Обсуждение: Re: Add CASEFOLD() function.

Hi

2024年12月12日(木) 18:00 Jeff Davis <pgsql@j-davis.com>:
>
> Unicode case folding is a way to convert a string to a canonical case
> for the purpose of case-insensitive matching.
>
> Users have long used LOWER() for that purpose, but there are a few edge
> case problems:
>
> * Some characters have more than two cased forms, such as "Σ" (U+03A3),
> which can be lowercased as "σ" (U+03C3) or "ς" (U+03C2). The CASEFOLD()
> function converts all cased forms of the character to "σ".
>
> * The character "İ" (U+0130, capital I with dot) is lowercased to "i",
> which can be a problem in locales that don't expect that.
>
> * If new lower case characters are added to Unicode, the results of
> LOWER() may change.
>
> The CASEFOLD() function solves these problems.
>
> Patch attached.

I took a quick look at this as it sounds useful for the described issue,
and it seems to work as advertised, except the function is named "FOLDCASE()"
in the patch, so I'm wondering which is intended? A quick search indicates
there are no functions of either name in other databases; Python has a
"casefold()"
function [1] and PHP a "foldCase()" function [2], so it doesn't seem there's a
de-facto standard for this.

[1] https://docs.python.org/3/library/stdtypes.html#str.casefold
[2] https://www.php.net/manual/en/intlchar.foldcase.php

Regards


Ian Barwick

On Thu, 2024-12-12 at 21:52 +0900, Ian Lawrence Barwick wrote:
> and it seems to work as advertised, except the function is named
> "FOLDCASE()"
> in the patch, so I'm wondering which is intended?

Thank you for looking into this, I went back and forth on the name, and
mistyped it a few times.

ICU seems to use "foldcase":

https://unicode-org.github.io/icu-docs/apidoc/dev/icu4c/ucasemap_8h.html

and it seems to be slightly more proper grammatically (with "fold"
being the verb). However, "case folding" is common terminology in
Postgres and Unicode, so "casefold" can be seen as the verb instead.

I don't have a strong opinion here so I will just go with whatever
seems like the popular choice.

Regards,
    Jeff Davis

On 12/12/24 13:30, Jeff Davis wrote:
> On Thu, 2024-12-12 at 21:52 +0900, Ian Lawrence Barwick wrote:
>> and it seems to work as advertised, except the function is named
>> "FOLDCASE()"
>> in the patch, so I'm wondering which is intended?
> 
> Thank you for looking into this, I went back and forth on the name, and
> mistyped it a few times.
> 
> ICU seems to use "foldcase":
> 
> https://unicode-org.github.io/icu-docs/apidoc/dev/icu4c/ucasemap_8h.html
> 
> and it seems to be slightly more proper grammatically (with "fold"
> being the verb). However, "case folding" is common terminology in
> Postgres and Unicode, so "casefold" can be seen as the verb instead.
> 
> I don't have a strong opinion here so I will just go with whatever
> seems like the popular choice.


FWIW I prefer casefold()

-- 
Joe Conway
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com

On 12/16/24 12:49, Jeff Davis wrote:
> One question I have is whether we want this function to normalize the
> output.
> 
> I believe most usecases would want the output normalized, because
> normalization differences (e.g. "a" U+0061 followed by "combining
> acute" U+0301 vs "a with acute" U+00E1) are more minor than differences
> in case.
> 
> Of course, a user could wrap it with the normalize() function, but
> that's verbose and easy to forget. I'm also not sure that it can be
> made as fast as a combined function that does both.

Perhaps a one arg version that always casefolds and a two arg version 
that accepts nfc, nfd, none (or something similar)?

> And a follow-up question: if it does normalize, the second parameter
> would be the requested normal form. But to accept the keyword forms
> (NFC, NFD in gram.y) rather than the string forms ('NFC', 'NFD') then
> we'd need to also need to add CASEFOLD to gram.y (like NORMALIZE). Is
> that a reasonable thing to do?

SQL 2023 seems to include the NORMALIZE syntax, but the only case 
folding considered is UPPER and LOWER. As such, I think it ought to be a 
function but not part of the grammar.

