From: Tom Lane <tgl@sss.pgh.pa.us>
To: Peter Eisentraut <peter.eisentraut@2ndquadrant.com>
cc: Marco Atzeri <marco.atzeri@gmail.com>,
        PostgreSQL mailing lists <pgsql-hackers@postgresql.org>
Subject: Re: initdb initalization failure for collation "ja_JP"
In-reply-to: <28708.1497797324@sss.pgh.pa.us>
References: <20c74bc3-d6ca-243d-1bbc-12f17fa4fe9a@gmail.com>
 <28708.1497797324@sss.pgh.pa.us>
Comments: In-reply-to Tom Lane <tgl@sss.pgh.pa.us>
	message dated "Sun, 18 Jun 2017 10:48:44 -0400"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <9153.1497973022.1@sss.pgh.pa.us>
Date: Tue, 20 Jun 2017 11:37:02 -0400
Message-ID: <9154.1497973022@sss.pgh.pa.us>
Precedence: bulk
Sender: pgsql-hackers-owner@postgresql.org

I wrote:
> Marco Atzeri <marco.atzeri@gmail.com> writes:
>> Building on Cygwin latest 10  beta1 or head sourece,
>> make check fails as:
>> ...
>> performing post-bootstrap initialization ... 2017-05-31 23:23:22.214 
>> CEST [16860] FATAL:  collation "ja_JP" for encoding "EUC_JP" already exists

> Hmph.  Could we see the results of "locale -a | grep ja_JP" ?

Despite the lack of followup from the OP, I'm pretty troubled by this
report.  It shows that the reimplementation of OS collation data import
as pg_import_system_collations() is a whole lot more fragile than the
original coding.  We have never before trusted "locale -a" to not produce
duplicate outputs, not since the very beginning in 414c5a2e.  AFAICS,
the current coding has also lost the protections we added very shortly
after that in 853c1750f; and it has also lost the admittedly rather
arbitrary, but at least deterministic, preference order for conflicting
short aliases that was in the original initdb code.

I suppose the idea was to see whether we actually needed those defenses,
but since we have here a failure report after less than a month of beta,
it seems clear to me that we do.  I think we need to upgrade
pg_import_system_collations to have all the same logic that was there
before.

Now the hard part of that is that because pg_import_system_collations
isn't using a temporary staging table, but is just inserting directly
into pg_collation, there isn't any way for it to eliminate duplicates
unless it uses if_not_exists behavior all the time.  So there seem to
be two ways to proceed:

1. Drop pg_import_system_collations' if_not_exists argument and just
define it as adding any collations not already known in pg_collation.

2. Significantly rewrite it so that it de-dups the collation set by
hand before trying to insert into pg_collation.

#2 seems like a lot more work, but on the other hand, we might need
most of that logic anyway to get back deterministic alias handling.
However, since I cannot see any real-world use case at all for
if_not_exists = false, I figure we might as well do #1 and take
whatever simplification we can get that way.

I'm willing to do the legwork on this, but before I start, does
anyone have any ideas or objections?

			regards, tom lane


-- 
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers