Re: Speed up transaction completion faster after many relations are accessed in a transaction

On Thu, 25 Jul 2019 at 05:49, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> On the whole, I don't especially like this approach, because of the
> confusion between peak lock count and end-of-xact lock count. That
> seems way too likely to cause problems.

Thanks for having a look at this. I've not addressed the points
you've mentioned due to what you mention above. The only way I can
think of so far to resolve that would be to add something to track
peak lock usage. The best I can think of to do that, short of adding
something to dynahash.c is to check how many locks are held each time
we obtain a lock, then if that count is higher than the previous time
we checked, then update the maximum locks held, (probably a global
variable). That seems pretty horrible to me and adds overhead each
time we obtain a lock, which is a pretty performance-critical path.

I've not tested what Andres mentioned about simplehash instead of
dynahash yet. I did a quick scan of simplehash and it looked like
SH_START_ITERATE would suffer the same problems as dynahash's
hash_seq_search(), albeit, perhaps with some more efficient memory
lookups. i.e it still has to skip over empty buckets, which might be
costly in a bloated table.

For now, I'm out of ideas. If anyone else feels like suggesting
something of picking this up, feel free.

--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services