Horton

Re: A performance issue with Memoize

David Rowley <dgrowleyml@gmail.com>

View thread View on postgr.es
From: David Rowley <dgrowleyml@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Richard Guo <guofenglinux@gmail.com>, PostgreSQL-development <pgsql-hackers@postgresql.org>
Date: 2024-01-26T21:02:51Z
Lists: pgsql-hackers 
On Sat, 27 Jan 2024 at 09:41, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> David Rowley <dgrowleyml@gmail.com> writes:
> > I've adjusted the comments to what you mentioned and also leaned out
> > the pretty expensive test case to something that'll run much faster
> > and pushed the result.
>
> drongo and fairywren are consistently failing the test case added
> by this commit.  I'm not quite sure why the behavior of Memoize
> would be platform-specific when we're dealing with integers,
> but ...

Maybe snprintf(buf, "%.*f", 0, 5.0 / 2.0); results in "3" on those
rather than "2"?

Looking at the code in fmtfloat(), we fallback on the built-in snprintf.

I can try changing the unique1 < 5 to unique1 < 4 to see that's more stable.

David

Commits

  1. Attempt to fix newly added Memoize regression test

  2. Compare varnullingrels too in assign_param_for_var().

  3. De-dupicate Memoize cache keys

  4. Improve NestLoopParam generation for lateral subqueries

  5. Avoid sharing PARAM_EXEC slots between different levels of NestLoop.