[ClusterLabs] Apache graceful restart not supported by heartbeat apache control script

[Ken Gaillot]

On Mon, 2019-03-25 at 10:42 -0700, Cole Miller wrote:
> Hi users at clusterlabs.org,
> 
> My current project at work is a two node cluster running apache and
> virtual IPs on CentOS 7. I found in my testing that apache when run
> by corosync does not have a reload or graceful restart. Before the
> cluster, when apache was a single instance, I previously used
> graceful restart to be able to add configurations with out affecting
> current users of the system.
> 
> Initially, I tried adding the capability myself as a resource action
> in the pcs configuration, I was able to add the action, but it would
> have no effect when run. Digging deeper, I eventually found the
> apache control script at /usr/lib/ocf/resource.d/heartbeat/apache and
> was able to see that there are no commands available for a graceful
> restart, only graceful stop. 
> 
> Currently, I have found two ways of getting around this limitation.
> The first is “/bin/kill -USR1 `cat /var/run/httpd.pid`", and the
> second is "/bin/test -f /var/run/httpd.pid && /usr/sbin/httpd -f
> /etc/httpd/conf/httpd.conf -c "PidFile /var/run/httpd.pid" -k
> graceful”. Both seem to work with out adversely affecting corosync.

Hi,

apachectl (apache2ctl on some distros) is an easier interface, i.e.
"apachectl graceful"

> My primary question is; what is the reason that graceful restart is
> not supported by the heartbeat apache control script? It seems like a
> pretty common usage of apache, so it seems a strange omission from
> the script. I am concerned that I’m trying to work against the grain
> since it is absent from the script.

Like an init script, an OCF script is mainly intended for the system
(the cluster in this case) to bring a service up or down, and then
secondarily as a user interface for doing such actions. The cluster
won't ever need or use a graceful restart command, so there hasn't been
a demand for an OCF interface to it.

I would just use apachectl for reloads and graceful restarts. Neither
should affect pacemaker's monitoring (as long as they succeed). (By
contrast, you shouldn't use apachectl to do a full start or stop if
apache is being controlled by the cluster, because the cluster will see
that as a failure.)

A side note: there is a "reload" action supported by some OCF agents,
but its use is muddled in pacemaker right now. There is an effort to
clear that up, after which we will likely implement a user interface
for natively reloading a service. It would make sense for the apache
agent to use a graceful restart for that action. So, in the longer
term, there might be a way to do what you want via cluster tools.

> My second question is; are the two work arounds presented above safe
> for use with corosync/heartbeat/pcs? Is one preferable to the other?
> If neither are safe, how would you recommend being able to update
> apache configurations with no downtime?
> 
> I am new to this list and could not find a way to search the
> archives, so if this question has already been answered, could you
> point me to the search area and to the answer as well?
> 
> Thank you in advance for your advice and recommendations.
> -Cole
-- 
Ken Gaillot <kgaillot at redhat.com>