Designing a Cost Model - Automated Physical Database Design and Tuning - page 96

5.5 Summary

•

Materializing a configuration to obtain the estimated cost of a query is
not scalable.

• What-if calls allow optimizing queries under arbitrary configurations by

simulating the metadata of hypothetical indexes along with supporting
statistical information.
• What-if optimization calls can be avoided in some scenarios by inferring
costs from previous optimizations of atomic configurations.
• INUM and CPQO are techniques that can reoptimize queries under differ-
ent configurations much more eciently than by doing separate what-if
optimization calls.

5.6 Additional Reading
After initial work on parametric cost models, a new line of work used the opti-
mizer cost model itself to evaluate the goodness of a configuration. 6 The what-
if optimization concept was later extended for the case of complex engines and
multicolumn indexes, acknowledging the diculties involved in materializing
statistics, 5 whose creation cost can be reduced by using sampling. 3 Other per-
formance improvements include the notion of atomic configurations, 6 which
can be computed based on assumptions of the query engine 4 or by leveraging
index benefit graphs. 7 , 9 More advanced strategies like INUM 8 and CPQO 2 further
reduce the cost to perform what-if optimization calls.

References
1. Nicolas Bruno and Surajit Chaudhuri. Automatic physical database tun-
ing: A relaxation-based approach. In Proceedings of the ACM Interna-
tional Conference on Management of Data (SIGMOD) , 2005.
2. Nicolas Bruno and Rimma Nehme. Configuration-parametric query op-
timization for physical design tuning. In Proceedings of the ACM Inter-
national Conference on Management of Data (SIGMOD) , 2008.
3. Surajit Chaudhuri, Rajeev Motwani, and Vivek Narasayya. Ran-
dom sampling for histogram construction: How much is enough? In