On Fri, Jun 21, 2019 at 6:54 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> > That's not a bad goal, although invoking a user-supplied callback
> > while holding a buffer lock is a little scary.
>
> I nominate Robert for Understater of the Year. I think there's pretty
> much 0 chance of that working reliably.
It's an honor to be nominated, although I am pretty sure this is not
my best work in category, even for 2019. There are certainly useful
things that could be done by such a callback without doing anything
that touches shared memory and without doing anything that consumes
more than a handful of CPU cycles, so it doesn't seem utterly crazy to
think that such a design might survive. However, the constraints we'd
have to impose might chafe.
I am more inclined to ditch the callback model altogether in favor of
putting any necessary looping logic on the caller side. That seems a
lot more flexible, and the only trick is figuring out how to keep it
cheap. Providing some kind of context object that can hold onto one or
more pins seems like the most reasonable approach. Last week it seemed
to me that we would need several, but at the moment I can't think of a
reason why we would need more than one. I think we just want to
optimize the case where several undo lookups in quick succession are
actually reading from the same page, and we don't want to go to the
expense of looking that page up multiple times. It doesn't seem at all
likely that we would have a chain of undo records that leaves a
certain page and then comes back to it later, because this is a log
that grows forward, not some kind of random-access thing. So a cache
of size >1 probably wouldn't help.
Unless I'm still confused.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company