Re: Order changes in PG16 since ICU introduction - Mailing list pgsql-hackers

On Fri, 2023-04-21 at 13:28 -0400, Tom Lane wrote:
> I am wondering however whether this doesn't mean that all our
> carefully
> coded fast paths for C locale just went down the drain.

The code still exists. You can test it by using the built-in collation
"C" which is correctly specified with collprovider=libc and
collcollate=C.

For my test dataset, ICU 72, glibc 2.35:

  -- ~07s
  explain analyze select t from a order by t collate "C";

  -- ~15s
  explain analyze select t from a order by t collate "en-US-x-icu";

  -- ~21s
  explain analyze select t from a order by t collate "en-US-u-va-posix-
x-icu";

  -- ~34s
  explain analyze select t from a order by t collate "en_US";

I believe the confusion in this thread comes from:

* The syntax of CREATE DATABASE (the same as v15 but still confusing)
* The fact that you need provider=libc to get memcmp() behavior (same
as v15 but still confusing)

Regards,
    Jeff Davis