The one-paragraph summary of the lecture
Go from a SQL query → logical plan → physical plan. That physical plan can be produced either by cost-based search or purely by rules. Along the way, rewrite expressions and flatten nested queries. Cost-model estimates come from summarizations collected from the real data (1:18:42, 1:19:09).
Pavlo's own caveat: this was a surface-level brain dump of what you must consider when building an optimizer, not a complete treatment (1:18:42).
Course-specific notes
- Project 3 requires touching the query optimizer, but it is rule-based, not cost-based (
1:18:42). - BusTub has no statistics infrastructure (
1:14:04). - 15-721 is not running next semester; instead there's a special topics course, 15-799, entirely on query optimization. Pavlo's claim: do well in it and you'll have no problem finding a job (
1:19:22,1:19:34).
Next lecture
Transactions and concurrency control — described as the second hardest part of database systems (query optimization being the first). The difficulty: reasoning about correctness when the order in which queries are submitted need not be the order in which they run (1:19:55).