Re: Q: documentation improvement re collation version mismatch

On Thu, Nov 10, 2022 at 11:47:01AM +0100, Karsten Hilbert wrote:
> Thanks, Julien, for your explanation.
> 
> > > regarding changed collation versions this
> > >
> > >     https://www.postgresql.org/docs/devel/sql-altercollation.html
> > >
> > > says:
> > >
> > >     The following query can be used to identify all
> > >     collations in the current database that need to be
> > >     refreshed and the objects that depend on them:
> > >
> > >     SELECT pg_describe_object(refclassid, refobjid, refobjsubid) AS "Collation",
> > >            pg_describe_object(classid, objid, objsubid) AS "Object"
> > >       FROM pg_depend d JOIN pg_collation c
> > >            ON refclassid = 'pg_collation'::regclass AND refobjid = c.oid
> > >       WHERE c.collversion <> pg_collation_actual_version(c.oid)
> > >       ORDER BY 1, 2;
> > >
> > > I feel the result of that query can be slightly surprising
> > > because it does not return (to my testing) any objects
> > > depending on the database default collation, nor the database
> > > itself (as per a collation version mismatch in pg_database).
> >
> > Indeed.  The default collation is "pinned", so we don't record any dependency
> > on it.
> 
> Indirectly we do, don't we ?  Or else
> 
> > >     WHERE
> > >         collprovider IN ('d', 'c')
> 
> would not make much sense, right ?

What I meant is that we don't insert record in pg_depend to track dependencies
on pinned object, including the default collation.  The collprovider here comes
from pg_index.indcollation which is a different thing.  It can indeed store the
default collation, but it's only a small step toward less false negative.

Try that query with e.g.

CREATE INDEX ON sometable ( (somecol > 'somevalue') );

or

CREATE INDEX ON sometable (someid) WHERE somecol > 'somevalue';

Both clearly can get corrupted if the underlying collation library changes the
result of somecol > 'somevalue', but wouldn't be detected by that query.  There
are likely a lot more cases that would be missed, you can refer to the
discussions from a couple years ago when we tried to properly track all index
collation dependencies.

> The comment above the query in the official documentation is rather assertive
> (even if may true to the letter) and may warrant some more cautionary
> wording ?   Added, perhaps, some variation of this:
> 
> > For now, the only safe way to go is either reindex everything, or everything
> > except some safe cases (non-partial indexes on plain-non-collatable datatypes
> > only).

I think the comment is very poorly worded, as it leads readers to believe that
objects with a pg_depend dependency on a collation are the only one that would
get corrupted in case of glibc/ICU upgrade.

I agree that there should be a big fat red warning saying something like
"reindex everything if there's any discrepancy between the recorded collation
version and the currently reported one unless you REALLY know what you're
doing."