On Mon, Aug 7, 2023 at 3:02 PM Tomas Vondra
<tomas.vondra@enterprisedb.com> wrote:
> > I would also argue that the results are actually not that great,
> > because once you get past 64 partitions you're right back where you
> > started, or maybe worse off. To me, there's nothing magical about
> > cases between 16 and 64 relations that makes them deserve special
> > treatment - plenty of people are going to want to use hundreds of
> > partitions, and even if you only use a few dozen, this isn't going to
> > help as soon as you join two or three partitioned tables, and I
> > suspect it hurts whenever it doesn't help.
>
> That's true, but doesn't that apply to any cache that can overflow? You
> could make the same argument about the default value of 16 slots - why
> not to have just 8?
Yes and no. I mean, there are situations where when the cache
overflows, you still get a lot of benefit out of the entries that you
are able to cache, as when the frequency of access follows some kind
of non-uniform distribution, Zipfian or decreasing geometrically or
whatever. There are also situations where you can just make the cache
big enough that as a practical matter it's never going to overflow. I
can't think of a PostgreSQL-specific example right now, but if you
find that a 10-entry cache of other people living in your house isn't
good enough, a 200-entry cache should solve the problem for nearly
everyone alive. If that doesn't cause a resource crunch, crank up the
cache size and forget about it. But here we have neither of those
situations. The access frequency is basically uniform, and the cache
size needed to avoid overflows seems to be unrealistically large, at
least given the current design. So I think that in this case upping
the cache size figures to be much less effective than in some other
cases.
It's also a bit questionable whether "cache" is even the right word
here. I'd say it isn't, because it's not like the information in the
fast-path locking structures is a subset of the full information
stored elsewhere. Whatever information is stored there is canonical
for those entries.
> Yes, I agree. I don't know if this particular design would be the right
> one (1000 elements seems a bit too much for something included right in
> PGPROC). But yeah, something that flips from linear search to something
> else would be reasonable.
Yeah ... or there could be a few slots in the PGPROC and then a bit
indicating whether to jump to a larger shared memory structure located
in a separate array. Not sure exactly.
--
Robert Haas
EDB: http://www.enterprisedb.com
