[anarazel]

Where I think ML would be much better than what we (postgres) do, is iteratively improving selectivity estimation. In today's postgres there's zero feedback from noticing at runtime that the collected statistics lead to bad estimates. In a better world we'd use that knowledge to improve future selectivity estimates.

> In order to do that, they purposefully don't explore the true planning space...they might explore 3-10 alternative ways of executing, whereas there might be hundreds or thousands of ways to do the same thing.

FWIW, often postgres' planner explores many more plan shapes than that (although not as complete plans, different subproblems are compared on a cost basis).

> While Postgres has explicitly chosen to not implement planning pragmas to override planner behavior, it would be really cool if you could have multiple planners optimized for different types of workloads,

FWIW, it's fully customizable by extensions. There's a hook to take over planning, and that can still invoke postgres' normal planner if the query isn't applicable. Obviously that's not the same as actually providing pragmas.