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

On Tue, Dec 8, 2020 at 6:23 AM Kyotaro Horiguchi
<horikyota(dot)ntt(at)gmail(dot)com> wrote:
>
> At Tue, 08 Dec 2020 09:45:53 +0900 (JST), Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> wrote in
> > At Mon, 7 Dec 2020 17:18:31 +0530, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote in
> > > Hmm, how is it possible if Insert is done before Truncate? The insert
> > > should happen in old RelFileNode only. I have verified by adding a
> > > break-in (while (1), so that it stops there) heap_xlog_insert and
> > > DropRelFileNodesAllBuffers(), and both get the same (old) RelFileNode.
> > > How have you verified what you are saying?
> >
> > You might be thinking of in-transaction sequence of
> > Inert-truncate. What *I* mention before is truncation of a relation
> > that smgrnblocks() has already been called for. The most common way
> > to make it happen was INSERTs *before* the truncating transaction
> > starts.

What I have tried is Insert and Truncate in separate transactions like below:
postgres=# insert into mytbl values(1);
INSERT 0 1
postgres=# truncate mytbl;
TRUNCATE TABLE

After above, manually killed the server, and then during recovery, we
have called heap_xlog_insert() and DropRelFileNodesAllBuffers() and at
both places, RelFileNode is the same and I don't see any reason for it
to be different.

> > It may be a SELECT on a hot-standby. Sorry for the confusing
> > expression.
>
> And ,to make sure, it is a bit off from the point of the discussion as
> I noted. I just meant that the proposition that "smgrnblokcs() always
> returns false for "cached" when it is called in
> DropRelFileNodesAllBuffers()" doesn't always holds.
>

Right, I feel in some cases the 'cached' won't be true like if we
would have done Checkpoint after Insert in the above case (say when
the only WAL to replay during recovery is of Truncate) but I think
that should be fine. What do you think?

--
With Regards,
Amit Kapila.