Re: pg_upgrade and rsync

> 6 марта 2015 г., в 6:11, Bruce Momjian <bruce(at)momjian(dot)us> написал(а):
>
> On Thu, Mar 5, 2015 at 10:55:28AM +0300, Vladimir Borodin wrote:
>> You are correct that a pg_controldata file is copied over that has
>> wal_level=minimal, but that should not be a problem.
>>
>>
>> I suppose, this is the root cause of why replica does not start as hot standby.
>> It it enough to start it as warm standby, but not hot standby.
>> See CheckRequiredParameterValues function in xlog.c which is called inside of
>> StartupXLOG function.
>
> Yes, you are correct. I spent all day building a test harness so I
> could automate this setup and test various failures. I was able to
> reproduce your failure, and you are correct that the proper fix is to
> set wal_level=hot_standby on the new master, and then start and stop the
> new cluster just before rsync.
>
> The root cause is that pg_upgrade calls pg_resetxlog -o on the new
> cluster _after_ the new cluster stopped for the final time, so rsync is
> copying the incorrect pg_controldata wal_level value. Also, even if
> pg_resetxlog preserved wal_level in the control file, there is no
> guarantee that the user configured the new cluster's wal_level for
> hot_standby anyway.
>
> What I have done is to update the pg_upgrade instructions to add this
> required step. Updated doc patch attached. (I also added the --delete
> flag to rsync.) Thanks so much for your detailed report.

It seems to work fine now. The only thing that would be nice to change is to explicitly write that these instructions are correct for hot standby installations too.

+ <para>
+ If you have Log-Shipping Standby Servers (<xref
+ linkend="warm-standby">), follow these steps to upgrade them (before
+ starting any servers):
+ </para>

Actually, I’ve entered this thread because it is not obvious from the paragraph above or any other places.

>
>> But it could not be done with --size-only key, because control-file is
>> of fixed
>> size and rsync would skip it. Or may be everything should be copied
>> with
>> --size-only and control-file should be copied without this option.
>>
>>
>> Well, what happens is that there is no _new_ standby pg_controldata
>> file, so it is copied fully from the new master. Are you running initdb
>> to create the new standby --- you shouldn't be doing that as the rsync
>> will do that for you.
>>
>>
>> No, I don’t. The scenario of the problem with copying control-file was in case
>> when I:
>> 1. ran pg_upgrade on master and got control-file with "wal_level = minimal",
>> 2. did rsync --size-only to replica (and it got this control-file with
>> "wal_level = minimal"),
>> 3. started and stopped postgres on master to get «good» control-file with
>> "wal_level = hot_standby»,
>> 4. did rsync --size-only to replica one more time. And this time control-file
>> is not copied because of the same size of control-file.
>>
>> Actually, if you don’t do step 2, everything works as expected. Sorry for
>> bothering you.
>
> Ah, yes, I think doing rsync twice is never a good suggestion. It can
> lead to too many failures. Doing the start/stop before rsync seems like
> the best solution.
>
> --
> Bruce Momjian <bruce(at)momjian(dot)us> http://momjian.us
> EnterpriseDB http://enterprisedb.com
>
> + Everyone has their own god. +
> <rsync.diff>

--
May the force be with you…
https://simply.name