[Hawkular-dev] Integrating APM data into H-Metrics

Jay Shaughnessy jshaughn at redhat.com
Mon Jan 9 11:02:41 EST 2017 

For those of us not overly familiar with APM let me see if I can 
summarize what I've read so far and see if I'm close...

Trace:  Made up of a set of spans, a trace represents a high level unit 
of work, probably a complete piece of business logic, or service 
endpoint.  And whose duration is likely more reflective of 
user-experience, like a response-time.  As such it would seem like if a 
trace duration is within expectation then there isn't too much to worry 
about and the spans making up that trace are likely not overly relevant, 
unless you want to look at response-time tuning and improvement.   But, 
a "kpi" (Key Performance Indicator, right?) tag on a trace could in 
general be useful, as you may want to be able to establish a baseline 
and perform alerting on durations that reflect user experience.

If the above is fairly on target it would seem like storing trace 
duration metrics would be a good idea but storing span duration metrics 
would be a waste, until such time that the trace is not well behaved.  
It said in one of the replies that all spans were stored in ES, I assume 
this is just for a brief time to support some ad-hoc analysis in the 
short term.

As for long term storage I'd agree that it should be user-directed, tags 
on the APM side to decide what data to store seems like a good idea, 
especially at the trace level. And if the trace behaves badly (perhaps 
based on an alert) then spans for that trace perhaps should get tagged 
(at least until the issue is resolved).

I don't really know why the solution requires ES but likely it was just 
available  and easy to use.  If we can start to incorporate Hmetrics 
that seems good, a complete migration to Hmetrics seems better ;-)   As 
Gary mentioned, certainly there is a benefit of having Halerting built 
in for use with metrics being stored.   From an alerting perspective 
we're also working with Gary/APM to perform alerting on APM events, 
which could also potentially trigger some automated tagging.



