Hi,
On 2019-02-18 19:13:31 -0500, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2019-02-18 19:01:06 -0500, Tom Lane wrote:
> >> Mmm ... AIUI, the patches currently proposed can only help for what
> >> David called "point lookup" queries. There are still going to be
> >> queries that scan a large proportion of a partition tree, so if you've
> >> got tons of partitions, you'll be concerned about this sort of thing.
>
> > Agreed - but it seems not unlikely that for those the rest of the
> > planner / executor overhead will entirely swamp any improvement we could
> > make here. If I understand correctly the benchmarks here were made with
> > "point" update and select queries, although the reference in the first
> > post in this thread is a bit vague.
>
> I think what Maumau-san is on about here is that not only does your
> $giant-query take a long time, but it has a permanent negative effect
> on all subsequent transactions in the session. That seems worth
> doing something about.
Ah, yes, that makes sense. I'm inclined to think however that the
original approach presented in this thread is better than the
reset-the-whole-hashtable approach. Because:
> I think David's on the right track --- keep some kind of moving average of
> the LOCALLOCK table size for each transaction, and nuke it if it exceeds
> some multiple of the recent average. Not sure offhand about how to get
> the data cheaply --- it might not be sufficient to look at transaction
> end, if we release LOCALLOCK entries before that (but do we?)
Seems too complicated for my taste. And it doesn't solve the issue of
having some transactions with few locks (say because the plan can be
nicely pruned) interspersed with transactions where a lot of locks are
acquired.
Greetings,
Andres Freund