[mod_cluster-issues] [JBoss JIRA] (MODCLUSTER-427) mod_cluster can break stickiness for the first request on new child processes

[Aaron Ogburn (JIRA)]

    [ https://issues.jboss.org/browse/MODCLUSTER-427?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12999025#comment-12999025 ] 

Aaron Ogburn commented on MODCLUSTER-427:
-----------------------------------------

Ah, thanks, JF, you are right.  I hadn't looked into proxy_cluster_child_init() yet.  But checking it out, I see currently that it creates the proxy_cluster_watchdog thread, but doesn't do anything itself to fill the balancer.  The created proxy_cluster_watchdog thread will fill the balancer after its first two seconds of sleep though.

Further testing around this matched up.  If I waited 2+ seconds for a subsequent request, sticky sessions can work since the proxy_cluster_watchdog executed and filled the balancer in the new process.  If the next request comes in less than 2 seconds later, proxy_cluster_watchdog didn't execute yet and so the balancer is not yet filled in the new process, and sticky sessions are not maintained as mentioned in my original description.

I modified my PRs to address proxy_cluster_child_init() since it's the origin of the issue.   I essentially added the update_workers_node calls here to ensure it is done within init rather than lazily 2 seconds later when the created proxy_cluster_watchdog first executes.  This fixes my reproduction as well.  What do you think?

> mod_cluster can break stickiness for the first request on new child processes
> -----------------------------------------------------------------------------
>
>                 Key: MODCLUSTER-427
>                 URL: https://issues.jboss.org/browse/MODCLUSTER-427
>             Project: mod_cluster
>          Issue Type: Bug
>      Security Level: Public(Everyone can see) 
>          Components: Native (httpd modules)
>    Affects Versions: 1.2.9.Final, 1.3.1.Alpha1
>         Environment: JBoss EAP 6.3.0
>            Reporter: Aaron Ogburn
>            Assignee: Jean-Frederic Clere
>             Fix For: 1.2.10.Final, 1.3.1.Final
>
>
> mod_cluster can break stickiness for the first request on new child processes.  It looks like this occurs specifically when "CreateBalancers 1" is used.  Prefork typically makes this much worse as well.
> My debugging showed that the proxy_balancer would exist, but it would essentially be empty in the new child process.  find_session_route/find_route_worker would be called as expected, but the for loop in find_route_worker wasn't even doing anything because balancer->workers->nelts was 0.  The balancer would then finally be populated in the new child when the first request hits internal_find_best_byrequests and calls update_workers_node.



--
This message was sent by Atlassian JIRA
(v6.3.1#6329)