On 1/9/2017 5:10 AM, Gary Brown wrote:
> Apologies I didn't set the context better :)
>
> There are a couple of reasons to investigate integration with H-Metrics.
>
> 1) Long term storage of metrics, as mentioned - primarily because we are currently collecting all information for use in short term analysis. If some of that information is of long term interest to an end user, then having the ability to identify the relevant information is necessary. Two options are (a) mark the data at the point it is generated (i.e. tag it), or (b) define server side configuration to extract the relevant data. Option (a) just seems the most straightforward.
>
> 2) Integration with H-Alerts for free. So although we can (and have) integrated directly, it may be better from an integrated architecture perspective to feed the metrics through H-Metrics.
>
> Regards
> Gary
>
> ----- Original Message -----
>>
>> On Mon, Jan 9, 2017 at 10:05 AM, Gary Brown < gbrown at redhat.com > wrote:
>>
>>
>>
>>
>> ----- Original Message -----
>>> Hello,
>>>
>>> let me start with terminology explanation:
>>> Span is one unit of work. It can be a method/REST endpoint invocation or
>>> just
>>> any unit of work.
>>> List of spans form a trace which is currently tree like structure(every
>>> span
>>> has its parent, except root span).
>>> Span captures its start and end, therefore it is possible to calculate a
>>> duration.
>>>
>>> By default all spans reported to the server are persisted in elasticsearch.
>>>
>>> I think there are two different things in this email:
>>>
>>> 1. OpenTracing standard "kpi" tag. Tags are key-value pairs, so I guess
>>> value
>>> would be a metric name (e.g. "user-registration": duration of user
>>> registration which consist of multiple service calls). I think that any
>>> duration is considered as KPI. Typically when users want to show durations
>>> for a particular service they can use filtering by tags(custom/standard).
>>> Is
>>> there really need for KPI tag?
>> This is about long term metric storage and analysis, rather than short term
>> user based filtering. So if we want to be able to identify the key metrics
>> that are of long term interest, it would need to either be done through some
>> configuration on the server, or as suggested in this proposal by allowing
>> the application to tag the relevant spans.
>>
>>
>> Long term storage wasn't mentioned in the original email. I ask for a better
>> explanation "KPI" tag. If there is a really need for it.
>>
>>
>>> 2. integration with h-metrics. What are the benefits for users? Integration
>>> with grafana? Cannot be grafana directly connected to elastic APM is using?
>>> All durations (span, trace) are stored in our backend and published to JMS
>>> so an integration with other systems should be possible.
>> Long term storage of specific metrics. APM is collecting a large amount of
>> information, so it is likely that it will only be retained for a limited
>> time frame.
>>
>>
>> I think that tracing tool is mostly used for a real time analysis and
>> debugging or tool for solving performance problems and testing
>> "canary"/"blue-green" deployments, but long term storage can useful or have
>> an access to aggregated results from older periods of time. Statistics
>> (average/median) of service response times. We should probably look at
>> zipkin ML and look if somebody was looking for this feature.
>>
>> It would actually make sense to provide some kind of reports as APM does
>> higher level aggregations. Question is if there is an interest of this
>> feature. I'm not very familiar with grafana but from what I read it can
>> connect to elasticsearch so maybe it can be used with APM to show dasboards
>> and reports.
>>
>>
>>
>>
>> Regards
>> Gary
>>
>>>
>>> Regards,
>>>
>>>
>>> On Sat, Jan 7, 2017 at 5:09 PM, John Sanda < jsanda at redhat.com > wrote:
>>>
>>>
>>> I want to make sure I understand the difference between trace and span. A
>>> trace captures the call between two services whether those are REST
>>> endpoints, EJBs, etc. Let’s say we want a trace of A calling B. Within that
>>> A calls C which in turn calls D, etc. and then finally B is called. Would a
>>> span measure the duration of the call between C and D?
>>>
>>> Are all spans captured/recorded by default?
>>>
>>> Would this be optional functionality? I am wondering because it obviously
>>> requires having Hawkular Metrics running which also means Cassandra.
>>>
>>> Under what tenant would you store these metrics?
>>>
>>>> On Jan 7, 2017, at 7:36 AM, Gary Brown < gbrown at redhat.com > wrote:
>>>>
>>>> Hi
>>>>
>>>> Wanted to discuss a proposal for recording some metric data captured from
>>>> Hawkular APM in Hawkular Metrics.
>>>>
>>>> For those not familiar with Hawkular APM, it captures the end to end
>>>> trace
>>>> instance (think of it as a distributed call stack), for each invocation
>>>> of
>>>> an application. This trace can include information about the
>>>> communications between services, but can also include details about
>>>> internal components used within the services (e.g. EJBs, database calls,
>>>> etc).
>>>>
>>>> First point is that if we were to record duration metrics for each 'span'
>>>> captured (i.e. scope within which a task is performed), for each
>>>> invocation of an application, then it would result in a large amount of
>>>> data that may be of no interest. So we need to find a way for end
>>>> users/developers to express which key points within an application they
>>>> do
>>>> want recorded as metrics.
>>>>
>>>> The proposal is to allow the application/services to define a
>>>> tag/property
>>>> on the spans of interest, e.g. 'kpi', that would indicate to the server
>>>> that the duration value for the span should be stored in H-Metrics. The
>>>> value for the tag should define the name/description of the KPI.
>>>>
>>>> If considered a suitable solution, then we can also propose it as a
>>>> standard tag in the OpenTracing standard.
>>>>
>>>> There are a couple of metrics that we could record by default, first
>>>> being
>>>> the trace instance completion time, and the second possibly being the
>>>> individual service response times (although this could potentially also
>>>> be
>>>> governed by the 'kpi' tag).
>>>>
>>>> Thoughts?
>>>>
>>>> Regards
>>>> Gary
>>>>
>>>>
>>>> _______________________________________________
>>>> 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
>>>
>>>
>>>
>>> --
>>> Pavol Loffay
>>> Cell: +421948286055
>>> Red Hat
>>> Purkyňova 111 TPB-B
>>> 612 45 Brno
>>>
>>> _______________________________________________
>>> 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
>>
>>
>>
>> --
>> Pavol Loffay
>> Cell: +421948286055
>> Red Hat
>> Purkyňova 111 TPB-B
>> 612 45 Brno
>>
>> _______________________________________________
>> 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

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/hawkular-dev/attachments/20170109/a43353b1/attachment-0001.html

More information about the hawkular-dev mailing list