[Pgpool-general] RE: Bad degeneration when shutting down secondary backend

[Pablo Milano]

Hi Tatsuo, thanks for your response.

I am already using health check and I have also set
"replication_stop_on_mismatch = true". The only difference is the pgpool
version. The problem can be reproduced as follows:

1) Setup a pgpool host with two backends in linux boxes with the settings
explained above.
2) Start "stressing" the pgpool host with a constant transaction rate (about
4 tps).
3) Halt or reboot the secondary backend.

PS: I know that 2.6.2 is the latest version, but the changelog about the EOF
error was in the 2.6.1 version


> -----Original Message-----
> From: Tatsuo Ishii [mailto:t-ishii at sra.co.jp] 
> Sent: Domingo, 14 de Agosto de 2005 09:10 a.m.
> To: Pablo G. Milano
> Cc: pgpool-general at pgfoundry.org
> Subject: Re: Bad degeneration when shutting down secondary backend
> 
> > I am using pgpool 2.5.1 with linux kernel 2.6.8
> > 
> > I noticed this repeteable behaviour in a heavily loaded environment:
> > When shutting down the linux box containing the backend 
> configured as
> > "secondary", while sending transactions to pgpool, 
> degeneration mode shuts
> > down primary backend instead of secondary. The log shows 
> the following:
> > 
> > ---------------------------------------
> > Aug 12 15:04:14 asimov pgpool: 2005-08-12 15:04:14 ERROR: pid 5888:
> > pool_process_query: kind does not match between backends master(C)
> > secondary(E)
> > Aug 12 15:04:14 asimov pgpool: 2005-08-12 15:04:14 LOG:   pid 5888:
> > notice_backend_error: master: 1 fail over request from pid 5888
> > Aug 12 15:04:14 asimov pgpool: 2005-08-12 15:04:14 LOG:   
> pid 5876: starting
> > degeneration. shutdown master host asimov(5432)
> > Aug 12 15:04:15 asimov pgpool: 2005-08-12 15:04:15 LOG:   pid 5876:
> > degeneration done. shutdown master host asimov(5432)
> > ---------------------------------------
> > 
> > It seems that pgpool encounters a data mismatch between 
> backends, when what
> > really happens is that one of them is dying.
> > I've found this in pgpool 2.6.1 changelog:
> > 
> > 	* do not raise failover when read() read encounters 
> EOF. It seems it
> > 	  is a oversight. Backend crush may cause that but it does not
> > 	  immediately mean postmaster crush
> > 
> > Does this solve the problem I'm having? Should I move to 
> pgpool 2.6.1?
> > Thanks in advance.
> 
> Maybe.
> 
> But for particular error like this:
> 
> > pool_process_query: kind does not match between backends master(C)
> > secondary(E)
> 
> I don't think upgrading to 2.6.2(the latest version is 2.6.2, not
> 2.6.21, FYI) solves the problem. The query results do not agree
> between two servers. pgpool cannot judge which one is healthy. So it
> just shuts down the master. 
> 
> What I would sugget here is:
> 
> 1) set following in pgpoo.conf. This will avoid unwanted degeneration.
> 
>    replication_stop_on_mismatch = true
> 
> 2) upgrade to 2.6.2. and use "health check" functionality. Heal check
>    will occasionally checks if the servers are healthy and if
>    something goes wrong, automatically starts the degeneration.
> --
> Tatsuo Ishii
>