On Sat, Apr 24, 2010 at 6:29 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
> On Fri, 2010-04-23 at 22:52 -0400, Robert Haas wrote:
>> Thoughts?
>
> Only a requirement: that we design this in a way that will allow temp
> tables to be used during Hot Standby. I make not other comment.
For so long as local temporary tables put their catalog entries in
pg_class, we're not going to be able to use them during Hot Standby.
See the email I just sent elsewhere on this thread for a long term
roadmap to getting out of that pickle. At least under the
implementation I'm proposing here, making global temporary tables
usable would be an easier nut to crack, because the catalog entries
are a non-issue. There is one major problem, though: assigning a
scratch relfilenode to the temporary table requires generating an OID,
which we currently have no way to allow on the standby. Upthread I
also proposed an implementation for unlogged tables (that is, contents
don't survive a server bounce) which wouldn't have that problem
either, although I haven't fully thought it through and there may very
well be other issues.
So in short: I don't think anything we're talking about it would make
HS use harder, and some of it might make it easier. But probably some
additional engineering effort dedicated specifically to solving the
problems unique to HS would still be needed.
...Robert