Tom,
> I'm not excited about it either, but I think if we're going to make
> this adjustable it does need its own knob. I can easily believe
> that a large list of precompiled GUCs could be counterproductive
> given a workload where you don't get much reuse, so I don't want
> the list size going to the moon just because someone cranked up
> work_mem for other purposes.
Yes. I was just trying to avoid thinking about it. ;-)
>
> (I'm not real sure that that "self-organizing list" data structure
> would work well beyond 1000 or so entries even if you did have
> enough re-use to justify them all. Anyone want to try to do some
> performance testing? In particular I think we might want to drop
> the move-to-front approach in favor of move-up-one, just to avoid
> O(N^2) memmove costs.)
Hmmm. Yeah, I can see that.
Well, I have a test case here (the PostgreSQL download logs), or I
wouldn't have brought up the issue. I just need to find a way to
multi-thread it so I can get the effect of multiple clients.
--Josh