On Thu, Aug 12, 2010 at 2:06 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Thu, Aug 12, 2010 at 12:27 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>> Lastly, it bothers me that you've put in code to delete files belonging
>>> to temp rels during crash restart, without any code to clean up their
>>> catalog entries. This will therefore lead to dangling pg_class
>>> references, with uncertain but probably not very nice consequences.
>
>> I thought about this pretty carefully, and I don't believe that there
>> are any unpleasant consequences. The code that assigns relfilenode
>> numbers is pretty careful to check that the newly assigned value is
>> unused BOTH in pg_class and in the directory where the file will be
>> created, so there should be no danger of a number getting used over
>> again while the catalog entries remain. Also, the drop-object code
>> doesn't mind that the physical storage doesn't exist; it's perfectly
>> happy with that situation.
>
> Well, okay, but I'd suggest adding comments to the drop-table code
> pointing out that it is now NECESSARY for it to not complain if the file
> isn't there. This was never a design goal before, AFAIR --- the fact
> that it works like that is kind of accidental. I am also pretty sure
> that there used to be at least warning messages for that case, which we
> would now not want.
That seems like a good idea. I'll post an updated patch.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company