XML Query Evaluation

1
XML Query Evaluation
Based on Slides by Alan Halverson Raghav
Kaushik University of Wisconsin, Madison
2
Evaluating XML queries

• Native XML database
• Store and query XML in a native form
• Tamino, Excelon, XIndice
• Niagara, Timber
• XML-enabled databases
• Databases with extensions for transferring data
  between XML documents and themselves
• relational/object-oriented/
• Focus on native XML databases

3
Native XML database

• An XML document is stored in two formats in
  Niagara
• As a DOM tree, making a walk through all elements
  in document order easy
• Actual data, can reconstruct original document
• Data Manager
• As inverted lists, mapping terms to document
  locations
• Index on tag name
• Index Manager

4
Niagara System Architecture
5
Numbering Scheme

• Assign to each element a start number, end
  number, and level
• Allows for easy relationship testing
• Example B is a child element of A if and only
  if B.start gt A.start B.end lt A.end
  
• B.level A.level 1

6
Data Manager
start

• XKey (docid, elementid) pair

7
Data Manager detail

• Two types of cursors
• Child Axis (CA) given an XKey, enumerates child
  elements with optional filtering on a tag name
• Descendant Axis (DA) enumerates all proper
  descendant elements of an XKey

8
Unnest Algorithm

• Algorithm uses the Child and Descendant cursors
  provided by the Data Manager
• Top picture shows in Yellow the elements
  enumerated by a Child cursor opened at element B
• Bottom picture shows the same for a Descendant
  cursor opened at B

9
Unnest for Path expression queries
10
Unnest for Path expression queries

• Existential predicate can be handled as a
  separate state machine
• Positional predicate can be handled by augmented
  state information

11
Index Manager

• Path expression queries comprise of a series of
  steps
• Each step has a tag name and a relationship with
  context node
• Parend-child or Ancestor-descendant
• Build an inverted index on tagname
• Associate information with each occurrence to
  check containment relationship

12
Index Manager detail

• B-tree index keyed by TermID, DocID
• Each leaf entry contains a 2nd-level index
  pointing to pages of postings
• Posting (DocId, Start, End, Level)

13
ZigZag Join Example

• Each list is sorted on (document id, start)
• Merge based algorithm for evaluating path
  expression queries

Evaluate path A//B
A1 A2 A3 A4
B1 B2 B3 B4 B5 B6 B7
B1
B5
B6
B7
Output
14
ZigZag Join Algorithm

• Secondary index on start number in the inverted
  lists
• Utilize this index on the posting lists to skip
  forward over postings in one input list which
  are guaranteed not to match any future postings
  on the other list

15
ZigZag Example
Evaluate path A//B
A1 A2 A3 A4
B1 B2 B3 B4 B5 B6 B7
B1
B5
B6
B7
Output
16
Structure Indexes

• Inverted lists completely answer single step path
  expression queries
• Structure indexes can be used to answer more
  complex queries efficiently
• Can be viewed as a summary of the data graph

17
Data Model Directed Graphs
article
1
2
title
Data on the Web
citedby
section
5
3
section
4

section
title
p
figure
section
6
7
8
9
p
title
10
12
11
15
16
p
title
title
p
17
18

Introduction

img
14
13
title

Graph Representation
18
In-coming paths Backward Direction
article
1
/article//figure/title
citedby
section
5
title
3,4
2
9,12
p
section
figure
title
8
6,10
7,11
p
title
15,17
16,18
title
img
13
14
19
Outgoing Paths Forward Direction
article
1
//section/figure
citedby
section
5
title
3,4
2
9,12
p
section
figure
title
8
6,10
7,11
p
title
15,17
16,18
title
img
13
14
20
Covering

• Index I
• Path expression P
• I covers P if the result of executing P on I and
  taking the union of the extents is correct

21
Path Index Definition

• Quotient graph obtained from a partition of the
  element nodes

1
5
citedby
extent
title
3,4
2
9,12
7,11
6,10
8
15,17
16,18
13
14
22
Challenge in Path Indexing

• What is the partition we begin with?
• How do we group together nodes into equivalence
  classes?
• Requirements
• Generic Make index as widely applicable as
  possible
• Small Make index as small as possible (even the
  data graph is an index according to our
  definition!)

23
Proposal For Index Specification

• Pick only a subset of tags S to index
• Attach priority to tree edges over non-tree edges
• Specify which non-tree (ref) edges R to index
• Exploit local similarity
• Tree depth

24
Local Similarity

• section/section/figure forward look-ahead
  of 2
• section/figure forward look-ahead of 1
• section forward look-ahead of 0
• The idea is to be able to say only care about
  path expressions where the forward look-ahead is
  at most k
• In other words, group nodes based on local
  structure

25
Integrating Structure Indexes and Inverted Lists
title
article
1
citedby
title
2
XML
section
3,4
5
XML
9,12
section
p
title
figure
E2 ..


8
6,10
7,11
p
title
15,17
16,18
title
img
13
14
26
Using the Modified Lists
article/citedby/XML
article
XML
E4?
E4



section
title
citedby
section
p
title
figure
title
p
title
img
27
Putting it all together

• A path expression can be split into multiple
  parts
• Different algorithm can be applied to each part
• Optimization problem
• Need good statistics and cost models