Re: Early WIP/PoC for inlining CTEs

On Fri, Jan 11, 2019 at 1:49 PM David Fetter <david(at)fetter(dot)org> wrote:
> I don't see those as the same thing even slightly. Functions are
> Turing complete, generally speaking, which means that unless we send
> along those descriptors, we're asking the planner to solve the Halting
> Problem.

So... your argument is that functions are Turing-complete, but the
queries which call those functions somehow aren't? Actually, there's
probably a decent argument that WITH RECURSIVE is Turing-complete even
without any fancy functions. See
https://wiki.postgresql.org/wiki/Turing_Machine_(with_recursive)

> > OK, I know that's a bit of a straw man -- you're talking about hints
> > within a query, not DDL. Still, I think our theory about not having
> > hints is that we should have the optimizer try to figure it out
> > instead of making the user specify the behavior that they want -- and
> > I think sometimes that's setting the bar at an impossible level.
>
> There is a worked example that's open source.
> https://github.com/ossc-db/pg_hint_plan
>
> Have we looked over it seriously for inclusion in PostgreSQL?

That really has very little to do with what's under discussion here,
unless you're proposing that the right strategy for determining
whether to materialize the CTE or not is to execute the query both
ways and then use that to construct the plan we use to execute the
query.

> When they're specifying it, are they specifying it globally, or
> per WITH clause, or...?

Per WITH clause. That's the proposal which is under discussion here,
not anything else. Did you read the thread?

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company