Horton
Menu

Re: Order changes in PG16 since ICU introduction

Robert Haas <robertmhaas@gmail.com>

View thread
From: Robert Haas <robertmhaas@gmail.com>
To: Jeff Davis <pgsql@j-davis.com>
Cc: Tom Lane <tgl@sss.pgh.pa.us>, Regina Obe <lr@pcorp.us>, Peter Eisentraut <peter.eisentraut@enterprisedb.com>, Sandro Santilli <strk@kbt.io>, pgsql-hackers@lists.postgresql.org
Date: 2023-04-21T20:33:48Z
Lists: pgsql-hackers 
On Fri, Apr 21, 2023 at 3:25 PM Jeff Davis <pgsql@j-davis.com> wrote:
> I am also having second thoughts about accepting "C" or "POSIX" as an
> ICU locale and transforming it to "en-US-u-va-posix" in v16. It's not
> terribly useful (why not just use memcmp()?), it's not fast in my
> measurements (en-US is faster), so maybe it's better to just throw an
> error and tell the user to use C (or provider=none as I suggest
> above)?

I mean, to renew a complaint I've made previously, how the heck is
anyone supposed to understand what's going on here?

We have no meaningful documentation of how to select an ICU locale
that works for you. We have a couple of examples and a suggestion that
you should use BCP 47. But when I asked before for documentation
references, the ones you provided were not clear, basically
incomprehensible. In follow-up discussion, you admitted you'd had to
consult the source code to figure certain things out.

And the fact that "C" or "POSIX" gets transformed into
"en-US-u-va-posix" is also completely documented. That string appears
twice in the code, but zero times in the documentation. There's code
to do it, but users shouldn't have to read code, and it wouldn't help
much if they did, because the code comments don't really explain the
rationale behind this choice either.

I find the fact that people are having trouble here completely
predictable. Of course if people ask for "C" and the system tells them
that it's using "en-US-u-va-posix" instead they're going to be
confused and ask questions, exactly as is happening here. glibc
collations aren't particularly well-documented either, but people have
some experience with, and they can get a list of values that have a
chance of working from /usr/share/locale, and they know what "C"
means. Nobody knows what "en-US-u-va-posix" is. It's not even
Googleable, really, whereas "C locale" is.

My opinion is that the switch to using ICU by default is ill-advised
and should be reverted. The compatibility break isn't worth whatever
advantages ICU may have, the documentation to allow people to
transition to ICU with reasonable effort doesn't exist, and the fact
that within weeks of feature freeze people who know a lot about
PostgreSQL are struggling to get the behavior they want is a really
bad sign.

-- 
Robert Haas
EDB: http://www.enterprisedb.com

Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →

  1. Introduce "builtin" collation provider.

  2. ICU: do not convert locale 'C' to 'en-US-u-va-posix'.

  3. initdb: change default --locale-provider back to libc.

  4. CREATE DATABASE: make LOCALE apply to all collation providers.

  5. Remove read-only server settings lc_collate and lc_ctype

  6. ICU: use uloc_getDefault() for initdb.

  7. Doc fixes for commit 1e16af8ab5.

  8. Doc: fix some rendering problems in the PDF docs build.

  9. Doc improvements for language tags and custom ICU collations.

  10. ICU: check for U_STRING_NOT_TERMINATED_WARNING.

  11. Reduce icu_validation_level default to WARNING.

  12. Revert "ICU: do not convert locale 'C' to 'en-US-u-va-posix'."

  13. Fix initdb --no-locale.

  14. Avoid character classification in regex escape parsing.