Overview of SPGiST

1
Overview of SP-GiST

• Walid G. Aref
• Department of Computer Science
• Purdue University

2
Indexing

• With the emergence of non-traditional database
  applications, the need for non-traditional types
  of indexes is inevitable
• Example applications

3
Challenges in Indexing

• Current database systems support
• B-trees, hash tables
• Very few systems support R-trees
• Fewer systems support a variant of the region
  quadtree

4
What is Wrong?

• Building and integrating an index type into the
  database system is an overwhelming task
• Integration with the query optimizer, when to use
  the index and when not, cost model, selectivity
  estimation
• Providing query operators that utilize the index
• Concurrency control and recovery techniques
• Very few index structures/research address all
  these issues

5
SP-GiST Space-partitioning Generalized Search
Trees

• Software engineering solution to support a wide
  class of indexes inside a database management
  system
• GiST Supports B-tree-like indexes, e.g., R-trees
• SP-GiST Supports the class of space-partitioning
  trees, e.g., variants of the quadtree, variants
  of the trie, k-d tree

6
SP-GiST Space-partitioning Generalized Search
Trees

• The framework provides the basic services inside
  a database system, e.g.,
• Concurrency control and recovery
• Query operators
• Bulk-loading and insertion
• Integration with the query optimizer, costing,
  selectivity estimation
• Node clustering
• SP-GiST supports the class of space-partitioning
  (SP) trees
• Suitable for emerging database applications
• An extensible index structure that can be
  instantiated to realize any member in the class
  of SP trees
• Example index structures realizable by SP-GiST
• Disk-based versions of variants of the trie, all
  variants of quadtrees and octrees, the k-d tree,
  the bin-tree,

7
SP-GiST Extensible Interfaces

• Internal Methods
• Supported by the DBMS system
• reflect similarities among the various SP-trees,
  e.g., insert, delete, search, bulk-load
  algorithms
• Interface Parameters and External Methods
• Extensible index interfaces
• Reflect structural and behavioral differences
  among various SP-trees
• Need to be supported by the user to instantiate a
  new type of SP-tree index

8
Examples on SP Trees
Data Driven SP Trees Space is decomposed based
on the input data
Space Driven SP Trees Space decomposition is
independent from the order of data insertion
k-d Tree
Trie
9
Examples on Index Realization inside SP-GiST
10
Main Characteristics of Space-Partitioning Trees

• Decompose the space recursively into a fixed
  number of disjoint partitions
• There are two types of space-partitioning trees
• Space-driven space-partitioning trees
• e.g., the trie and region quadtree
• Data-driven space-partitioning trees
• e.g., the point quadtree and the k-d tree

11
Index Realization using SP-GiST

• Interface Parameters
• Key type Type of the data in the leaf-level of
  the tree,
• E.g., word, point
• Number of space partitions
• E.g., 4 for a quadtree, 26 for a trie, etc.
• Node predicate gives the predicate to use when
  navigating the SP-tree
• E.g., letter a, (x,y) inside node.rectangle
• Bucket size determines the maximum number of
  items that a leaf node can hold

12
Index Realization using SP-GiST

• Resolution determines maximum number of space
  decomposition, set based on the space and the
  granularity required

13
Interface Parameters for SP-GiST

• PathShrink NeverShrink, LeafShrink, TreeShrink

14
Interface Parameters for SP-GiST

• NodeShrink Determines if empty partitions should
  be kept in the tree or not
• E.g., tree vs. forest

15
Index Realization using SP-GiST

• External Methods
• Consistent a boolean function to guide the
  search in the tree
• PickSplit defines the way for splitting nodes in
  the tree
• Cluster defines how tree nodes are clustered
  into disk pages (SP-GiST provides a default
  clustering algorithm)

16
SP-GiST Internal Methods

• Internal methods are provided by the SP-GiST
  framework and does not need recoding
• Search Traverses the tree using the Consistent
  external method
• Insert Delete Builds the tree using the
  PickSplit and Cluster external methods
• Bulk Load Uses a general bulk loading algorithm
  ICDE 2004, Direct Buffering Bulk Loading
  (DBDL), that can bulk load any SP-tree
• Bulk Insert Uses a general bulk insertion
  algorithm ICDE 2004, Buffer Tree Bulk Insertion
  (BTBI) that can bulk insert a group of objects in
  any SP-tree

17
SP-GiST Details

• More detail
• Ilyas and Aref, JIIS 2001, SSDBM 2001
• www.cs.purdue.edu/faculty/aref.html