Re: In-placre persistance change of a relation

At Wed, 6 Jul 2022 08:44:18 -0700, Jacob Champion <jchampion(at)timescale(dot)com> wrote in
> On Thu, Mar 31, 2022 at 2:36 AM Kyotaro Horiguchi
> <horikyota(dot)ntt(at)gmail(dot)com> wrote:
> >
> > At Thu, 31 Mar 2022 18:33:18 +0900 (JST), Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> wrote in
> > > I don't think this can be commited to 15. So I post the fixed version
> > > then move this to the next CF.
> >
> > Then done. Thanks!
>
> Hello! This patchset will need to be rebased over latest -- looks like
> b74e94dc27f (Rethink PROCSIGNAL_BARRIER_SMGRRELEASE) and 5c279a6d350
> (Custom WAL Resource Managers) are interfering.

Thank you for checking that! It got a wider attack by b0a55e4329
(RelFileNumber). The commit message suggests "relfilenode" as files
should be replaced with "relation storage/file" so I did that in
ResetUnloggedRelationsInDbspaceDir.

This patch said that:

> * INIT forks are never loaded to shared buffer so no point in
> * dropping buffers for such files.

But actually some *buildempty() functions use ReadBufferExtended() for
INIT_FORK. So that's wrong. So, I did that but... I don't like that.
Or I don't like that some AMs leave buffers for INIT fork after. But I
feel I'm misunderstanding here since I don't understand how the INIT
fork can work as expected after a crash that happens before the next
checkpoint flushes the buffers.

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center