Re: [BUG] pg_basebackup from disconnected standby fails

On Wed, Oct 26, 2016 at 1:53 PM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> On Mon, Oct 24, 2016 at 12:25 PM, Michael Paquier
> <michael(dot)paquier(at)gmail(dot)com> wrote:
>> On Mon, Oct 24, 2016 at 1:39 PM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>>>
>>> I think what you are saying is not completely right, because we do
>>> update minRecoveryPoint when we don't perform a new restart point.
>>> When we perform restart point, then it assumes that flushing the
>>> buffers will take care of updating minRecoveryPoint.
>>
>> Yep, minRecoveryPoint still gets updated when the last checkpoint
>> record is the last restart point to avoid a hot standby to allow
>> read-only connections at a LSN-point earlier than the last shutdown.
>> Anyway, we can clearly reject 1. in the light of
>> https://www.postgresql.org/message-id/CAA4eK1KmjtsXqF0cav7Cs4d4vwv2H_pc8d8q1BUCqDzAF+7EzQ@mail.gmail.com
>> when playing with different stop locations at recovery.
>>
>
> That point holds good only for cases when we try to update minimum
> recovery point beyond what is required (like earlier we were thinking
> to update it unconditionally), however what is being discussed here is
> to update only if it is not updated by flush of buffers. I think that
> is okay, but I feel Kyotaro-San's fix is a good fix for the problem
> and we don't need to add some more code (additional update of control
> file) to fix the problem.

We have the choice between a solution that requires any backup taken
from standbys to need more WAL than it needs, not sure if would be a
lot, and a solution that prevents minRecoveryPoint to get older than
the point where standby would start recovery anyway at next startup.
For correctness the latter solution looks more robust to me to be
honest.
--
Michael