Re: ALTER TABLE lock strength reduction patch is unsafe Reply-To:

On Fri, Mar 21, 2014 at 06:53:27PM +0000, Simon Riggs wrote:
> On 21 March 2014 17:49, Noah Misch <noah@leadboat.com> wrote:
> 
> >> >   alter table information_schema.triggers set (security_barrier = true);
> >>
> >> I find it hard to justify why we accept such a statement. Surely its a
> >> bug when the named table turns out to be a view? Presumably ALTER
> >> SEQUENCE and ALTER <other stuff> has checks for the correct object
> >> type? OMG.
> >
> > We've framed ALTER TABLE's relkind leniency as a historic artifact.  As a move
> > toward stricter checks, ALTER TABLE refused to operate on foreign tables in
> > 9.1 and 9.2.  9.3 reversed that course, though.  For better or worse, ALTER
> > TABLE is nearly a union of the relation ALTER possibilities.  That choice is
> > well-entrenched.
> 
> By "well entrenched", I think you mean undocumented, untested, unintentional?

It's deliberate; a -hackers discussion revisits it perhaps once a year.  The
ALTER VIEW documentation says:
 For historical reasons, ALTER TABLE can be used with views too; but the only variants of ALTER TABLE that are allowed
withviews are equivalent to the ones shown above.
 

ALTER INDEX and ALTER SEQUENCE say something similar.  

> Do we think anyone *relies* on being able to say the word TABLE when
> in fact they mean VIEW or SEQUENCE?

pg_dump emits statements that exercise it:
 psql -c 'create view v as select 1 as c; alter view v alter c set default 0;' pg_dump --table v | grep ALTER

> How is that artefact anything but a bug? i.e. is anyone going to stop
> me fixing it?

It's not the behavior I would choose for a new product, but I can't see
benefits sufficient to overturn previous decisions to keep it.

-- 
Noah Misch
EnterpriseDB                                 http://www.enterprisedb.com