2 Old-Fashioned Problems

1
2 Old-Fashioned Problems

• Surajit Chaudhuri
• Microsoft

2
Query Optimization Revisited

• Great engineering, poorly understood
• What makes our search algorithm unique?
• Is our Cardinality Estimation consistent?
• What are the limits of effective query
  optimization?
• Separate first order and second order
  assumptions..

3
Need for Evaluation Frameworks

• DB community working on too many exciting
  problems
• Hard to know if we are making progress
• Isnt this true of Query Optimizers as well?
• Correctness of transformation can be verified
• Impact on real workload/search efficiency
  unclear
• Automated Physical design
• A Technical definition of optimality/goodnes
  s exists
• - DB2 and Oracle sucks, but I cannot convince
  you?

4
Need for Evaluation Frameworks (2)

• Example Data Integration (Data Cleaning/Schema
  Matching)
• Fuzzy problem specification
• No notion of goodness except for user
  validation
• Call to Action
• Resist temptation to define another new problem
  with overlapping definition
• Separate heuristic/knobs from a concrete
  server/middleware API
• Define evaluation infrastructure (TREC like?)
• Otherwise
• We will be such like the AI folks..