Use POSIX collation in the Postgres registry

[theory]

Applies also to Yugabyte, and separately add for Cockroach. Using the POSIX collation ensures that index ordering never changes when the database is upgraded, which is a particular problem with glibc collations, but since POSIX/C collation is strictly byte-ordered, it should be fine. Of course, any of use of ORDER BY on such columns will return unexpected results when users are used to other locales, but since Sqitch itself only ever orders by timestamp, it should not be an issue in its own use.

Closes #763.

[sjstoelting]

As I would always have to reindex all indexes in a database with reindexdb, I don't see any advantage.
In addition outside of certain countries (US, CA, UK) email addresses with characters outside of the POSIX character set are possible and in fact used.

[theory]

As I would always have to reindex all indexes in a database with reindexdb, I don't see any advantage.

That's what I was wondering about.

In addition outside of certain countries (US, CA, UK) email addresses with characters outside of the POSIX character set are possible and in fact used.

I believe you can still use UTF-8 bytes, it's just that things are indexed purely in byte order, not a character encoding order. I think that's right, but I'm not positive, admittedly.

[theory]

The more I think about it, the more convinced I am that this isn't the right thing to do. We shouldn't be making collation decisions for the people using the database. It might even be surprising that the registry tables use a different collation than the rest of the database. And as @sjstoelting points out, one ought to expect to reindex all indexes in a database where the collation has changed; and including Sqitch's indexes should be pretty obvious (and generally low-overhead, it's not usually much data).

What do you think, @datafoo?

[datafoo]

The more I think about it, the more convinced I am that this isn't the right thing to do. We shouldn't be making collation decisions for the people using the database. It might even be surprising that the registry tables use a different collation than the rest of the database.

Sqitch should use whatever collation is best for itself. As far as users are concerned, this is an implementation detail so, as I see it, Sqitch would not be "making decisions for the people using the database".

And as @sjstoelting points out, one ought to expect to reindex all indexes in a database where the collation has changed; and including Sqitch's indexes should be pretty obvious (and generally low-overhead, it's not usually much data).

It might be obvious for you because you knows Sqitch inside out but I can tell you it is was not for me. All I had was "Rebuild all objects affected by this collation" when faced with the error:

WARNING: collation "xx-x-icu" has version mismatch
DETAIL: The collation in the database was created using version 1.2.3.4, but the operating system provides version 2.3.4.5.
HINT: Rebuild all objects affected by this collation and run ALTER COLLATION pg_catalog."xx-x-icu" REFRESH VERSION, or build PostgreSQL with the right library version.

I think we should do what's right. PostgreSQL indicates:

For this reason use locales only if you actually need them.

Is there any reason to use locales?

[theory]

Is there any reason to use locales?

The thing is, we don't. We just default to whatever the DBA has configured things for. It might be more surprising to them if it wasn't the same.

[datafoo]

The thing is, we don't. We just default to whatever the DBA has configured things for.

Then I believe we should follow PostgreSQL recommendation.

It might be more surprising to them if it wasn't the same.

Again, this is an implementation detail so it should not impact the rest of the user's database. Can you think of anything that might be problematic and/or more difficult for a DBA if Sqitch were to use POSIX collation?

[sjstoelting]

And again, there's a world outside the USA which do need and use different collations. If you don't take the characters used in other countries seriously, I boycott your products wherever possible. Leave it to the users and don't make decisions, that don't fit elsewhere. And again, I even have email addresses with domains containing umlauts.

[theory]

Yeah, that's the thing, I think we have no business choosing a collation.

[datafoo]

And again, there's a world outside the USA which do need and use different collations. If you don't take the characters used in other countries seriously, I boycott your products wherever possible. Leave it to the users and don't make decisions, that don't fit elsewhere. And again, I even have email addresses with domains containing umlauts.

I believe there is a misunderstanding. In issue #763, I am suggesting for Sqitch to explicitly use a collation not a character set, not a character encoding.

