Обсуждение: B-tree unique index duplicate key error happens only in SUSE 9.3

This bug happens in SUSE 9.3 on both Pentium 4 and AMD64, whether the
binaries are from  postgresql-8.0.1 RPMs on the SUSE 9.3 DVD or are
built from 8.0.3 source code. However this bug does NOT happen with a
Debian box (unstable) running 8.0.3 on an x86 (Athlon XP, whether
binary or built from source). The problem is Postgresql claims two
records has the same value for one string attribute that has a unique
constraint, while in fact the two string values are different. To  see
this bug, just do a restore from the pg_dump'ed file attached to this
email. Sample steps and error message follow:

-------- begin command ---------------
createdb -E utf8 pg_bug
psql pg_bug < pg_dup_key_bug.dump
NOTICE:  CREATE TABLE / UNIQUE will create implicit index
"gaocanusers_userid_key" for table "gaocanusers"
CREATE TABLE
ERROR:  duplicate key violates unique constraint "gaocanusers_userid_key"
CONTEXT:  COPY gaocanusers, line 2: "129406     ���ズ�
27324923@nowhere.com.kk f U\N      \N      \N      \N      --      f
    2002-09-12 00:00:00     \N      \\3031\\3..."

----------- end ------------

FYI, Works just fine on gentoo with the UTF8 and ICU patches.

... John

> This bug happens in SUSE 9.3 on both Pentium 4 and AMD64,=20
> whether the binaries are from  postgresql-8.0.1 RPMs on the=20
> SUSE 9.3 DVD or are built from 8.0.3 source code. However=20
> this bug does NOT happen with a Debian box (unstable) running=20
> 8.0.3 on an x86 (Athlon XP, whether binary or built from=20
> source). The problem is Postgresql claims two records has the=20
> same value for one string attribute that has a unique=20
> constraint, while in fact the two string values are=20
> different. To  see this bug, just do a restore from the=20
> pg_dump'ed file attached to this email. Sample steps and=20
> error message follow:

Zhenlei Cai <jpenguin@gmail.com> writes:
> This bug happens in SUSE 9.3 on both Pentium 4 and AMD64, whether the
> binaries are from  postgresql-8.0.1 RPMs on the SUSE 9.3 DVD or are
> built from 8.0.3 source code. However this bug does NOT happen with a
> Debian box (unstable) running 8.0.3 on an x86 (Athlon XP, whether
> binary or built from source). The problem is Postgresql claims two

What makes you think this is a Postgres bug, rather than a bug in the
locale definition you are using on the SUSE box?  Try feeding the two
strings in question to strcoll() and see what happens.

One way that you can get inconsistent results from strcoll() is if you
feed it strings that are invalid according to the character set encoding
that strcoll() thinks you are using, which is to say the encoding
implied by the current LC_CTYPE locale setting.  So it's possible that
the real problem is that you have Postgres' database encoding set to
something that's incompatible with the postmaster's LC_CTYPE locale.
(Try "show lc_ctype" to see what that is exactly.)

            regards, tom lane