Re: Broken hint bits (freeze)

On Wed, Jun 14, 2017 at 1:01 AM, Bruce Momjian <bruce(at)momjian(dot)us> wrote:
> On Mon, Jun 12, 2017 at 06:31:11PM +0300, Vladimir Borodin wrote:
>> What about the following sequence?
>>
>> 1. Run pg_upgrade on master,
>> 2. Start it in single-user mode and stop (to get right wal_level in
>> pg_control),
>> 3. Copy pg_control somewhere,
>> 4. Start master, run analyze and stop.
>> 5. Put the control file from step 3 to replicas and rsync them according to the
>> documentation.
>>
>> And I think that step 10.f in the documentation [1] should be fixed to mention
>> starting in single-user mode or with disabled autovacuum.
>>
>> [1] https://www.postgresql.org/docs/devel/static/pgupgrade.html
>
> First, I want to apologize for not getting involved in this thread
> earlier, and I want to thank everyone for the huge amount of detective
> work in finding the cause of this bug.
>
> Let me see if I can replay how the standby server upgrade instructions
> evolved over time.
>
> Initially we knew that we had to set wal_level to replica+ so that when
> you reconnect to the standby servers, the WAL would have the right
> contents. (We are basically simulating pg_start/stop backup with
> rsync.)
>
> There was a desire to have those instructions inside a documentation
> block dedicated to standby server upgrades, so the wal_level adjustment
> and new server start/stop was added to that block. I assumed a
> start/stop could not modify the WAL, or at least nothing important would
> happen, but obviously I was wrong. (pg_upgrade takes steps to ensure
> that nothing happens.) Adding ANALYZE in there just made it worse, but
> the problem always existed. I sure hope others haven't had a problem
> with this.
>
> Now, it seems we later added a doc section early on that talks about
> "Verify standby servers" so I have moved the wal_level section into that
> block, which should be safe. There is now no need to start/stop the new
> server since pg_upgrade will do that safely already.
>

! <para>
! Also, if upgrading standby servers, change <varname>wal_level</>
! to <literal>replica</> in the <filename>postgresql.conf</> file on
! the new cluster.

I think it is better to indicate that this is required for the master
cluster (probably it is clear for users) /"on the new cluster."/"on
the new master cluster.". Do we need something different for v10 where
default wal_level is 'replica'

--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com