Re: superuser unable to modify settings of a system table

On Fri, Jun 4, 2010 at 4:53 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> Personally, I think it would be better to put some work into making
>> allow_system_table_mods a little less simple-minded. =A0Right now,
>> !allow_system_table_mods prohibits you from doing perfectly sensible
>> things (as in the OP's original example) yet still allows you to do
>> things that are totally nuts (like DELETE FROM pg_class, which causes
>> every subsequent connection attempt for that database to panic).
>> Perfection may be too much to ask for but I'd take "modest
>> improvement"...
>
> Nope, that is the wrong viewpoint entirely. =A0allow_system_table_mods
> is intended to prevent you from modifying the *structure* of the
> system catalogs, which is fairly critical because the backend C code
> tends to depend on that. =A0Modifying the *content* of the catalogs
> is another matter, and in fact we let any superuser do that without
> having set allow_system_table_mods. =A0There is no practical way to
> distinguish a benign catalog-content change from a disastrous one,
> so we don't try.
>
> It's possible that reloptions is a special case and we should treat it
> as being more nearly in the content than structure category. =A0Not sure.

The backend C code also depends on the critical system indexes being
present in the system catalogs, yet we still allow them to be deleted.
 Is there really a use case for users fiddling with pg_proc, pg_class,
etc. directly?

At any rate, I'd be happy to drop that part of the proposal.  It would
be a step forward just to permit (even without
allow_system_table_mods) those changes which don't alter the structure
of the catalog.  For ALTER TABLE, the SET STATISTICS, (RE)SET
(attribute_option), SET STORAGE, CLUSTER ON, SET WITHOUT CLUSTER, and
(RE)SET (reloptions) forms are all things that fall into this
category, I believe.

--=20
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company