Message-ID: <34c910bf-50bf-c33f-940d-3e5fe029347e@enterprisedb.com>
Date: Tue, 11 Jan 2022 12:18:58 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:91.0)
 Gecko/20100101 Thunderbird/91.4.1
Subject: Re: ICU for global collation
Content-Language: en-US
To: Daniel Verite <daniel@manitou-mail.org>
Cc: pgsql-hackers <pgsql-hackers@postgresql.org>
References: <809c0fba-b0e6-44bf-b110-0886cb462aab@manitou-mail.org>
From: Peter Eisentraut <peter.eisentraut@enterprisedb.com>
In-Reply-To: <809c0fba-b0e6-44bf-b110-0886cb462aab@manitou-mail.org>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Archived-At: 
 <https://www.postgresql.org/message-id/34c910bf-50bf-c33f-940d-3e5fe029347e%40enterprisedb.com>
Precedence: bulk

On 10.01.22 12:49, Daniel Verite wrote:
> With the current patch, it's not possible, AFAICS, because the user
> can't tell that the collation is non-deterministic. Presumably this
> would require another option to CREATE DATABASE and another
> column to store that bit of information.

Adding this would be easy, but since pattern matching currently does not 
support nondeterministic collations, if you make a global collation 
nondeterministic, a lot of system views, psql, pg_dump queries etc. 
would break, so it's not practical.  I view this is an orthogonal 
project.  Once we can support this without breaking system views etc., 
then it's easy to enable with a new column in pg_database.

> The "daticucol" column also suggests that we don't expect to add
> other collation providers in the future. Maybe a pair of columns like
> (datcollprovider, datcolllocale) would be more future-proof,
> or a (datcollprovider, datcolllocale, datcollisdeterministic)
> triplet if non-deterministic collations are allowed.

I don't expect many new collation providers.  So I don't think an 
EAV-like storage would be helpful.  The other problem is that we don't 
know what we need.  For example, the libc provider needs both a collate 
and a ctype value, so that wouldn't fit into that scheme nicely.

> Also, pg_collation has "collversion" to detect a mismatch between
> the ICU runtime and existing indexes. I don't see that field
> for the db-wide ICU collation, so maybe we currently miss the capability
> to detect the mismatch for the db-wide collation?

Yeah, I think I need to add a datcollversion field and the associated 
checks.