Re: Improving executor performance

Hi,

On 2016-11-04 15:25:58 -0700, Doug Doole wrote:
> On 07/13/2016 06:18 PM, Andres Freund wrote:
> > Attached (in patch 0003) is a proof-of-concept implementing an
> > expression evalution framework that doesn't use recursion. Instead
> > ExecInitExpr2 computes a number of 'steps' necessary to compute an
> > expression. These steps are stored in a linear array, and executed one
> > after another (save boolean expressions, which can jump to later steps).
> > E.g. to compute colname = 1 the steps are 1) fetch var, 2) evaluate
> > const, 3) call function.
>
> We've been having trouble with the performance of simple expressions in
> PLpgSQL so I started playing with this patch. (No sense re-inventing the
> wheel after all.) It was straightforward to extend to simple expressions and
> showed an immediate improvement (~10% faster on a simple test).

That's cool.

> Running in our full environment highlighted a few areas that I think
> are worth more investigation.

There indeed should be many of those.

> However, before I tackle that, is the posted proof-of-concept still the
> latest and greatest? If not, any chance of getting the latest?

It's definitely *not* the latest and greatest. My latest version
actually passes all regression tests, and has a few additional
expression types handled natively (NULL tests, CASE IIRC), and more than
few bugs fixed.

I'm away at the moment, but I plan to post a new version end of this or
beginning of next week.

> Going forward, I'd like to collaborate on our efforts if you're interested.

That sounds good!

Regards,

Andres