At Tue, 4 Oct 2022 17:15:31 -0700, Nathan Bossart <nathandbossart@gmail.com> wrote in
> On Tue, Oct 04, 2022 at 07:53:11PM -0400, Tom Lane wrote:
> > I wrote:
> >> PFA a quick-hack fix that solves this issue by making per-transaction
> >> subsidiary hash tables. That's overkill perhaps; I'm a little worried
> >> about whether this slows down normal cases more than it's worth.
> >> But we ought to do something about this, because aside from the
> >> duplication aspect the current storage of these lists seems mighty
> >> space-inefficient.
> >
> > After further thought, maybe it'd be better to do it as attached,
> > with one long-lived hash table for all the locks. This is a shade
> > less space-efficient than the current code once you account for
> > dynahash overhead, but the per-transaction overhead should be lower
> > than the previous patch since we only need to create/destroy a hash
> > table entry not a whole hash table.
First one is straight forward outcome from the current implement but I
like the new one. I agree that it is natural and that the expected
overhead per (typical) transaction is lower than both the first one
and doing the same operation on a list. I don't think that space
inefficiency in that extent doesn't matter since it is the startup
process.
> This feels like a natural way to solve this problem. I saw several cases
> of the issue that was fixed with 6301c3a, so I'm inclined to believe this
> usage pattern is actually somewhat common.
So releasing locks becomes somewhat slower? But it seems to still be
far faster than massively repetitive head-removal in a list.
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center
