I am not sure what is best solution here but same test passes in Boron, so this 
may be seen as a regression :(


> On Apr 25, 2017, at 2:17 AM, Jozef Bacigál <[email protected]> 
> wrote:
> 
> Hi Luis,
> 
> I did some investigation to the bug 6459 [1]. First of all before it was 
> working in some strange way. From the description from you the test is 
> running like this:
> - Connect switch to 1 node out of the 3 that conforms the cluster.
> - Kill the karaf process of that node.
> - Check switch operational info (inv & topo) is still available where there 
> is no switch connected to cluster anymore.
> - Disconnect switch from killed node.
> - Start killed node.
> - Check switch operational info (inv & topo) is still available where there 
> is no switch connected to cluster anymore.
> 
> It means we are running cluster with several controllers but connect one 
> switch to one controller. This controller will create a record in DS inv for 
> this device. Then you kill controller, thats mean plugin maybe get or maybe 
> not get the event of switch disconnection but anyway we cant delete it from 
> DS because the controller is already down. Other instances of plugin in 
> cluster do not know about this switch, it was not connected to them. 
> Afterward you disconnect the switch, plugin, controller already down. Then 
> you start the controller again. Plugin could search the DS for any device 
> which is in DS inv but not connected, but plugin doesn't know if that switch 
> is not created by another cluster node or even if plugin is in cluster at 
> all. So plugin can't delete it anyway. 
> 
> Before is was correct because we deleted all nodes from DS inv when 
> disconnected, this leaded plugin in unstable state in case we deleted the 
> switch from DS after another cluster node was created/updated it in DS and 
> started working. 
> 
> So there is two possibilities in my mind. 
> - Plugin clear all nodes immediately (which may or may not solve this 
> problem, if cluster went down quicker the we get the info about disconnecting 
> and may lead to the bug when plugin deleted working node for another cluster 
> node). 
> - We accept that there is not possibility to clear inv if switch was 
> connected only to one node and cluster went down unexpected.
> 
> Guys if you got any other idea let my know, need to fresh thoughts 
> <OutlookEmoji-😊.png>
> 
> Jozef
> 
> 
> 
> [1] https://bugs.opendaylight.org/show_bug.cgi?id=6459 
> <https://bugs.opendaylight.org/show_bug.cgi?id=6459>
_______________________________________________
openflowplugin-dev mailing list
[email protected]
https://lists.opendaylight.org/mailman/listinfo/openflowplugin-dev

Reply via email to