Обсуждение: pg_dumpall and owner of the extension

On Tue, 2024-01-23 at 16:31 +0200, kaido vaikla wrote:
> pg_pumpall (13.4) changes owner of the extension.
>
> [Extension created by user A, dump restored by user B -> Extension objects belong to B]
>
> Manual says "The user who runs CREATE EXTENSION becomes the owner of the extension".
> So pg_dumpall is not aware  who ran "CREATE EXTENSION"?

The user who runs CREATE EXTENSION is the user who restores the dump, hence the
changed ownership of the extension objects.

There is no ALTER EXTENSION ... OWNER TO to change the owner to the original user,
which is probably what would be needed in this case.

This is arguably a bug.

Trawling the archives, I found this pertinent thread:
https://www.postgresql.org/message-id/flat/4E1F1072.3010300%40enterprisedb.com

I looked for the referenced discussion, and all I found was this:
https://www.postgresql.org/message-id/flat/21857.1297101968%40sss.pgh.pa.us#e2697a586e2aecbe175d3ce1b7f70bf2

That sounds to me like ALTER EXTENSION ... OWNER TO was originally planned,
but never implemented:

 "BTW, on trying this I notice that pg_dump's default approach to
  ownership doesn't work because of the lack of an ALTER EXTENSION
  OWNER TO command.  I'm going to go ahead and add extowner to the catalog
  anyway, because it's easy and I'm convinced we're going to want it
  later.  But I don't feel like writing ALTER EXTENSION OWNER TO right
  now, so pg_dump will continue its current behavior of creating the
  extension as the user running the script."

Obviously nobody has felt like implementing ALTER EXTENSION ... OWNER TO
in the dozen years that followed that statement...

This was probably not a real problem for anybody, because traditionally
you needed a superuser to run CREATE EXTENSION, and functions are executable
by everybody by default, so you normally don't notice a difference.

Now that we have trusted extensions that don't require a superuser to
create, things might have changed.  It would be interesting to know if you
have a real problem with the current behavior, which would be an argument
in favor of fixing the omission.

Yours,
Laurenz Albe

kaido vaikla <kaido.vaikla@gmail.com> writes:
> So pg_dumpall is not aware  who ran "CREATE EXTENSION"?

Yeah, pg_dump makes no effort to restore extensions with the same
owner.  This didn't matter back when the feature was designed,
because you had to be superuser to install most interesting
extensions.  I suppose we should revisit the choice sometime.
There's no ALTER EXTENSION OWNER command, though, so it would
need to be done in a different way from other object types.
Not sure offhand what the downsides of that would be.

            regards, tom lane

kaido vaikla <kaido.vaikla@gmail.com> writes:
> Yes, i have a real problem. To keep things clear in database, my design is:
> - every application has own schema
> - every schema has owner who is not a user postgres
> - only schema owner can do DDL's on schema

> So if some application needs some extensions, then i give temporary
> suppersuser privilege to schema owner,
> if "create extension"  needs it, extensions are installed "with schema"
>  And revoke suppersuser privilege after extension inatall.
> I'm not sure, is it my design against postgres concept or not :(.

Well, it's certainly creating a problem for pg_dump: the alleged owner
of the extension doesn't have enough privilege to install it.  The
easiest way to make pg_dump support this would be to have it issue
something like

SET ROLE extension_owner;
CREATE EXTENSION foo;
RESET ROLE;

but that would fail for you.

            regards, tom lane

On Tue, 2024-01-23 at 18:01 +0200, kaido vaikla wrote:
> Yes, i have a real problem. To keep things clear in database, my design is:
> - every application has own schema
> - every schema has owner who is not a user postgres
> - only schema owner can do DDL's on schema
>
> So if some application needs some extensions, then i give temporary suppersuser privilege to schema owner, 
> if "create extension"  needs it, extensions are installed "with schema"  And revoke suppersuser privilege after
extensioninatall.  
> I'm not sure, is it my design against postgres concept or not :(.
>
> But in same time, i expect that pg_dumpall will not change ownership or at least gives some error or something.

I think your expectation is founded, and the current behavior is surprising.

At first, I thought that would be a security problem in combination with
trusted extensions, but then the objects from a trusted extension always
belong to the bootstrap superuser.
So, all extension objects always belong to a superuser, and a change in
ownership to a different after a restore won't be a privilege escalation.

In practice, it doesn't matter much to which superuser extension objects
belong.

Your usage pattern sort of breaks that.  I think that your practice of
temporarily elevating the prospective extension owner to superuser is
a questionable one.  You end up having extension objects owned by a
non-superuser.  That probably won't make much difference in most cases,
but it could matter for SECURITY DEFINER functions or views that use
restricted objects.  And if the extension object owner changes after a
restore, the behavior could change.

I don't really see what you are expecting to gain from your design.
Instead of temporarily elevating the schema owner's privileges, why
don't you create the extension as superuser right away?
(As an aside: you need a superuser to turn sumebody into a superuser.
Don't tell me that your application is connecting as superuser...)

So while I think that the current behavior is less than ideal, I don't
see it as a massive problem that needs to be fixed.  What is it that
stops working for you after the restore?

Perhaps some documentation for the surprising behavior might be in place.

Yours,
Laurenz Albe

On Wed, 2024-01-24 at 19:04 +0200, kaido vaikla wrote:
> CREATE EXTENSION doc should be change like:
> "The user who runs CREATE EXTENSION becomes the owner of the extension until first export-import" ;)

Here is a patch that implements something like that.

Yours,
Laurenz Albe