If you run

initdb --auth=trust --pgdata="./mydata" --encoding=UTF-8 --no-locale --username="postgres"

... you end up with UTF8 encoding and C collation:

postgres=# \l
                                            List of databases
   Name    |  Owner   | Encoding | Collate | Ctype | ICU Locale | Locale Provider |   Access privileges   
-----------+----------+----------+---------+-------+------------+-----------------+-----------------------
 postgres  | postgres | UTF8     | C       | C     |            | libc            | 
 template0 | postgres | UTF8     | C       | C     |            | libc            | =c/postgres          +
           |          |          |         |       |            |                 | postgres=CTc/postgres
 template1 | postgres | UTF8     | C       | C     |            | libc            | =c/postgres          +
           |          |          |         |       |            |                 | postgres=CTc/postgres
(3 rows)

UTF8 being an encoding for unicode, you should be able to use any character you like.

[sjstoelting]

What is the goal about having data stored as UTF-8 and sorted by something that doesn't contain most of the characters that are available? You can change your own stuff to whatever you like, if you love POSIX, do it in your own stuff. But dictating that to everyone is still something I believe to be a decision by it's users. Am 15. August 2023 11:52:37 MESZ schrieb datafoo ***@***.***>:

…

> And again, there's a world outside the USA which do need and use different collations. If you don't take the characters used in other countries seriously, I boycott your products wherever possible. Leave it to the users and don't make decisions, that don't fit elsewhere. And again, I even have email addresses with domains containing umlauts. I believe there is a misunderstanding. In issue #763, I am suggesting for Sqitch to explicitly use a _collation_ not a _character set_, not a _character encoding_. If you run ```shell initdb --auth=trust --pgdata="./mydata" --encoding=UTF-8 --no-locale --username="postgres" ``` ... you end up with UTF8 encoding and C collation: ```shell postgres=# \l List of databases Name | Owner | Encoding | Collate | Ctype | ICU Locale | Locale Provider | Access privileges -----------+----------+----------+---------+-------+------------+-----------------+----------------------- postgres | postgres | UTF8 | C | C | | libc | template0 | postgres | UTF8 | C | C | | libc | =c/postgres + | | | | | | | postgres=CTc/postgres template1 | postgres | UTF8 | C | C | | libc | =c/postgres + | | | | | | | postgres=CTc/postgres (3 rows) ``` UTF8 being an encoding for unicode, you should be able to use any character you like. -- Reply to this email directly or view it on GitHub: #765 (comment) You are receiving this because you were mentioned. Message ID: ***@***.***>

[datafoo]

What is the goal about having data stored as UTF-8 and sorted by something that doesn't contain most of the characters that are available?

As a user of Sqitch, I do not interact with Sqitch using SQL, I interact with sqitch using the sqitch command. So, I do not care about how sqitch sort things as long as it is working for everyone.

You can change your own stuff to whatever you like, if you love POSIX, do it in your own stuff. But dictating that to everyone is still something I believe to be a decision by it's users.

I am not dictating anything to anyone, I am interacting on this issue to make sqitch even better than it already is. If you share this goal, you could perhaps try to explain how this proposal is creating a problem.

[theory]

I'm beginning to come back around to @datafoo's point of view, because using C/POSIX collation has no effect on the data you can store: it's all UTF-8. What it affects is the sort order:

david=# CREATE TABLE try (name TEXT COLLATE "POSIX", email TEXT COLLATE "POSIX");
CREATE TABLE
david=# INSERT INTO try VALUES('Stölting', 'stölting@example.com'), ('Stõlting', 'stõlting@example.com'), ('Stolting', 'stolting@example.com');
INSERT 0 3
david=# select * from try order by name;
   name   |        email         
----------+----------------------
 Stolting | stolting@example.com
 Stõlting | stõlting@example.com
 Stölting | stölting@example.com
(3 rows)

All the character encodings remain, although the sort ordering might not be expected for some locale or another: it's simply sorted in lexicographic order — that is, by the UTF-8 byte orders.