Michael Paquier <michael@paquier.xyz> wrote on 06/27/2019 11:17:10 PM:
> From: Michael Paquier <michael@paquier.xyz>
> To: Brad Nicholson <bradn@ca.ibm.com>
> Cc: Grigory Smolkin <g.smolkin@postgrespro.ru>, pgsql-
> general@lists.postgresql.org
> Date: 06/27/2019 11:17 PM
> Subject: [EXTERNAL] Re: Re: Enabling checksums on a streaming replica
>
> On Wed, Jun 26, 2019 at 02:08:19PM -0400, Brad Nicholson wrote:
> > In testing, it doesn't appear to matter. I've ensured that I've generated
> > some full page writes (confirmed via pg_waldump), and those apply
> > fine.
>
> Full pages writes are first written from shared buffers to WAL, where
> their checksums does not actually apply. When the WAL records are
> read an applied, a full page image is recovered in shared buffers.
> The checksum of the page would get updated once the shared buffer page
> used is evicted and written back to disk.
>
So if all the checksums are being recalculated on the replica, this approach should be relatively safe, should it not?
Assuming pg_checksums is doing the right thing (and it looks to me like it should be).
Brad.