Horton

Re: BUG #14999: pg_rewind corrupts control file global/pg_control

Michael Paquier <michael@paquier.xyz>

View thread View on postgr.es
From: Michael Paquier <michael@paquier.xyz>
To: Masahiko Sawada <sawada.mshk@gmail.com>
Cc: Dmitry Dolgov <9erthalion6@gmail.com>, TipTop Labs <office@tiptop-labs.com>, Michael Paquier <michael.paquier@gmail.com>, PostgreSQL mailing lists <pgsql-bugs@postgresql.org>
Date: 2018-03-05T09:01:11Z
Lists: pgsql-bugs, pgsql-hackers 
On Mon, Mar 05, 2018 at 05:47:06PM +0900, Masahiko Sawada wrote:
> Since files other than relation files such as vm, fsm, WAL are
> categorize as FILE_ACTION_COPY_CONFIG, I think the name would cause
> misreading. For example, we can use FILE_ACTION_COPY_DATA and
> FILE_ACTION_COPY_OTHER?

I am not stopped on a given name.

> As you mentioned before, with this patch we end up ignoring file-open
> error of database files other than relation files. I guess it's a bit
> risky. We can enter them to COPY_DATA but I'd defer it to committer.

COPY_DATA is aimed at being used on files which can be parsed by
blocks.  You cannot do that with VMs and FSMs.  Now you are true that we 
could complain for non-configuration files which need to be
fully-copied, still Postgres issues a fsync on all the files of the data
folder when beginning recovery, so you would bump on problems anyway
after restarting the rewound instance.
--
Michael

Commits

  1. Add note in pg_rewind documentation about read-only files