[ 
https://issues.apache.org/jira/browse/MESOS-1458?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14019079#comment-14019079
 ] 

Kevin Sweeney commented on MESOS-1458:
--------------------------------------

First, a car analogy. The accelerometer is usually more useful than the 
odometer, even though you can theoretically compute the reading on the 
accelerometer by sampling the odometer on a tight enough interval. I want the 
snapshot of the dashboard (what I get from hitting the metrics endpoint) to 
contain both an odometer (counter) and an accelerometer (precomputed rate) 
reading.

Back to computers, I do want the ability to use the endpoint as a rudimentary 
{{top}}. Computing the aggregate closer to the source of the data has the 
advantage of being simpler to reason about and lets a snapshot have independent 
meaning. Also, while it is true that a polling interval needs to be exposed, I 
think a hardcoded interval of something like 1sec [1] (or something more 
empirically calculated) would be fine for this use case.

[1] The framework we use in Aurora makes it configurable - 
https://github.com/twitter/commons/blob/master/src/java/com/twitter/common/application/modules/StatsModule.java#L66-L68

> Restore "cpus" metric in addition to cpu_*_secs counters in container stats
> ---------------------------------------------------------------------------
>
>                 Key: MESOS-1458
>                 URL: https://issues.apache.org/jira/browse/MESOS-1458
>             Project: Mesos
>          Issue Type: Story
>          Components: isolation, slave
>    Affects Versions: 0.18.0
>            Reporter: Kevin Sweeney
>
> While a "cpu-seconds per second" metric can be computed with 2 samples of 
> cpu_*_secs at different times, it's often convenient to have a single 
> self-contained value, for example if sample collection interval is long, or 
> so that an operator can glance at the cpu usage of a task with a single curl 
> command.
> Add user_cpus, system_cpus, and total_cpus metrics to the available container 
> stats.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Reply via email to