Re: WAL replay bugs

Thanks for the tests.

On Wed, Nov 5, 2014 at 5:21 AM, Alvaro Herrera <alvherre(at)2ndquadrant(dot)com>
wrote:

> Michael Paquier wrote:
> > On Mon, Jul 14, 2014 at 6:14 PM, Kyotaro HORIGUCHI
> > <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> wrote:
> > > Although I doubt necessity of the flexibility seeing the current
> > > testing framework, I don't have so strong objection about
> > > that. Nevertheless, perhaps you are appreciated to put a notice
> > > on.. README or somewhere.
> > Hm, well... Fine, I added it in this updated series.
>
> FWIW I gave this a trial run and found I needed some tweaks to test.sh
> and the Makefile in order to make it work on VPATH; mainly replace ./
> with `dirname $0` in a couple test.sh in a couple of places, and
> something similar in the Makefile. Also you have $PG_ROOT_DIR somewhere
> which doesn't work.
>
Ah thanks, forgot that.

> Also you have the Makefile checking for -DBUFFER_CAPTURE exactly but for
> some reason I used -DBUFFER_CAPTURE=1 which wasn't well received by your
> $(filter) stuff. Instead of checking CFLAGS it might make more sense to
> expose it as a read-only GUC and grep `postmaster -C buffer_capture` or
> similar.
>
Yes that's a good idea.

Now, do we really want this feature in-core? That's somewhat a duplicate of
what is mentioned here:
http://www.postgresql.org/message-id/CAB7nPqQMq=4eJAK317mxZ4Has0i+1rSLBQU29zx18JwLB2j1OA@mail.gmail.com
Of course both things do not have the same coverage as the former is for
buildfarm and dev, while the latter is dedidated to production systems, but
could be used for development as well.

The patch sent there is a bit outdated, but a potential implementation gets
simpler with XLogReadBufferForRedo able to return flags about each block
state during redo. I am still planning to come back to it for this cycle,
though I stopped for now waiting for the WAL format patches finish to shape
the APIs this feature would rely on.
Regards,
--
Michael