[pgpool-general: 1301] Re: pgpool stopped accepting client connections after 1 node hung

[Tatsuo Ishii]

> On Mon, Jan 7, 2013 at 3:05 PM, Tatsuo Ishii <ishii at postgresql.org> wrote:
>>>>>> I don't understand why pgpool stopped accepting client connections.
>>>>>> I'd expect that if any single node goes down, pgpool should continue
>>>>>> to work and accept connections, and simply mark the unresponsive node
>>>>>> as unavailable.
>>>>>
>>>>> That is my question too. Do you see this kind of message in the pgpool log?
>>>>>
>>>>>                 degenerate_backend_set: 2 fail over request from pid xxxx
>>>>>
>>>>> If you see this, pgpool should initiate the failover and mark cuda-db5 down.
>>>>
>>>> Nope, that message was not present at any time.
>>>
>>> There was a bug report regarding pgpool-II 3.2 (or higher)'s
>>> connect_inet_domain_socket():
>>> http://www.pgpool.net/mantisbt/view.php?id=46
>>>
>>> In the report the error message was same as you
>>> (connect_inet_domain_socket: connect() failed: Connection timed out).
>>> and I have created a patch to fix it:
>>> http://www.pgpool.net/mantisbt/file_download.php?file_id=55&type=bug
>>>
>>> Can you try it out? Still I am investigating why you did not see fail
>>> over but I think you want to try the patch to avoid the error first.
>>
>> Oh, I think I see the reason why you do not see fail over.
>> You have this:
>> fail_over_on_backend_error = off
>>
>> In this case new_connection() does not trigger fail over.
>>
>>                         /* If fail_over_on_backend_error is true, do failover.
>>                          * Otherwise, just exit this session.
>>                          */
>>                         if (pool_config->fail_over_on_backend_error)
>>                         {
>>                                 notice_backend_error(i);
>>                         }
>>                         child_exit(1);
> 
> Thanks.
> 
> Just so that I'm clear, your recommendation is that I set
> fail_over_on_backend_error=on *AND* apply the patch you provided, and
> that will allow pgpool to failover an unresponsive postgres server,
> such that pgpool does not stop accepting client connections?

Yes, exactly.
--
Tatsuo Ishii
SRA OSS, Inc. Japan
English: http://www.sraoss.co.jp/index_en.php
Japanese: http://www.sraoss.co.jp