Обсуждение: Query regarding function cleanup in extension upgrade path

Ayush Vatsa <ayushvatsa1810@gmail.com> writes:
> To ask the question let me give a hypothetical example:-
> Suppose we have an extension named xyz with version 1.0. It has
> xyz--1.0.sql and xyz.c file. I have declared a function named fun() in the
> xyz--1.0.sql file and its definition in the xyz.c file.
> Now I want to drop this function in the next upgrade i.e. xyz--1.0--1.1 so
> I will use DROP FUNCTION fun(); in it and remove the definition from the
> xyz.c file.
> Here my doubt is wouldn't xyz--1.0 complain about the missing definition of
> fun() and if yes how can I clean up my function definition in the xyz.c
> file?

Yeah, you can't really remove the C extern symbol ever.  You can
reduce the C function to a stub that just throws a not-supported
error, perhaps, but your users won't necessarily appreciate that.
It's usually better to make the shlib capable of supporting both
the 1.0 and 1.1 APIs, so that users aren't forced into updating
the extension's SQL declarations immediately.

If you look at the standard contrib modules, you'll find a number
of cases where there are backwards-compatibility functions that
just exist to support people who're still using an older version
of the extension's SQL declarations.  Those are likely to remain
there indefinitely.

            regards, tom lane

Ayush Vatsa <ayushvatsa1810@gmail.com> writes:
> Just two follow up queries regarding this -
> 1. Suppose I created a new version 1.1 in which I reduce the C function to
> throw an error then ship it, will users get the .c latest file immediately
> and their old function will throw error but they have to use ALTER
> EXTENSION xyz UPGRADE TO 1.1 to use the latest objects defined in 1.1.sql.
> Is this the correct understanding?

Yes, if you do it like that then once they install the new shlib their
function will be broken until they do ALTER EXTENSION UPGRADE.

> 2. While going through the contrib folder I find that in the regress test
> there are two .out files with respect to a single .sql file, example
> citext.out and citext_1.out wrt citext.sql. Why is it so? Even in git blame
> , I couldn't find much!

We use that when the expected test output is environment-dependent.
If the diff between the .out files isn't pretty self-explanatory,
you can try checking the git log for the "_1.out" file to see why it
was created.

            regards, tom lane

> On 14 Feb 2024, at 21:56, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Ayush Vatsa <ayushvatsa1810@gmail.com> writes:

>> 2. While going through the contrib folder I find that in the regress test
>> there are two .out files with respect to a single .sql file, example
>> citext.out and citext_1.out wrt citext.sql. Why is it so? Even in git blame
>> , I couldn't find much!
>
> We use that when the expected test output is environment-dependent.
> If the diff between the .out files isn't pretty self-explanatory,
> you can try checking the git log for the "_1.out" file to see why it
> was created.

Some further details on this may be gleaned from the pg_regress source code
where it checks for an alternate file (_1 through _9) in case the main expected
output file created a diff:


https://git.postgresql.org/gitweb/?p=postgresql.git;a=blob;f=src/test/regress/pg_regress.c;h=c894005dac06bc233896520c83c18e51af8ace9c;hb=HEAD#l1450

--
Daniel Gustafsson