Re: pg_basebackup blocking all queries with horrible performance

On Sun, Jun 10, 2012 at 4:08 PM, Fujii Masao <masao(dot)fujii(at)gmail(dot)com> wrote:
> On Sun, Jun 10, 2012 at 10:34 PM, Fujii Masao <masao(dot)fujii(at)gmail(dot)com> wrote:
>> On Sun, Jun 10, 2012 at 9:25 PM, Fujii Masao <masao(dot)fujii(at)gmail(dot)com> wrote:
>>> On Sun, Jun 10, 2012 at 7:43 PM, Magnus Hagander <magnus(at)hagander(dot)net> wrote:
>>>> On Sat, Jun 9, 2012 at 2:51 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>>>>> Fujii Masao <masao(dot)fujii(at)gmail(dot)com> writes:
>>>>>> This seems a bug. I think we should prevent pg_basebackup from
>>>>>> becoming synchronous standby. Thought?
>>>>>
>>>>> Absolutely.  If we have replication clients that are not actually
>>>>> capable of being standbys, there *must* be a way for the master
>>>>> to know that.
>>>>
>>>> I thought we fixed this already by sending InvalidXlogRecPtr as flush
>>>> location? And that this only applied in 9.2?
>>>>
>>>> Are you saying we picked pg_basebackup *in backup mode* (not log
>>>> streaming) as synchronous standby?
>>>
>>> Yes.
>>>
>>>> If so then yes, that is
>>>> *definitely* a bug that should be fixed. We should never select a
>>>> connection that's not even streaming log as standby!
>>>
>>> Agreed. Attached patch prevents pg_basebackup from becoming sync
>>> standby. Also this patch fixes another problem: currently only walsender
>>> which reaches STREAMING state can become sync walsender. OTOH,
>>> sync walsender thinks that walsender with higher priority will be sync one
>>> whether its state is STREAMING, and switches to potential sync walsender.
>>> So when the standby with higher priority connects to the master, we
>>> might have no sync standby until it reaches the STREAMING state.
>>> To fix this problem, the patch switches walsender's state from sync to
>>> potential *after* walsender with higher priority has reached the
>>> STREAMING state.
>>>
>>> We also should not select (1) background stream process forked from
>>> pg_basebackup and (2) pg_receivexlog as sync standby because they
>>> don't send back replication progress. To address this, I'm thinking to
>>> introduce new option "NOSYNC" in "START_REPLICATION" command
>>> as follows, and to change (1) and (2) so that they specify NOSYNC.
>>>
>>>    START_REPLICATION XXX/XXX [NOSYNC]
>>>
>>> If the standby specifies NOSYNC option, it's never assigned as sync
>>> standby even if its name is in synchronous_standby_names. Thought?
>>
>> The standby which always sends InvalidXLogRecPtr back should not
>> become sync one. So instead of NOSYNC option, by checking whether
>> InvalidXLogRecPtr is sent, we can avoid problematic sync standby.
>
> We should not do this because Magnus is proposing the patch
> (http://archives.postgresql.org/pgsql-hackers/2012-06/msg00348.php)
> which breaks the above assumption at all. So we should introduce
> something like NOSYNC option.

Wouldn't the better choice there in that case be to give a switch to
pg_receivexlog if you *want* it to be able to become a sync replica,
and by default disallow it? And then keep the backend just treating
InvalidXlogRecPtr as don't-become-sync-replica.

--
 Magnus Hagander
 Me: http://www.hagander.net/
 Work: http://www.redpill-linpro.com/