Unicode character default collation table

Question

I don't know which site this question belongs exactly, so posting it here.

I use Postgresql 9.2 on RHEL 6.4 and observe the following:

select foo 
from unnest('{а,ә,б,в,г,д,е,ж}'::text[]) as foo 
order by foo collate "kk_KZ.utf8"

gives

а
ә
б
в
г
д
е
ж

BUT

select foo 
from unnest('{а,ә,б,в,г,д,е,ж}'::text[]) as foo 
order by foo collate "en_US.utf8"

gives

а
б
в
г
д
е
ә -- misplaced
ж

Further, I found that there is the Default Unicode Collation Element Table [1], which lists the character in question (04D9 ; [.199D.0020.0002.04D9] # CYRILLIC SMALL LETTER SCHWA) in proper order.

I understand that it is silly to expect the cyrillic characters be handled properly by "en_US.utf8" locale, but what is the correct behavior by Unicode or any other relevant standards in cases, where a character does not normally belong to language/locale used for collation?

[1] http://www.unicode.org/Public/UCA/latest/allkeys.txt

Answer 1

Postgresql uses the locales provided by the operating system. In your setup, locales are provided by glibc. Glibc uses a heavily modified version of an "ancient" version of ISO 14651 (see glibc Bug 14095 - Review / update collation data from Unicode / ISO 14651 for information on current pains in trying to update glibc locale data).

As of glibc 2.28, to be released on 2018-08-01, glibc will use data from ISO 14651:2016 (which is synchronized to Unicode 9), and will give the order the OP expects for en_US.

ISO 14651 is Method for comparing character strings and description of the common template tailorable ordering and it is similar to the UCA, with some differences. The CTT (Common Template Table) is the ISO14651 equivalent of the DUCET, and they are aligned.

The first time CYRILLIC SMALL LETTER SCHWA appeared in a collation table in glibc was for the az_AZ locale (Azerbaijani), where it is ordered after CYRILLIC SMALL LETTER IE. This corresponds to:

commit fcababc4e18fee81940dab20f7c40b1e1fb67209
Author: Ulrich Drepper <drepper@redhat.com>
Date:   Fri Aug 3 08:42:28 2001 +0000

    Update.

    2001-08-03  Ulrich Drepper  <drepper@redhat.com>

        * locale/iso-639.def: Add Tigrinya.

From there, that ordering was eventually moved to the file iso14651_t1 as per Bug 672 - Include iso14651_t1 in collation rules, which was an effort to simplify glibc locale data. This corresponds to:

commit 5d2489928c0040d2a71dd0e63c801f2cf98e7efc
Author: Ulrich Drepper <drepper@redhat.com>
Date:   Sun Feb 18 04:34:28 2007 +0000

    [BZ #672]

    2005-01-16  Denis Barbier  <barbier@linuxfr.org>
        [BZ #672]
        * locales/ca_ES: Replace current collation rules by including
        iso14651_t1 and adding extra rules if needed.  There should be
        no noticeable changes in sorted text. only ligatures and
        ignoreable characters have modified weights.
        * locales/da_DK: Likewise.
        * locales/en_CA: Likewise.
        * locales/es_US: Likewise.
        * locales/fi_FI: Likewise.
        * locales/nb_NO: Likewise.

        [BZ #672]
        * locales/iso14651_t1: Simplified.  Extended.

Most locales in glibc start from iso14651_t1, and tailor it, which is what you are seeing with en_US.

While glibc based its default ordering in Azerbaijani, the DUCET instead bases it on the ordering for Kazakh and Tatar, which is where the difference comes from.

Answer 2

The Unicode Collation Algorithm allows any tailorings to be made to the DUCET.

There isn't a "correct" behaviour. There are various behaviours one could expect, and the most appropriate depends on the context, the audience. Sometimes any behaviour could be correct, since there isn't really a reason to force any order of cyrillic betters in an American English collation.

The Common Locale Data Repository provides locale-specific tailorings to the DUCET. The CLDR uses LDML (Locale Data Markup Language) to specify the tailorings, and the syntax is given by the Unicode Technical Specification #35, part 5.

The latest version of the data provided by the CLDR for en_US has no tailorings: it uses a modified version of the DUCET (as stated in UTS#35 under "Root collation"). It lists the cyrillic schwa after the cyrillic A, i.e., the order you were expecting.

There is also data for an en_US_POSIX locale, and that one includes some modifications, but none changes anything that isn't in ASCII.

It appears the en_US locale installed in your system uses a tailoring that puts the schwa next to E probably because of their similar form. It could be argued that would cause fewer surprises to an American English audience than sorting the schwa after A: ask people what that is and see how many will just tell you it is an "upside-down E". It isn't right or wrong, but if you ask me, it seems more appropriate than the collation found in the CLDR.

Answer 3

It's not misplaced. It might be to you, but it's not to me. :-) In all seriousness, there is no correct behavior by Unicode; there simply cannot be. A character set is a mapping; the collation is a locale-specific set of rules to sort the characters in that set -- and even within the same locale there can be multiple collations.

The ICU docs has colorful examples of how thorny this kind of stuff gets, in case you're curious. Quoting extensively:

http://userguide.icu-project.org/collation

[H]ere are some of the ways languages vary in ordering strings:

The letters A-Z can be sorted in a different order than in English. For example, in Lithuanian, "y" is sorted between "i" and "k".

Combinations of letters can be treated as if they were one letter. For example, in traditional Spanish "ch" is treated as a single letter, and sorted between "c" and "d".

Accented letters can be treated as minor variants of the unaccented letter. For example, "é" can be treated equivalent to "e".

Accented letters can be treated as distinct letters. For example, "Å" in Danish is treated as a separate letter that sorts just after "Z".

Unaccented letters that are considered distinct in one language can be indistinct in another. For example, the letters "v" and "w" are two different letters according to English. However, "v" and "w" are considered variant forms of the same letter in Swedish.

A letter can be treated as if it were two letters. For example, in traditional German "ä" is compared as if it were "ae".

Thai requires that the order of certain letters be reversed.

French requires that letters sorted with accents at the end of the string be sorted ahead of accents in the beginning of the string. For example, the word "côte" sorts before "coté" because the acute accent on the final "e" is more significant than the circumflex on the "o".

Sometimes lowercase letters sort before uppercase letters. The reverse is required in other situations. For example, lowercase letters are usually sorted before uppercase letters in English. Latvian letters are the exact opposite.

Even in the same language, different applications might require different sorting orders. For example, in German dictionaries, "öf" would come before "of". In phone books the situation is the exact opposite.

Sorting orders can change over time due to government regulations or new characters/scripts in Unicode.