Andres Freund <andres@anarazel.de> writes:
> On 2022-11-21 15:12:15 -0500, Tom Lane wrote:
>> If I were trying to find a better fix I'd be looking for ways for
>> parallel workers to be able to read the parent's temp tables.
>> (Perhaps that could tie in with the blue-sky discussion we had
>> the other day about allowing autovacuum on temp tables??)
> That'd be a nontrivial change, because we explicitly don't use any
> locking for anything relating to localbuf.c. One possible benefit could
> be that we could substantially reduce the code duplication between
> "normal" bufmgr.c and localbuf.c.
I didn't say this was easy ;-). Aside from locking, the local buffers
are inside the process's address space and not accessible from outside.
Maybe they could be mapped into a shared memory region instead?
And there are optimizations like commit a7212be8b that depend on the
assumption that nothing else is accessing our process's temp tables.
That'd need a lot of thought, if we don't want to give up all the
performance benefits of temp tables.
regards, tom lane
