Re: Unlogged tables cannot be truncated twice

On Wed, Jun 1, 2011 at 1:15 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> 2011/5/31 Andres Freund <andres(at)anarazel(dot)de>:
>> On Tuesday, May 31, 2011 03:27:22 Alvaro Herrera wrote:
>>> Excerpts from Andres Freund's message of lun may 30 20:47:49 -0400 2011:
>>> > On Tuesday, May 31, 2011 02:35:58 AM Andres Freund wrote:
>>> > > On Tuesday, May 31, 2011 02:14:00 AM Andres Freund wrote:
>>> > > > On Tuesday, May 31, 2011 01:56:05 AM Cédric Villemain wrote:
>>> > > > > I remove my own explanations as we conclude on the same thing.
>>> > > > > Attached is the fix by adding a (!reindex)  in the index.c if().
>>> > > >
>>> > > > Thats imo wrong because it will break a plain REINDEX?
>>> > >
>>> > > > I think one possible correct fix would be the attached:
>>> > > My version was wrong as well because it  did not observe
>>> > > RelationTruncate's nblocks argument. That function is used to
>>> > > "shorten" the relation in vacuum. So dropping the init fork there is
>>> > > not a good idea.
>>> > >
>>> > > So I think it is the simpler version of simply checking the existance
>>> > > of the fork before creating is ok.
>>>
>>> Hmm, I wonder if what we should be doing here is observe isreindex in
>>> index_build to avoid creating the init fork.  Doing smgr access at that
>>> level seems wrong.
>> isreindex doesn't contain the necessary informormation as its set doing a
>> REINDEX even though a new relfilenode is created and thus the fork needs to be
>> created.
>>
>> It doesn't seem terribly clean do do the !smgrexists(), I aggree with you
>> there. On the other hand we are calling smgrcreate() two lines down anyway. I
>> personally don't realy like the placement of that piece of code very much.
>> Doing it in index_build seems to be the wrong place. I don't think there
>> really is a good place for it right now.
>
> I'm open to suggestions on how to rearrange this, but I think for
> right now the approach you proposed upthread (add a smgrexists() test)
> is probably the simplest way to fix this.

Done. Your patch tested for FSM_FORKNUM instead of INIT_FORKNUM,
which seemed wrong, so I changed it. I also added comments.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company