-- 
Joe Conway
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com

On Mon, 2024-12-16 at 16:27 -0500, Joe Conway wrote:
>
> SQL 2023 seems to include the NORMALIZE syntax, but the only case
> folding considered is UPPER and LOWER. As such, I think it ought to
> be a
> function but not part of the grammar.

Should the standard support something like the Unicode idea of case
folding? If so, do we need to be careful of conflicts?

Regards,
    Jeff Davis

On 16.12.24 18:49, Jeff Davis wrote:
> One question I have is whether we want this function to normalize the
> output.
> 
> I believe most usecases would want the output normalized, because
> normalization differences (e.g. "a" U+0061 followed by "combining
> acute" U+0301 vs "a with acute" U+00E1) are more minor than differences
> in case.

Can you explain this in further detail?  I don't quite follow why this 
would be required.

> Of course, a user could wrap it with the normalize() function, but
> that's verbose and easy to forget. I'm also not sure that it can be
> made as fast as a combined function that does both.
> 
> And a follow-up question: if it does normalize, the second parameter
> would be the requested normal form. But to accept the keyword forms
> (NFC, NFD in gram.y) rather than the string forms ('NFC', 'NFD') then
> we'd need to also need to add CASEFOLD to gram.y (like NORMALIZE). Is
> that a reasonable thing to do?

That's maybe one reason to keep it separate.

Another might be that's not entirely clear how this should work in 
encodings other than UTF-8.  For example, the normalized string might 
not be representable in the encoding.

On Thu, 2024-12-19 at 17:18 +0100, Peter Eisentraut wrote:
> Can you explain this in further detail?  I don't quite follow why
> this
> would be required.

I am unsure now.

My initial reasoning was based on the idea that users would want to use
CASEFOLD(t) in a unique expression index as an improvement over
LOWER(t). And if you do that, you'd be surprised if some equivalent
strings ended up in the index. I don't think that's a huge problem,
because in other contexts we leave it up to the user to keep things
normalized consistently, and a CHECK(t IS NFC NORMALIZED) is a good way
to do that.

But there's a problem: full case folding doesn't preserve the normal
form, so even if the input is NFC normalized, the output might not be.
If we solve this problem, then we can just say that CASEFOLD()
preserves the normal form, consistently with how the spec defines
LOWER()/UPPER(), and I think that would be the best outcome.

I'm not sure if that problem is solvable, though, because what if the
input string is in both NFC and NFD, how do we know which normal form
to preserve?

We could tell users to use an expression index on
NORMALIZE(CASEFOLD(t)) instead, but that feels like inefficient
boilerplate.

>
> Another might be that's not entirely clear how this should work in
> encodings other than UTF-8.  For example, the normalized string might
> not be representable in the encoding.

That's a good point.

Regards,
    Jeff Davis

On Thu, 2024-12-19 at 09:51 -0800, Jeff Davis wrote:
> But there's a problem: full case folding doesn't preserve the normal
> form, so even if the input is NFC normalized, the output might not
> be.
> If we solve this problem, then we can just say that CASEFOLD()
> preserves the normal form, consistently with how the spec defines
> LOWER()/UPPER(), and I think that would be the best outcome.
>
> I'm not sure if that problem is solvable, though, because what if the
> input string is in both NFC and NFD, how do we know which normal form
> to preserve?

The options as I see it are:

1. Normalize the output (either by using an extra parameter or just
always normalizing to NFC). As you said, the problem here is that the
encoding might not work with normalization. One solution might be that
CASEFOLD() only works in UTF8, like NORMALIZE().

2. Try to preserve normalization as long as the encoding supports it.
The problem here is that we don't know what normal form to preserve,
because the input string might be in both NFC and NFD. We could
document that it preserves NFC form iff the input is NFC.

3. Allow CASEFOLD() to break the normal form of the input string. The
problem here is that the user may be surprised that the output is not
normalized even when all of their data is normalized. It's not clear to
me whether it still works for caseless matching -- it might if the
string is in a consistent form, even if not normalized.

Out of those I think #1 is the most appealing. Most users, and
especially users that care about these edge cases enough to use Full
Case Folding, are almost certainly going to be on UTF8 anyway. It's
also the most user-friendly.

Regards,
    Jeff Davis

Jeff Davis <pgsql@j-davis.com> writes:
> v6 attached. I plan to commit this soon.

