Re: [HACKERS] Use of non-restart-safe storage by temp_tablespaces

> On May 31, 2017, at 7:58 AM, Bruce Momjian  wrote:
> 
> On Wed, May 31, 2017 at 07:53:22AM -0700, Mark Dilger wrote:
>>> Just to clarify, the TEMPORARY clause would allow the tablespace to
>>> start up empty, while normal tablespaces can't do that, right?  One big
>>> problem is that we don't have any way to prevent non-temporary
>>> tablespaces from being created on transient storage.  I wonder if we
>>> should document this restriction, but it seems awkward to do.
>> 
>> It depends what you mean by allowing the tablespace to start up empty.
>> It must not be empty once users can run queries, since the catalogs will still
>> have entries for the tablespace and its dependent objects.  So, what must
>> happen is that during startup the tablespace and its temporary tables and
>> indexes get recreated if they are missing.
> 
> Uh, I thought only the sessions that created the temporary objects could
> see them, and since they are not in WAL and autovacuum can't see them,
> their non-existence in a temporary tablespace would not be a problem.

You are correct.  I was thinking about an extension to allow unlogged
tablespaces on temporary filesystems, but got the words "unlogged" and
"temporary" mixed up in my thinking and in what I wrote.  I should have
written that unlogged tablespaces would only host unlogged tables and
unlogged indexes, such that users are not surprised to find their data
missing.

On reflection, I think both features are worthwhile, and not at all exclusive
of each other, though unlogged tablespaces is probably considerably more
work to implement.

Mark Dilger