A Fast Index For Semistructured Data

1
A Fast Index For Semistructured Data

• Presented by
• Alexandra Martinez
• CIS 6930 - Indexing Large Databases

2
Indexing Semistructured Data

• Introduction
• Detour Tries and Patricia Tries
• The Index Fabric Structure
• Indexing XML with the Index Fabric
• Results and Conclusion

3
Introduction

• Semistructured Data
• Data with an irregular or changing organization.
• Often represented as a graph (elems relations
  schema)
• Queries over semistructured data
• Navigating paths through graph
• Indexes usually built for efficient access.
• Conventional techniques
• Use a relational database (translation, querying
  is costly)
• Use a native semistructured repository (new, bad
  query perf)

4
Introduction

• Proposed approach
• Relies on relational db but provides better
  performance.
• Encodes data paths as strings and insert them
  into an index that is optimized for string
  searching.
• For evaluating a query, the path is encoded as a
  search key string, then we do a lookup on index.
• () No need for knowing data schema a priori.
• () High perf. with changing and irregular
  structure.
• () Can accelerate queries along diff. access
  paths.

5
Indexing Semistructured Data

• Introduction
• Detour Tries and Patricia Tries
• The Index Fabric Structure
• Indexing XML with the Index Fabric
• Results and Conclusion

6
Detour Tries Patricia Tries

• A Trie is a tree that stores strings it
  represents each character as an edge on the path
  from the root to a leaf.
• Patricia tries (PT's) are a more compact form of
  tries. A PT is similar to a trie, except that
  nodes with only one child have been removed.
• The numbers inside the nodes (depth) indicate the
  character position in the string to compare to
  the labels on the outgoing edges.
• PTs achieve compression at the cost of no longer
  storing the complete keys, but rather the
  differences between keys.
• PTs are unbalanced structures.

7
Detour Tries PTs Example
key2
A trie indexing one string
A trie indexing mulitple strings
A Patricia trie indexing mulitple strings
8
Indexing Semistructured Data

• Introduction
• Detour Tries and Patricia Tries
• The Index Fabric Structure
• Balancing Patricia tries
• Two kinds of links
• Searching
• Updates
• Indexing XML with the Index Fabric
• Results and Conclusion

9
Fabric Balancing Patricia Tries

• PTs are not balanced in large dbs, unbalance
  can be large result in performance degradation
• Problem is solved by introducing multiple layers
  into the PT.
• Horizontal layers are added to skip some of the
  vertical levels.
• Horizontal structure is always balanced.
• Balancing the PT allows for searches (and
  updates) in time proportional to the number of
  layers instead of the length of the indexed keys.

10
Fabric Balancing PTs Example
Layer 1 indexes common prefixes of each subtrie
(block) in Layer 0
Root is always at leftmost layer

c
11
Fabric Balancing Patricia Tries
12
Fabric Two kinds of links
a

• Labeled Far Link ( )
• This link is the same as an edge between a parent
  and a child in a normal trie, except that the
  parent is in layer i1 and the child is in layer
  i.
• Unlabeled Direct Link ( )
• Connects a node in layer i1 with a node
  representing the same prefix in layer i.

13
Fabric - Searching

• Start at the root node of the block at leftmost
  layer (layer N)
• Within a block, compare characters in search key
  to edge labels, and keep following the edges.
• If edge is a far link search proceeds to a
  block in next layer (N-1)
• If this block miss, backtrack (follow direct link
  instead of far link).
• If no labeled edge matches follow a direct link
  to a new block in the next layer (N-1)
• Eventually layer 0 is reached.
• If no labeled edge matches key not found.
• Otherwise, a path if followed to the data.

14
Fabric Search Examples castle, fast
15
Fabric Searching One I/O

• Search accesses one block per layer.
• Since horizontal layers are balanced, all
  searches traverse same layers gt access same
  blocks.
• Compact storage of keys - blocks have very high
  fan out PT has low height
• Ex 3 layers - sufficient to store a billion keys
• 2 upper layers in memory, layer 0 on disk.
• Thus searches require a single index I/O.