The documentation for this function is giving the PDF docs build
indigestion:

[WARN] FOUserAgent - Glyph "?" (0x3a3, Sigma) not available in font "Courier".
[WARN] FOUserAgent - Glyph "?" (0x3c3, sigma) not available in font "Courier".
[WARN] FOUserAgent - Glyph "?" (0x3c2, sigma1) not available in font "Courier".

Found characters that cannot be output in the PDF document;  see README.non-ASCII

Not sure about a good workaround for this.  Are there any characters
within LATIN-1 that have interesting case-folding behavior?

            regards, tom lane

On Sat, 2025-01-25 at 00:00 -0500, Tom Lane wrote:
> Found characters that cannot be output in the PDF document;  see
> README.non-ASCII

Thank you, fixed.

> Not sure about a good workaround for this.  Are there any characters
> within LATIN-1 that have interesting case-folding behavior?

I just removed that example. There's already another example using ß
(U+00DF), though that only applies to PG_UNICODE_FAST (the new
collation that performs full case mapping and now full case folding).

Regards,
    Jeff Davis

One question I have is whether we want this function to normalize the
output.

Yes, we do.

I am sorry that I am so late to the party, but I am currently writing the Change Proposal for the SQL Standard for this function.

For <fold> (which includes LOWER() and UPPER()), the text says in Section 6.35 GR 7.e:

If the character set of <character factor> is UTF8, UTF16, or UTF32, then FR is replaced by
    Case:
        i) If the <search condition> S IS NORMALIZED evaluates to True, then NORMALIZE (FR)
        ii) Otherwise, FR.

Here, FR is the result of the function and S is its argument.

It does not appear to me that our LOWER and UPPER functions obey this rule, so there is a valid argument that we should continue to ignore it. Or, we can say that we have at least one of three compliant.

--

Vik Fearing

On Tue, 2025-06-17 at 17:37 +0200, Vik Fearing wrote:
> If the character set of <character factor> is UTF8, UTF16, or UTF32,
> then FR is replaced by
>      Case:
>          i) If the <search condition> S IS NORMALIZED evaluates to
> True, then NORMALIZE (FR)
>          ii) Otherwise, FR.

I read that as "if the input is normalized, then the output should be
normalized", IOW preserve the normalization. But does it mean "preserve
whatever the input normal form is" or "preserve NFC if the input is
NFC, otherwise the normalization is undefined"?

The above wording seems to mean "preserve NFC if the input is NFC",
because that's what NORMALIZE(FR) does when the normal form is
unspecified.

> It does not appear to me that our LOWER and UPPER functions obey this
> rule,

You are correct:

   WITH s(t) AS
   (SELECT NORMALIZE(U&'\00C1\00DF\0301' COLLATE "en-US-x-icu"))
   SELECT UPPER(t) = NORMALIZE(UPPER(t)) FROM s;
    ?column?
   ----------
    f

>  so there is a valid argument that we should continue to ignore it.
> Or, we can say that we have at least one of three compliant.

What do other databases do?

Given how costly normalization can be, imposing that on every caller
seems like a bit much. And favoring NFC for the user unconditionally
might not be the best thing. Then again, NFC is good most of the time,
and there are patches to speed up normalization.

I tend to think that a lot of users who want casefolding would also
want normalization, but it's hard to weigh that against the performance
cost. It might not matter outside of a few edge cases, though I'm not
sure exactly how many.

Regards,
    Jeff Davis

On 17/06/2025 20:14, Jeff Davis wrote:
> On Tue, 2025-06-17 at 17:37 +0200, Vik Fearing wrote:
>> If the character set of <character factor> is UTF8, UTF16, or UTF32,
>> then FR is replaced by
>>       Case:
>>           i) If the <search condition> S IS NORMALIZED evaluates to
>> True, then NORMALIZE (FR)
>>           ii) Otherwise, FR.
> I read that as "if the input is normalized, then the output should be
> normalized", IOW preserve the normalization. But does it mean "preserve
> whatever the input normal form is" or "preserve NFC if the input is
> NFC, otherwise the normalization is undefined"?
>
> The above wording seems to mean "preserve NFC if the input is NFC",
> because that's what NORMALIZE(FR) does when the normal form is
> unspecified.


Yes, and that is also the default for <normalized predicate>.


