Sam Idicula, Oracle XML DB Development Team

1
Binary XML Storage and Query Processing in Oracle
VLDB 2009

• Sam Idicula, Oracle XML DB Development Team

2
Outline

• Motivation
• Binary XML Overview
• Storage Format Details
• Query Processing
• Performance Evaluation
• Conclusion

3
Previous Oracle XML Storage Models

• CLOB Storage
• Text representation preserves exact form of
  original document (including white spaces)
• Very good performance for insert full retrieval
• Size bloat (including tags, string representation
  of dates, numbers etc)
• Need to parse the document for all XML processing
• Query DML processing are not efficient
• Memory overhead with DOM
• Mid-tier does not take advantage of parsing and
  validation already done on DB tier (and
  vice-versa)

4
Previous Oracle XML Storage Models

• Object Relational Storage (OR)
• XML Schema-based mapping to object-relational
  tables
• Preserves DOM fidelity (more than traditional
  shredding)
• Simple XPaths translate to table/column access
• Very good query performance for highly structured
  use cases
• Flexibility is limited due to schema dependency
• Insert, full retrieval etc are poor (expand on
  this separate these into 2 slides)

5
Motivation/Goals for Binary XML

• Bridge the gap between two extremes
• Structure-unaware text representation Full
  flexibility, poor query performance
• Object-relational mapping Heavily dependent on
  rigid structure
• Several customer use cases fall in between these
  extremes
• Native format that can
• Handle full spectrum of XML database use cases
• Optimized semi-structured use cases
• Provide good performance for a wide variety of
  operations
• Retain flexibility advantage of XML data model
  while providing good performance

6
Customer use cases
High Flexibility
Majority of semi-structured customer use cases
Low Flexibility
7
Motivation/Goals for Binary XML

• XML Schema usage
• Need to be efficient for query processing on
  schemaless loosely structured schemas
• Ability to use schema constraints for more
  efficient processing
• Provide good performance for a wide range of
  operations
• Query
• DML Insert/Load, Partial (piecewise) update
• Full-document fragment retrieval
• Schema Validation Evolution
• Mid-tier integration

8
Oracle Binary XML Overview

• Compact Schema-aware XML Format
• Pre-parsed tokenized binary representation
• Addresses space-bloat associated XML 1.x
  serialization
• Intended for use in all tiers of Oracle stack
• Oracle XML DB
• Oracle iAS / XDK Java
• Exploits XML Schema information if available
• Also supports non-schema-based encoding
• Preserves Infoset or Data Model fidelity Not
  bytes
• Can create an XML Index for query optimization

9
Oracle Binary XML
WebCache
AppServer
Database
Client
Oracle Binary XML

• Mid-tier Processing Oracle XDK Java support
• Binary XML allows direct access to
  fragments/sub-trees
• XML processing optimization Scalable mid-tier DOM

10
Format Details

• Opcodes roughly corresponding to SAX events
• Each opcode has fixed number of operands
• Document-ordered serialization of opcodes
• Stored as a BLOB
• Tag names are tokenized into qname IDs
• Central repository (or)
• Inlined definitions
• Optimized opcodes for simple elements, repeating
  elements etc.
• Uses native data-types in the presence of XML
  schema

11
Streaming Capabilities

• Streaming XPath evaluation
• XPathTable with NFA Multiple XPaths evaluated in
  a single pass
• Forward axes
• Streaming partial updates
• Most common update scenarios handled in streaming
  manner
• eg updateXML( /purchaseOrder/Reference/text()
  , XXXX)
• Can be directly applied on disk avoiding
  expensive DOM construction
• Takes advantage of the Oracle SecureFile LOB
  storage to perform delta update

12
Query Processing Architecture
XQuery
SQL/XML
DB XQuery Rewrite
XMLIndex
Functional Evaluation (Streaming XPath)
Path-based XMLIndex
Table-based XMLIndex
Binary XML
13
Document-level Summary

• Long-term goal Efficient tree-oriented
  navigation
• Important for query execution
• Pure streaming is too costly over large documents
• Current Implementation
• Start end offsets for large subtrees
• Threshold for large can be adjusted
• Used for skipping to end of subtree
• Working on significant enhancements
• Handling all axes

14
Search-based Decoder

• Goal Search for a simple XPath or XPath location
  step in a Binary XML stream
• Main search params are (axis, qname ID)
• Supports wild cards
• OR of multiple qnameIDs allowed
• Return only when theres a result or search is
  done
• Skip irrelevant subtrees
• Using summary if possible
• Schema-aware search
• Can search for kidnum or child-position instead
  of qname ID
• Can terminate search earlier based on schema

15
Schema-aware NFA

• Goal Evaluate multiple XPaths in single pass
  over document
• Uses Y-Filter-like approach to build NFA
• Works in conjunction with search-based decoder
• Translates transitions to searches when possible
• Push unbranched linear state transition paths
  into search-based decoder
• Uses XML schema when available
• Use of kidnum instead of qname ID
• Sequence Occurrence constraints
• Derives a strict sequential constraint

16
Performance Query - XMark

• Ratio of elapsed time geometric mean for 100M
  XMark doc
• SB Schema-based
• NSB Non-schema-based
• CLOB is 144x
• No indexes

17
Performance Insert

• Ratio of elapsed time for XMark 10M doc
• SB Schema-based

18
Performance Full Retrieval

• Ratio of elapsed time for XMark 10M doc
• SB Schema-based

19
Performance Compression

• D1 Structured
• D2 Semi-structured
• D3 Document-centric
• Based on actual customer datasets mix of XML
  document sizes
• Further compression possible via SecureFile LOB
  compression

20
Summary

• Binary XML
• Native XML storage format
• Handle full spectrum of XML use cases
• Schema-aware
• Query Processing Optimizations
• Search-based Decoder
• Document-level Summary
• Performance Results

21
For more information

• Contact
• Sam.Idicula_at_Oracle.com
• Downloads, technical documentation
• http//www.oracle.com/technology/tech/xml/xmldb/
  index.html