On Sat, Jun 14, 2025 at 11:20:31AM -0400, Tom Lane wrote:
> Tomas Vondra <tomas@vondra.me> writes:
>> The OIDs for user-defined objects (e.g. those from extensions) are not
>> stable, and this will not change. The only way is to prevent the test
>> output, e.g. by not including OIDs in the results, and eliminating all
>> other types of non-determinism - eg. by enforcing ordering, etc.
>
> You could also make your tests print database object OIDs in
> symbolic form, e.g. cast function OIDs to regprocedure, use
> pg_describe_object(), etc. Then it doesn't matter if the
> underlying numbers change.
Definitely never expose an OID for something that users should be able
to trigger by themselves in the output. First, it is usually not
helpful because only the OID does not tell one which object is wrong.
When it comes to check the state of the OIDs in the catalogs, or when
it comes to manipulate the generation of some commands, another trick
that we are relying on quite a bit in the regression tests is to save
an OID in a \gset command, then reuse it in a query for equality
comparison. For utility commands generated, it is also possible
(well, kind of), to generate command names and pass them to a PL/pgsql
function triggered by an EXECUTE. This would not work for commands
like VACUUM or REINDEX commands on toast relations, because these
cannot run inside a transaction context, still that can be really
useful.
A last thing I've been doing in this area in the last few years which
is a interesting trick is the isolation test
reindex-concurrently-toast.spec, that checks a case with TOAST
relations and concurrent REINDEX CONCURRENTLY: I've been abusing of
allow_system_table_mods to rename the TOAST relations to a fixed name,
then issue REINDEX commands on fixed names. In this case, the test
case is able to become deterministic even if the OIDs vary in time.
--
Michael