16
Fabric Updates

• Similar to B-trees. Very efficient, like
  searches.
• Insertion
• Involves a change to single block in the lowest
  layer (layer 0). If a block has no space for
  insertion, it splits. Splits may cascade to
  higher layers.
• Deletion
• Find block to be updated, remove edge pointing to
  the key. Blocks might merge to compact trie.
• Updates
• Deletion followed by Insertion.

17
Indexing Semistructured Data

• Introduction
• Detour Tries and Patricia Tries
• The Index Fabric Structure
• Indexing XML with the Index Fabric
• Designators
• Raw Paths
• Refined Paths
• Results and Conclusion

18
Indexing XML Example XML
19
Indexing XML - Designators

• Designator - A unique special character(s)
  assigned to each tag that appears in the XML
• Designator Dictionary maintains mapping between
  tags and designators

Designator Dictionary
The designator-encoded XML strings are inserted
into the Index Fabric.
20
Indexing XML Raw Paths

• Raw paths index the hierarchical structure of the
  XML by encoding root-to-leaf paths as strings.
• Simple path expressions starting at the root
  requires a single index lookup.
• Ex the XML fragment
• ltAgtalphaltBgtbetaltCgtgammalt/Cgtlt/Bgtlt/Agt
• Can be represented as a tree with 3 root-to-leaf
  paths
• 1)ltAgtalpha 2)ltAgtltBgtbeta 3)ltAgtltBgtltCgtgamma
• which are encoded as
• 1)A alpha 2) A B beta 3) A B C gamma
• Under mapping f, where
• f (ltAgt) A, f (ltBgt) B , f (ltCgt) C

21
Indexing XML Raw Paths Example
1
alpha
A
A alpha A B beta A B C gamma
2
beta
B
3
ltAgtalpha
C
ltBgtbeta
gamma
ltCgtgamma
lt/Cgt
lt/Bgt
lt/Agt
22
Indexing XML Refined Paths

• Specialized paths through XML that optimize
  frequently occurring access patterns.
• Can support queries that have wildcards (,),
  alternates (), and constants.
• DBA decides which refined paths are appropriate

23
Refined Paths An Example

• Freq.Q Find the invoices where company X sold to
  company Y. Find ltbuyergt tags that are siblings
  of a ltsellergt tag.
• Assign a designator Z to such a path
• Encode info indexed by this ref. path in a key
• gt
• Insert the created keys into the Index Fabric.
• Keys refer to the XML fragments that answer Q.

ltinvoicegt ltbuyergtYlt/buyergt
ltsellergtXlt/sellergt lt/invoicegt
ltinvoicegt ltbuyergtABC Corplt/buyergt
ltsellergtAcme Inclt/sellergt lt/invoicegt
Z ABC Corp Acme Inc
24
Indexing Semistructured Data

• Introduction
• Detour Tries and Patricia Tries
• The Index Fabric Structure
• Indexing XML with the Index Fabric
• Results and Conclusion

25
Results

• Index Fabric compared to
• DBMSs native B-tree index over tables generated
  by STORED.
• DBMSs native B-tree index over tables generated
  by basic edge-mapping (roots and edges).
• Index Fabric outperforms B-tree indexes.
• Index Fabric offers significant optimization
  especially for complex queries, refined paths.

26
Conclusions

• Indexing for Semistructured data - significant
  challenges
• Many problems not yet solved - efficient
  processing of query involving complex regular
  expressions
• Index Fabric Indexing for XML stored in
  relational DB (can work for other models)
• Interesting features
• Combines aspects of Patricia Tries (scaling) and
  B-trees (balanced, optimized for disk access)
• No a priori knowledge of structure is needed

27
References

• B. Cooper et al. A Fast Index for Semistructured
  Data. In Proc.VLDB, 2001. Available at
• B. Cooper and M. Shadmon. The Index Fabric A
  mechanism for indexing and querying the same data
  in many different ways. Technical Report, 2000.
  Available at http//www.rightorder.com/technology
  /overview.pdf

28
Thanks