Cost Models DO Matter

1
Cost Models DO Matter

• Providing Cost information for Diverse Data
  Sources in A Federated System

2
Introduction

• Optimization cross-source queries in Federated
  Systems.
• Optimizer must consider
• Capabilities of the data sources.
• Costs of operations performed by the sources.
• Contribution of the paper
• Addresses how wrappers supply cost and
  cardinalities information to the optimizer.
• Proposes a Framework to ease the above task.

3
Review-Costing in Traditional Databases

• Cost of a query plan cost of the topmost
  operator (POP) in the query plan tree.
• Cardinality of a query result cardinality of the
  topmost operator.
• Costs associated with each POP
• Total cost (in sec.)
• Re-execution cost (in sec.)
• Cardinality
• Computation of these costs Use of Cost Formulas
• Cost Formulas variables are instantiated using
  statistics.

4
Flow of Information
5
Costing Query Plans in Federated Systems

• Cost-based optimizer needs cooperation from
  wrappers to cost plans.
• Creating a Framework to do so needs to
• Define a cost model that tells what information
  must a wrapper supply to the optimizer.
• Find cost formulas to compute the above
  information.
• Gather statistics from the wrapper data source to
  instantiate the cost formulas.

6
Cost Model

• Integration of the costs of work done by a remote
  data source into the cost of a query plan.
• Each POP (plan operator) executed by those data
  sources is represented as PUSHDOWN POP in the
  query plan tree.
• Wrappers supply total cost , re-execution cost
  and result cardinality of the POP to the
  optimizer.

7
Cost Formulas

• Needed from Wrappers to compute their plan costs.
• Cost formulas that captures the access costs to
  retrieve the attributes of the sources objects.
• Cost formulas to measure method costs total
  method cost, re-execution method cost.

8
Statistics

• Used as inputs to the cost formulas.
• Wrapper requires logical statistics about the
  data not physical characteristics.
• Base cardinalities.
• Attributes (uniform assumption / average length)
• Gathering statistics by executing a workload of
  queries against the data managed by the wrapper.

9
Adaptations of the Framework
10
Results

• Cost-based optimization vs. push down as much
  work as possible to the data sources.
• Query Family of joins (from equi-joins to cross
  products).
• In that family a Cross-Over point exists at which
  it no longer make sense to put the joins down.
• Main Result Wrapper Input is crucial for the
  optimizer to identify the Cross Over point and
  thus to choose good plans.

11
Schematically
12
Results on Adaptability of the Framework
13
Summary

• Results shows the need for cost-based
  optimization in Federated Systems.
• A Framework for this optimization was proposed.

14
(No Transcript)