Re: In-placre persistance change of a relation

At Fri, 25 Dec 2020 09:12:52 +0900 (JST), Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> wrote in
> Hello.
>
> At Thu, 24 Dec 2020 17:02:20 +0900 (JST), Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> wrote in
> > The patch is attached to the next message.
>
> The reason for separating this message is that I modified this so that
> it could solve another issue.
>
> There's a complain about orphan files after crash. [1]
>
> 1: https://www.postgresql.org/message-id/16771-cbef7d97ba93f4b9@postgresql.org
>
> That is, the case where a relation file is left alone after a server
> crash that happened before the end of the transaction that has created
> a relation. As I read this, I noticed this feature can solve the
> issue with a small change.
>
> This version gets changes in RelationCreateStorage and
> smgrDoPendingDeletes.
>
> Previously inittmp fork is created only along with an init fork. This
> version creates one always when a relation storage file is created. As
> the result ResetUnloggedRelationsInDbspaceDir removes all forks if the
> inttmp fork of a logged relations is found. Now that pendingDeletes
> can contain multiple entries for the same relation, it has been
> modified not to close the same smgr multiple times.
>
> - It might be better to split 0001 into two peaces.
>
> - The function name ResetUnloggedRelationsInDbspaceDir is no longer
> represents the function correctly.

As pointed by Robert in another thread [1], persisntence of (at least)
GiST index cannot be flipped in-place due to incompatibility of fake
LSNs with real ones.

This version RelationChangePersistence() is changed not to choose
in-place method for indexes other than btree. It seems to be usable
with all kind of indexes other than Gist, but at the mement it applies
only to btrees.

1: https://www.postgresql.org/message-id/CA+TgmoZEZ5RONS49C7mEpjhjndqMQtVrz_LCQUkpRWdmRevDnQ@mail.gmail.com

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center