Re: Unhappy about API changes in the no-fsm-for-small-rels patch

On Fri, Apr 19, 2019 at 10:38 AM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> On Thu, Apr 18, 2019 at 2:10 PM John Naylor <john(dot)naylor(at)2ndquadrant(dot)com> wrote:
> > Agreed. I suspect the most realistic way to address most of the
> > objections in a short amount of time would be to:
> >
> > 1. rip out the local map
> > 2. restore hio.c to only checking the last block in the relation if
> > there is no FSM (and lower the threshold to reduce wasted space)
> > 3. reduce calls to smgr_exists()
> >
>
> Won't you need an extra call to RelationGetNumberofBlocks to find the
> last block?

If I understand you correctly, no, the call now in
GetPageWithFreeSpace() just moves it back to where it was in v11. In
the corner case where we just measured the table size and the last
block is full, we can pass nblocks to RecordAndGetPageWithFreeSpace().
There might be further optimizations available if we're not creating a
local map.

> Also won't it be less efficient in terms of dealing with
> bloat as compare to current patch?

Yes. The threshold would have to be 2 or 3 blocks, and it would stay
bloated until it passed the threshold. Not great, but perhaps not bad
either.

> I think if we go this route, then
> we might need to revisit it in the future to optimize it, but maybe
> that is the best alternative as of now.

It's a much lighter-weight API, which has that much going for it.
I have a draft implementation, which I can share if it comes to that
-- it needs some more thought and polish first.

> I am thinking that we should at least give it a try to move the map to
> rel cache level to see how easy or difficult it is and also let's wait
> for a day or two to see if Andres/Tom has to say anything about this
> or on the response by me above to improve the current patch.

Since we have a definite timeline, I'm okay with that, although I'm
afraid I'm not quite knowledgeable enough to help much with the
relcache piece.

--
John Naylor https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services