>> It does not appear to me that our LOWER and UPPER functions obey this
>> rule,
> You are correct:
>
>     WITH s(t) AS
>     (SELECT NORMALIZE(U&'\00C1\00DF\0301' COLLATE "en-US-x-icu"))
>     SELECT UPPER(t) = NORMALIZE(UPPER(t)) FROM s;
>      ?column?
>     ----------
>      f
>
>>   so there is a valid argument that we should continue to ignore it.
>> Or, we can say that we have at least one of three compliant.
> What do other databases do?


I don't know.  I am just pointing out what the Standard says.  I think 
we should either comply, or say that we don't do it for LOWER and UPPER 
so let's keep things implementation-consistent.


> Given how costly normalization can be, imposing that on every caller
> seems like a bit much.


How much does it cost to check for NFC?  I honestly don't know the 
answer to that question, but that is the only case where we need to 
maintain normalization.


> And favoring NFC for the user unconditionally
> might not be the best thing. Then again, NFC is good most of the time,
> and there are patches to speed up normalization.


It's not unconditionally, it's only if the input was NFC.


> I tend to think that a lot of users who want casefolding would also
> want normalization, but it's hard to weigh that against the performance
> cost. It might not matter outside of a few edge cases, though I'm not
> sure exactly how many.


I defer to you and others in the thread to make this decision.

-- 

Vik Fearing

On Thu, 2025-06-19 at 05:03 +0100, Thom Brown wrote:
> Late to the party, but is there an argument for porting this to the
> citext type? Or supplementing the extension with an additional type
> ("cftext"? *shrug*).

CASEFOLD() addresses a lot of the problems with using LOWER(), so that
sounds like a good idea. I'd be interested to hear from users of
citext.

Regards,
    Jeff Davis

On 17.06.25 17:37, Vik Fearing wrote:
> For <fold> (which includes LOWER() and UPPER()), the text says in 
> Section 6.35 GR 7.e:
> 
> 
> If the character set of <character factor> is UTF8, UTF16, or UTF32, 
> then FR is replaced by
>      Case:
>          i) If the <search condition> S IS NORMALIZED evaluates to True, 
> then NORMALIZE (FR)
>          ii) Otherwise, FR.
> 
> 
> Here, FR is the result of the function and S is its argument.
> 
> 
> It does not appear to me that our LOWER and UPPER functions obey this 
> rule, so there is a valid argument that we should continue to ignore it. 
> Or, we can say that we have at least one of three compliant.

The SQL standard also says in a few other places that normalization 
should be applied, and we do none of those, so this is probably not a 
reason to change CASEFOLD at this point.

On 19.06.25 06:03, Thom Brown wrote:
> Late to the party, but is there an argument for porting this to the 
> citext type? Or supplementing the extension with an additional type 
> ("cftext"? *shrug*). It currently uses lower(), so our current 
> recommendation for dealing with all unicode characters is to use 
> nondeterministic collations.

What is the motivation for wanting a citext variant instead of using 
nondeterministic collations?

On Thu, 19 Jun 2025 at 15:51, Peter Eisentraut <peter@eisentraut.org> wrote:
>
> On 19.06.25 06:03, Thom Brown wrote:
> > Late to the party, but is there an argument for porting this to the
> > citext type? Or supplementing the extension with an additional type
> > ("cftext"? *shrug*). It currently uses lower(), so our current
> > recommendation for dealing with all unicode characters is to use
> > nondeterministic collations.
>
> What is the motivation for wanting a citext variant instead of using
> nondeterministic collations?

Ease of use, perhaps. It seems easier to use:

column_name cftext

rather than:

CREATE COLLATION case_insensitive_collation (
    PROVIDER = icu,
    LOCALE = 'und-u-ks-level2',
    DETERMINISTIC = FALSE
);

column_name text COLLATE case_insensitive_collation

But I see the arguments against it. It creates an unnecessary
dependency on an extension, and if someone wants to ignore both case
and accents, they may resort to using 2 extensions (citext + unaccent)
when none are needed. I guess I don't feel strongly about it either
way.

