Date: 2025-05-14
Time: 11:30–12:20
Room: Breakout Room 3
Level: Advanced
Planner is the only Postgres subsystem using inferential statistics to adapt its behavior. But there are many other subsystems that would benefit from feedback-directed adaptation.
Using vacuum as a case study, this talk explores how we could use statistics and control theory to tune vacuum's aggressiveness. Based on over a year of failed experiments, this talk will explain the challenges involved in designing such a system.
There are two major sub-projects: introducing generic summary data structures to the cumulative statistics subsystem and building a flexible probabilistic model. Each of these poses a flurry of challenges. Any solution must work in most cases, self correct when it doesn't work, and, when all else fails, be debuggable.
This talk will review the concepts in statistics and control theory required to develop the predictive models and the APIs to use them. Finally, it will ask the audience what we can do differently to move forward.