Re: [HACKERS] CTE inlining

On Wed, May 03, 2017 at 02:33:05PM -0300, Alvaro Herrera wrote:
> David Fetter wrote:
> 
> > When we add a "temporary" GUC, we're taking on a gigantic burden.
> > Either we support it forever somehow, or we put it on a deprecation
> > schedule immediately and expect to be answering questions about it for
> > years after it's been removed.
> > 
> > -1 for the GUC.
> 
> Absolutely.
> 
> So ISTM we have three choices:
> 
> 1) we switch unmarked CTEs as inlineable by default in pg11.  What seems
> likely to happen for a user that upgrades to pg11 is that 5 out of 10
> CTE-using queries are going to become faster than with pg10, and they
> are going to be happy; 4 out of five are going to see no difference, but
> they didn't have to do anything about it; and the remaining query is
> going to become slower, either indistinguishably so (in which case they
> don't care and they remain happy because of the other improvements) or
> notably so, in which case they can easily figure where to add the
> MATERIALIZED option and regain the original performance.
> 
> 
> 2) unmarked CTEs continue to be an optimization barrier, but we add
> "WITH INLINED" so that they're inlineable.  Some users may wonder about
> it and waste a lot of time trying to figure out which CTEs to add it to.
> They see a benefit in half the queries, which makes them happy, but they
> are angry that they had to waste all that time on the other queries.
> 
> 
> 3) We don't do anything, because we all agree that GUCs are not
> suitable.  No progress.  No anger, but nobody is happy either.
> 

+1 for option 1. I just finished rewriting a well written CTE query to
avoid the optimization fence and get reasonable performance.

Regards,
Ken