Re: Truncation of mapped catalogs (whether local or shared) leads to server crash

Hi Robert, Andres, Tom,

Thank you for sharing your thoughts.

On Tue, Jun 18, 2024 at 8:02 PM Andres Freund <andres(at)anarazel(dot)de> wrote:
>
> Hi,
>
> On 2024-06-18 19:58:26 +0530, Ashutosh Sharma wrote:
> > On Tue, Jun 18, 2024 at 7:50 PM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> > >
> > > Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> > > > On Tue, Jun 18, 2024 at 8:10 AM Ashutosh Sharma <ashu(dot)coek88(at)gmail(dot)com> wrote:
> > > >> Executing below commands:
> > > >> -- set allow_system_table_mods TO on;
> > > >> -- truncate table pg_type;
> > >
> > > > If the operation isn't allowed without turning on
> > > > allow_system_table_mods, that means that doing it is probably a bad
> > > > idea and will probably break stuff, as happened here.
> > >
> > > Nothing good can come of truncating *any* core system catalog --- what
> > > do you think you'll still be able to do afterwards?
> > >
> > > I think the assertion you noticed is there because the code path gets
> > > traversed during REINDEX, which is an operation we do support on
> > > system catalogs. I have zero interest in making truncate work
> > > on them.
> > >
> >
> > I agree with you on that point. How about considering a complete
> > restriction instead?
>
> What's the point? There are occasional cases where doing something dangerous
> is useful, for debugging or corruption recovery. If we flat out prohibit this
> we'll just need a allow_system_table_mods_but_for_real option.
>

This is specifically about truncation of system catalogs, and does not
refer to any other DML operations on system catalogs, which I see are
necessary for many extensions that directly update catalogs like
pg_proc and others. Additionally, according to the comments in
truncate_check_rel(), we permit truncation of the pg_largeobject
catalog specifically during pg_upgrade. So, afaiu truncation of any
catalogs other than this can be restricted.

--
With Regards,
Ashutosh Sharma.