[pgpool-general: 2968] Re: backend_flag for streaming replication
nagata at sraoss.co.jp
Wed Jun 25 12:09:14 JST 2014
I'm sorry for the late reply.
On Thu, 15 May 2014 14:10:25 +0100
Bruce McAlister <bruce.mcalister at blueface.com> wrote:
> Hi All,
> We are currently testing a postgresql cluster and pgpool cluster. I
> would just like to clarify the backend_flag parameter suitable for this
> The postgresql cluster is managed as a pacemaker cluster resource using
> pgsql resource agent, this manages the startup/shutdown/failover of the
> postgresql master instace and synchronous slaves. The postgresql version
> is 9.3.4 and is configured in streaming replication mode.
> The pgpool cluster is setup using the pgpool-ii-ha resource agent (v2.2)
> which is also magaed by pacemaker.
> I have been reading the pgpool doc's and it appears that the
> 'DISALLOW_TO_FAILOVER' parameter is the recommended setting for
> backend_flag for a host definition. I presume that this is because
> pgpool will not need to perform any sort of recovery operations in case
> a postgresql node failed.
> If I set the DISALLOW_TO_FAILOVER parameter, will pgpool continue to
> probe for the new master if there was a failure. For example, if db node
> 0 was master and db node 1 & 2 were slaves, then postgresql fails on the
> master node, the cluster promotes db node 1 as the new master, will
> pgpool reconnect the sessions and active db node 1 as the new write master.
Yes. However, node 0 must be recovered by the cluster software before
pgpool accepts new connection from clients.
> Currently with the testing we have done thus far, pgpool is doing this
> correctly, but I have just noticed that my backen_flag setting is
> 'ALLOW_TO_FAILOVER' so pgpool is firing it's recovery steps.
> Additionally, in this mode, I need to pcp_attach_node once the failed db
> node is recovered.
In this case, pcp_attach_node is needed since down db is detached when
ALLOW_TO_FAILOVER is used.
> How does pgpool handle this scenario when DISALLOW_TO_FAILOVER is the
> parameter in use? Do I also need to run pcp_attach_node in this mode if
> the system failes over a db node?
Wnen DISALLOW_TO_FAILOVER is used:
(1) Even when a backend node goes down, the node never be detached by
connection error, healthcheck failure, or pcp_detachd_node.
(2) However, new connections from clients are rejected since pgpool can't
create connections to backends.
(3) pgpool can accept new connections **after** the node is recovered
by cluster softwre (e.g. pacemaker) or manually. pcp_attach_node isn't
needed since the node never be detached.
> Your thoughts and comments are appreciated :)
> pgpool-general mailing list
> pgpool-general at pgpool.net
Yugo Nagata <nagata at sraoss.co.jp>
More information about the pgpool-general