Thom

On Thu, Jun 19, 2025 at 11:37 AM Thom Brown <thom@linux.com> wrote:
> On Thu, 19 Jun 2025 at 15:51, Peter Eisentraut <peter@eisentraut.org> wrote:
> > On 19.06.25 06:03, Thom Brown wrote:
> > > Late to the party, but is there an argument for porting this to the
> > > citext type? Or supplementing the extension with an additional type
> > > ("cftext"? *shrug*). It currently uses lower(), so our current
> > > recommendation for dealing with all unicode characters is to use
> > > nondeterministic collations.
> >
> > What is the motivation for wanting a citext variant instead of using
> > nondeterministic collations?
>
> Ease of use, perhaps. It seems easier to use:
>
> column_name cftext
>
> rather than:
>
> CREATE COLLATION case_insensitive_collation (
>     PROVIDER = icu,
>     LOCALE = 'und-u-ks-level2',
>     DETERMINISTIC = FALSE
> );
>
> column_name text COLLATE case_insensitive_collation
>
> But I see the arguments against it. It creates an unnecessary
> dependency on an extension, and if someone wants to ignore both case
> and accents, they may resort to using 2 extensions (citext + unaccent)
> when none are needed. I guess I don't feel strongly about it either
> way.

Don't forget, if you have a defined insensitive / normalized
collations, you also enable on-the-fly collation based matching, a la
"SELECT 'Å' = 'A' COLLATE ignore_accent_case;" regardless of the
provided collations (which I think is much more common certain in
other databases)

Robert Treat
https://xzilla.net

On 19/06/2025 16:47, Peter Eisentraut wrote:
> On 17.06.25 17:37, Vik Fearing wrote:
>> For <fold> (which includes LOWER() and UPPER()), the text says in 
>> Section 6.35 GR 7.e:
>>
>>
>> If the character set of <character factor> is UTF8, UTF16, or UTF32, 
>> then FR is replaced by
>>      Case:
>>          i) If the <search condition> S IS NORMALIZED evaluates to 
>> True, then NORMALIZE (FR)
>>          ii) Otherwise, FR.
>>
>>
>> Here, FR is the result of the function and S is its argument.
>>
>>
>> It does not appear to me that our LOWER and UPPER functions obey this 
>> rule, so there is a valid argument that we should continue to ignore 
>> it. Or, we can say that we have at least one of three compliant.
>
> The SQL standard also says in a few other places that normalization 
> should be applied, and we do none of those, so this is probably not a 
> reason to change CASEFOLD at this point.
>

Works for me.

-- 

Vik Fearing.

On Thu, 2025-06-19 at 18:21 +0200, Vik Fearing wrote:
> >
> > The SQL standard also says in a few other places that normalization
> > should be applied, and we do none of those, so this is probably not
> > a
> > reason to change CASEFOLD at this point.
> >
>
> Works for me.

Sounds good. We can document compatibility notes around this point.

If normalization becomes important, we can take the time to work out
the performance implications more carefully, and potentially introduce
an NCASEFOLD() if needed.

Regards,
    Jeff Davis

On Thu, 19 Jun 2025 at 18:39, David E. Wheeler <david@justatheory.com> wrote:
>
> On Jun 19, 2025, at 12:59, Thom Brown <thom@linux.com> wrote:
>
> > No. But given the options, I would personally choose nondeterministic collations now that they are available. I
justwish they were more user-friendly as I suspect the majority of people either won't know about them, or won't know
howto use them. 
>
> I suspect there are a lot of uses of citext for databases created before nondeterministic collations existed and
peopleare unaware of them or unclear on the migration path from one to the other, let alone implications for any
infrastructurethey built around cutest (like function signatures and return values). As long as citext conteinues to be
maintainedthere and there’s no super clear path to migrate, I’d bet good money that few would bother to switch. 

Maybe the citext doc page should explain how to get unhooked from it.
Something like:

ALTER TABLE mytable
  ALTER COLUMN ci_column
  SET DATA TYPE TEXT COLLATE case_insensitive_collation;

or

CREATE DOMAIN ci_text AS text
  COLLATE case_insensitive_collation;

ALTER TABLE mytable
  ALTER COLUMN ci_column
  SET DATA TYPE ci_text;

And because they're binary-compatible, they should also be free. No
doubt a procedure could do this to every instance in the database,
although I guess it gets trickier when it comes to functions that
accept citext as a parameter type, and other similar examples.

Thom