-Values of the
  Element
• Action
• Look-up Element Name in Query Index
• Examine Nodes in Candidate List
• Check Level, etc.
• If All Checks Succeed AND Final Path Node of
  Query Then the Document is Deemed to Match the
  Query
• Else If All Checks Succeed Then Move the Query to
  its Next State
• Else Do Nothing

16
XFilter Implementation

• End Element Handler
• Input Element Name
• Action
• Delete the Corresponding Path Nodes from the
  Candidate List (for Restoring Purpose)
• Element Character Handler
• Input Data
• Action Similar to Start Element Handler

17
XFilter Implementation

• Example
• Start Document
• Start Element a Level 1
• Start Element b Level 2
• Start Element c Level 3
• End Element c
• End Element b
• End Element a

18
XFilter Implementation
19
XFilter Implementation

• Advanced Features
• Attribute Filter
• Start Element Event Handler
• Content Filter
• Element Character Handler
• Nested Path Expression
• Treat Nested Sub-Queries as Another Query

20
XFilter Optimizations

• List Balancing (LB)
• Basic Approach First Path Node for Each Query in
  the Candidate List
• Low Selectivity
• Instead, Apply Candidate List Balancing
• When Adding a New Query to Query Index the Path
  Node Who has the Shortest Candidate List is
  Chosen as the Pivot Node
• Prefix

21
XFilter Optimizations

• Prefiltering
• Eliminate Queries, which have Element Name(s)
  that are not Present in the Document
• Yan and Garcia-Molinas Key Based Algorithm
• Assign Key Element of the Queries
• Create Occurrence Table for Each Arriving
  Document
• Occurrence Table Hash Table
• Key Element Name
• Data Queries, Whose Key is this Element
• Only Queries in Occurrence Table are Checked
  Further
• Thus, Each Input Document is Parsed Twice

22
XFilter Experimental Setup
23
Experiment 1.1 The Effect of Number of Profiles

• Number of Profiles (Standing XPath Queries)
  Changes
• Basic Algorithm Gives the Worst Performance
• List Balance Improves
• Prefiltering Leads to a Greater Speed-Up Than LB
• 2.6 of Profiles Match a Given Document
• Basic Algorithm Examines 12 of Profiles
• Prefiltering Examines Only 3.5 of Profiles

24
Experiment 1.2 The Effect of Number of Profiles

• Number of Profiles Changes Same as Before
• Skewed Selection of Elements Leads to
  Unbalanced Query Index (Hash Table) in Basic
  Algorithm
• List Balance is Effective in Balancing the Hash
  Table

25
Experiment 2.1The Effect of Depth

• Maximum Depth of XML Documents and Queries Change
• More Depth -gt More Checking -gt Greater Filtering
  Time
• List Balance and Prefiltering Graphs cross at
  Depth 8. With Higher Depth,
• Less Prefiltering
• LB Benefits with More Choices of Pivot Elements

26
Experiment 2.2The Effect of Depth

• Maximum Depth of XML Documents and Queries Change
• Skewed Selection of Elements
• LB Effectively Balances the Skewed Hash Table
• After Level 4, the Presence of Element Names in
  the Queries does not Change Much Due to Skewed
  Distribution. Workload Characteristics Remain
  Similar.

27
Experiment 3The Effect of Wildcard

• Wildcard () Usage Probability in Queries
  Change
• Prefiltering is Slower with More Wildcards
• Prefiltering Takes Extra Time Trying Filtering,
  but Prefiltering cannot Filter Out the Wildcards
• However, it is Unlikely that Many Profiles will
  have such a High Proportion of Wildcards.

28
Experiment 4.1The Effect of Filter

• Injected a New Fixed Attribute Named dummy into
  the Documents with Certain Probability
• Created a Simple Element Node Filter Containing
  Only that Fixed Attribute
• (e.g. _at_dummytrue)
• In this Experiment, a Single Element Node Filter
  is Placed in Different Levels of the Query with
  Fixed Query Selectivity of 10
• The Deeper the Filter, the Longer it Takes to Test

29
Experiment 4.2The Effect of Filter

• Filters are Placed at Level 2, with Varying
  Selectivity.
• Logarithmic Scale on Selectivity
• For All Algorithms, Performance is not Heavily
  Affected by Filter Selectivity

30
Summary of Results

• These Experiments Demonstrate that,
• XFilter approach is scalable
• The Extensions Provide Substantial Improvements
• List Balance is Effective When the Distribution
  of Elements in Queries is Highly Skewed
• Prefiltering is Effective in Reducing the Number
  of Profiles to Examine
• Combination of LB-Prefiltering Provides the Best
  Performance in All Cases
• Considering that Distribution of Elements in
  Queries of SDI Applications is Highly Skewed, and
  Prefiltering Requires a Space Overhead, Simple LB
  is Preferable in Many Practical Cases

31
Conclusions

• XML Document Filtering System XFilter for
  Selective Dissemination of Information (SDI)
• Expressive Profiles in XPath Query Language
• Profile Indexing and Matching Algorithms Based on
  a FSM Approach
• Optimization Techniques
• List Balancing
• Prefiltering

32
Related Work XTrie

• Efficient Filtering of XML Documents with XPath
  Expression ICDE 2002
• Supports Complex XPath Expressions (As Opposed to
  Simple, Single-Path Specifications)
• e.g. /a/bc/d//eg//e/f////e/f
• Supports Both Ordered and Unordered Matching of
  XML Data
• Ordered Matching //a//b/following-siblingd/c
  
• Substring-Based Query Indexing
• 2 to 4 Times Faster Than XFilter