Re: [Patch] Optimize dropping of relation buffers using dlist

On Wed, Dec 23, 2020 at 1:07 PM k(dot)jamison(at)fujitsu(dot)com
<k(dot)jamison(at)fujitsu(dot)com> wrote:
>
> On Tuesday, December 22, 2020 9:11 PM, Amit Kapila wrote:
>
> > In this code, I am slightly worried about the additional cost of each time
> > checking smgrexists. Consider a case where there are many relations and only
> > one or few of them have not cached the information, in such a case we will
> > pay the cost of smgrexists for many relations without even going to the
> > optimized path. Can we avoid that in some way or at least reduce its usage to
> > only when it is required? One idea could be that we first check if the nblocks
> > information is cached and if so then we don't need to call smgrnblocks,
> > otherwise, check if it exists. For this, we need an API like
> > smgrnblocks_cahced, something we discussed earlier but preferred the
> > current API. Do you have any better ideas?
>
> Right. I understand the point that let's say there are 100 relations, and
> the first 99 non-local relations happen to enter the optimization path, but the last
> one does not, calling smgrexist() would be too costly and waste of time in that case.
> The only solution I could think of as you mentioned is to reintroduce the new API
> which we discussed before: smgrnblocks_cached().
> It's possible that we call smgrexist() only if smgrnblocks_cached() returns
> InvalidBlockNumber.
> So if everyone agrees, we can add that API smgrnblocks_cached() which will
> Include the "cached" flag parameter, and remove the additional flag modifications
> from smgrnblocks(). In this case, both DropRelFileNodeBuffers() and
> DropRelFileNodesAllBuffers() will use the new API.
>

Yeah, let's do it that way unless anyone has a better idea to suggest.
--
With Regards,
Amit Kapila.