Hi guys, /wrt the GCM Topics support PR https://github.com/aerogear/aerogear-unifiedpush-server/pull/626 me and summersp have discussed how JMS should be used to route messages. Current implementation loads tokens and conditionally sends either message with registration IDs or topics. There are two things yet to be solved: - topics can be used up to 1 million registrations, otherwise you have to fall back to enumerating registration IDs - implementation with one sender (GCMPushNotificationSender) is suboptimal - the utilization of the topics can be increased by prioritizing topic message sending first (covering multiple, potentially thousands registrations) - followed by sending registration IDs out That's why I suggested to split implementation to two JMS queues talking to two sender impls (e.g. GCMTopicSender and GCMRegistrationIdsSender). - first we send out topic based messages (for efficiency) - we collect those topics that fail and resend them for processing as registration IDs (fail over) - registration ids sending can start in parallel, but it can't end until we sent out all topics

Additionally we get an ability to configure these two message routes individually on the JMS level (better utilization, transact-ability, fail over).

What do you think? Cheers, ~ Lukas _______________________________________________ aerogear-dev mailing list aerogear-dev(a)lists.jboss.org https://lists.jboss.org/mailman/listinfo/aerogear-dev

On Wed, Aug 12, 2015 at 3:22 PM, Lukáš Fryč <lukas(a)fryc.eu&gt; wrote: > Hi guys, > > /wrt the GCM Topics support PR > https://github.com/aerogear/aerogear-unifiedpush-server/pull/626 > > me and summersp have discussed how JMS should be used to route messages. > > > Current implementation loads tokens and conditionally sends either > message with registration IDs or topics. > > There are two things yet to be solved: > > - topics can be used up to 1 million registrations, otherwise you > have to fall back to enumerating registration IDs > - implementation with one sender (GCMPushNotificationSender) is > suboptimal > - the utilization of the topics can be increased by prioritizing > topic message sending first (covering multiple, potentially thousands > registrations) > - followed by sending registration IDs out > > > That's why I suggested to split implementation to two JMS queues talking > to two sender impls (e.g. GCMTopicSender and GCMRegistrationIdsSender). > > - first we send out topic based messages (for efficiency) > - we collect those topics that fail and resend them for processing as > registration IDs (fail over) > - registration ids sending can start in parallel, but it can't end > until we sent out all topics > > I agree that we need a slip here - but does it (GCM) really work like that ? I was more thinking/guess that we have to do the math and keep the device nr. per topic - wondering now, summers, what actually happens if device 1.000.001 registers w/ the topic? Does GCM fail the device ?

> Additionally we get an ability to configure these two message routes > individually on the JMS level (better utilization, transact-ability, fail > over). > but generally, I think these are two completely different cases, therefore we need to different senders > > > What do you think? > > > Cheers, > > ~ Lukas > > _______________________________________________ > aerogear-dev mailing list > aerogear-dev(a)lists.jboss.org > https://lists.jboss.org/mailman/listinfo/aerogear-dev > -- Matthias Wessendorf blog: http://matthiaswessendorf.wordpress.com/ sessions: http://www.slideshare.net/mwessendorf twitter: http://twitter.com/mwessendorf _______________________________________________ aerogear-dev mailing list aerogear-dev(a)lists.jboss.org https://lists.jboss.org/mailman/listinfo/aerogear-dev

On Wed, Aug 12, 2015 at 3:36 PM, Summers Pittman <supittma(a)redhat.com&gt; wrote: > > > On Wed, Aug 12, 2015 at 9:31 AM, Matthias Wessendorf <matzew(a)apache.org&gt; > wrote: > >> >> >> On Wed, Aug 12, 2015 at 3:22 PM, Lukáš Fryč <lukas(a)fryc.eu&gt; wrote: >> >>> Hi guys, >>> >>> /wrt the GCM Topics support PR >>> https://github.com/aerogear/aerogear-unifiedpush-server/pull/626 >>> >>> me and summersp have discussed how JMS should be used to route messages. >>> >>> >>> Current implementation loads tokens and conditionally sends either >>> message with registration IDs or topics. >>> >>> There are two things yet to be solved: >>> >>> - topics can be used up to 1 million registrations, otherwise you >>> have to fall back to enumerating registration IDs >>> - implementation with one sender (GCMPushNotificationSender) is >>> suboptimal >>> - the utilization of the topics can be increased by prioritizing >>> topic message sending first (covering multiple, potentially thousands >>> registrations) >>> - followed by sending registration IDs out >>> >>> >>> That's why I suggested to split implementation to two JMS queues >>> talking to two sender impls (e.g. GCMTopicSender >>> and GCMRegistrationIdsSender). >>> >>> - first we send out topic based messages (for efficiency) >>> - we collect those topics that fail and resend them for processing >>> as registration IDs (fail over) >>> - registration ids sending can start in parallel, but it can't >>> end until we sent out all topics >>> >>> >> I agree that we need a slip here - but does it (GCM) really work like >> that ? >> I was more thinking/guess that we have to do the math and keep the >> device nr. per topic - wondering now, summers, what actually happens if >> device 1.000.001 registers w/ the topic? Does GCM fail the device ? >> > > As far as Google's docs say, the device will register with GCM but the > topics will fail to send on the server. > > Specifically when we send the post to Google we will receive a failure > messages TOO_MANY_DEVICES. See the GCM Sender in the PR for exact details. > ok - thanks, so in case of this error - nothing is delivered? or just the first 1M ?

> > >> >>> Additionally we get an ability to configure these two message routes >>> individually on the JMS level (better utilization, transact-ability, fail >>> over). >>> >> >> but generally, I think these are two completely different cases, >> therefore we need to different senders >> >>> >>> >>> What do you think? >>> >>> >>> Cheers, >>> >>> ~ Lukas >>> >>> _______________________________________________ >>> aerogear-dev mailing list >>> aerogear-dev(a)lists.jboss.org >>> https://lists.jboss.org/mailman/listinfo/aerogear-dev >>> >> >> >> >> -- >> Matthias Wessendorf >> >> blog: http://matthiaswessendorf.wordpress.com/ >> sessions: http://www.slideshare.net/mwessendorf >> twitter: http://twitter.com/mwessendorf >> >> _______________________________________________ >> aerogear-dev mailing list >> aerogear-dev(a)lists.jboss.org >> https://lists.jboss.org/mailman/listinfo/aerogear-dev >> > > > _______________________________________________ > aerogear-dev mailing list > aerogear-dev(a)lists.jboss.org > https://lists.jboss.org/mailman/listinfo/aerogear-dev > -- Matthias Wessendorf blog: http://matthiaswessendorf.wordpress.com/ sessions: http://www.slideshare.net/mwessendorf twitter: http://twitter.com/mwessendorf _______________________________________________ aerogear-dev mailing list aerogear-dev(a)lists.jboss.org https://lists.jboss.org/mailman/listinfo/aerogear-dev