Chengkai Li Kevin-Chen-Chuan Chang Ihab Ilyas Sumin Song

1
Chengkai Li Kevin-Chen-Chuan Chang
Ihab Ilyas Sumin Song
RankSQL Query Algebra and Optimization
for Relational Top- k Queries

• Presented by
  Mariam John
• CSE 6392
• 03/20/2006

2
Contents

• Introduction
• RankSQL
• Ranking Query Model
• Rank-Relational Algebra
• Ranking Query PlansExecution Model
• Conclusion

3
Introduction

• Top-k queries provides only the top k query
  results according to a user-specified ranking
  function.
• Most of the available solutions are in the
  middleware, or focus on specific operators and
  queries.
• Top-k queries are not treated as first class
  query type in RDBMS. Relational algebra has no
  notion for ranking.

4
RankSQL

• Provides seamless support and integration of
  top-k queries with the existing SQL query
  facility in RDBMS.
• Supports ranking as a first-class database
  construct.
• Extends relational algebra and query
  optimization.

5
Example of a Top-k Query

• SELECT
• FROM Hotel h, Restaurant r, Museum m
• WHERE c1 AND c2 AND c3
• ORDER BY p1p2p3
• LIMIT k
• c1 r.cuisineItalian p1
  cheap(h.price)
• c2 h.pricer.pricelt100 p2
  close(h.addr,r.addr)
• c3 r.aream.area p3
  related(m.collection,
• 
  dinosaur)

6
Rank Query Model
Ranking
Filtering

• Rank relational query has 4 types of predicates
• Filtering Boolean-selection
  predicates
• Boolean-join
  predicates
• Ranking rank-selection predicates
• rank-join predicates
• Goal is to support rank relational queries
  efficiently.

7
Rank-Relational Query

• Such queries add a ranking dimension to query
  processing and optimization.
• Filtering restricts tuple membership by
  applying a Boolean function of Boolean selection
  or join predicates.
• Ranking restricts order by applying a monotonic
  scoring function of ranking predicates.

8
Ranking as First-Class Construct

• Support for ranking as a first class construct in
  RDBMS is lacking.
• Relational algebra models Boolean filtering as a
  first class construct in query processing.
• c1 is a selection over R, and
• c2 is a join condition over R S

9
Filtering as a First-Class Construct

• Algebra framework supports the following for
  Boolean filtering
• - splitting
• - interleaving
• Enable query optimization to transform from
  canonical form to efficient query plans.

10
Ranking as First-Class Construct

• Algebraic support for optimization is lacking for
  ranking.
• The sorting operator is monolithic.
• It may be beneficial to evaluate ranking
  predicates one by one and interleave them with
  Boolean filtering.

11
Challenges

• First, we must extend relational algebra to do
  the following
• Handle ranking
• Define algebraic laws to handle equivalence
  transformation
• Second, we need to generalize query optimization
  techniques to integrate the parallel dimensions
  of Boolean filtering and ranking.

12
Rank-Relational Algebra

• Rank-Relation is a relation with its tuples
  scored and ordered accordingly
• How do we rank a relation, given

13
Ranking principle

• Maximum possible score of a tuple t, denoted by
  , is defined as
• if
• 1 otherwise

14
Examples of Rank-Relations
15
Operators

• Need to extend relational-algebra operators for
  manipulating rank-relations.
• For supporting ranking as a first-class
  construct, define a new operator µ.
• This new rank operator should satisfy the two
  requirements splitting and interleaving.

16
New Operator, µ

• Extend relational algebra by adding a new rank
  operator, µ. What does mean?
• Extend the original semantics of existing
  operators with rank-awareness, enabling
  interaction with the new rank operator.
• Extend relational algebra such that it gives
  several equivalences relevant to ranking.

17
Results of Operators
18
Ranking Query Plans Execution Model

• Extend the common execution model to handle rank
  query.
• Operators incrementally output rank relations.
• Query has an explicitly requested result size.
• Key capability of a rank-aware operator is to
  decide if enough information has been obtained
  from its input tuples in order to incrementally
  produce the next ranked output tuple.

19
Example
20
Conclusion

• RankSQL is a system that provides a systematic
  framework to support efficient evaluation of
  top-k queries in RDBMS.
• Extend relational algebra to make ranking a
  first-class construct.
• Query execution model is extended to handle
  ranking query.
• Rank-aware operators are selective and
  context-sensitive.