[Hawkular-dev] metrics on the bus

[Michael Burman]

Hi,

Every inserted metric has to be processed inside metrics anyway, so it's an event for us. And we can't tell if some system wants it. The filtering has to be done somewhere and it's certainly not the metrics' job and event bus shouldn't work in that way either. Event bus should be done the one doing the filtering and routing, not the components producing the data. Why would we do the things backwards? We should remove the whole metrics-bus component if we do something on the metrics side - it has no purpose anymore in any sane architectural view.

There's no reason to push this work to metrics without removing hawkular-bus. It makes more sense to integrate alerts to metrics directly then - and I thought we wanted separated components.

Is there a reason we can't use known-to-work-architecture?

https://github.com/google/guava/wiki/EventBusExplained

If we add the functionality to metrics, it should happen with this (as we already have Guava and vertx isn't an option). And then it should remove hawkular-bus as it's been replaced.

  - Micke

----- Original Message -----
From: "Jay Shaughnessy" <jshaughn at redhat.com>
To: hawkular-dev at lists.jboss.org
Sent: Tuesday, August 9, 2016 4:12:24 PM
Subject: Re: [Hawkular-dev] metrics on the bus


An event bus is for publishing events and I'm not sure every datum is an event of any significance. The storage of an interesting datum could be an event, and what is interesting can't be determined by the storage mechanism unless told. It's not unlike adding a trigger to a db, or registering a listener to be informed of significant happenings among a larger set of happenings. So, to a degree I agree with Micke, publishing all events to the bus is good. But that leaves the semantic of choosing what an event is, and I'd say every datum is not an event. And I'd also say that just because a bus can handle millions of events, there is a practicality to not doing a bunch of unnecessary work, and instead leaving that bandwidth available to be used for solving actual business problems. 

I agree with Heiko that there are drawbacks to the polling approach (as usual). But it has the advantage of not needing any further work on the metrics side. But since Metrics already has hooks for bus publishing in hawkular, perhaps it wouldn't be hard to add some selection as well. I'm also fine with some sort of listener approach that leverages the co-location of components. Maybe some sort of Observer. Perhaps Micke or other metrics guys could comment on that possibility. 


On 8/9/2016 6:18 AM, Michael Burman wrote: 



Hi,

Like I said in the IRC, I don't like either approach and both sort of piss on the well known EIP pattern. Since this problem has been solved ages ago, why are we going to reinvent it? Use a bus, and I mean - a real bus. Created by us, or someone else.

onMessage(msg) {
 if(subscriptionMap.contains(msg.getTarget())) {
   subscriptionMap.get(msg.getTarget()).forEach(sub -> sub.writeMsg()));
 } 
}

addSubscriber(Subscriber sub, String target) { subscriptionMap.add(target, sub); }

Millions of msgs per second should be quite normal performance per node. Add thread notify & waits and have fun with dispatcher/multiplexer nirvana . VertX/Guava/Hazelcast/etc have event busses also, but in the end this is very simple process and should be part of the bus - not part of the metrics.

Components should not create the event bus, that should have been the whole point of hawkular-bus. If it can't do that, then it needs to be replaced.

  - Micke

----- Original Message -----
From: "Jay Shaughnessy" <jshaughn at redhat.com> To: "Discussions around Hawkular development" <hawkular-dev at lists.jboss.org> Sent: Tuesday, August 9, 2016 12:00:24 AM
Subject: [Hawkular-dev] metrics on the bus


Lucas and I were talking over jira [1] which has to do with metrics/alerting scale. This was discussed a bit on IRC recently as well. Today, metrics publishes all datapoints to the bus (metrics and avail go to different topics). The only consumer of that data is alerting, and it consumes a small fraction of the total data (actually it consumes none of it OOB at the moment, but that will hopefully change as Lucas's alerting work comes on line in MIQ). 

Although in its purest form this publish-it-all is the essence of bus publishing, we both feel it's an unnecessary waste of resources, as metrics can reach very high volume. There are a few approaches to reducing the publishing/filtering that we're currently doing. The options we discussed boil down to: 


    * No Publishing 


        * Just query metrics for the data needed for alerting (or whatever other external use we may have for the data) 
        * This is essentially a polling approach with frequent polling 
    * Demand Publishing 


        * The "just tell me what movie you want to see" approach 
        * Let clients request the metric ids it wants published to the bus 


I'm purposefully not going into much detail at this point. I'd rather we talk out a preferred approach between these two, or something not presented. But we'd like to move away from the current publish-it-all approach. 
[1] https://issues.jboss.org/browse/HWKALERTS-118 . 

_______________________________________________
hawkular-dev mailing list hawkular-dev at lists.jboss.org https://lists.jboss.org/mailman/listinfo/hawkular-dev _______________________________________________
hawkular-dev mailing list hawkular-dev at lists.jboss.org https://lists.jboss.org/mailman/listinfo/hawkular-dev 


_______________________________________________
hawkular-dev mailing list
hawkular-dev at lists.jboss.org
https://lists.jboss.org/mailman/listinfo/hawkular-dev