Hi Vinod,
I checked the corresponding master code [1] and also the v1 api response
for GET_METRICS call.
Do you suggest adding another field called `type` or something into the
result, so instead of
{
"name": "master/cpus_total",
"value": 0.0
}
we have
{
"name": "master/cpus_total",
"value": 0.0,
"type": "gauge"
}
[1] https://github.com/apache/mesos/blob/master/src/master/http.cpp#L1359
On Mon, Jun 27, 2016 at 5:23 PM, Vinod Kone <vinodk...@apache.org> wrote:
> We can definitely distinguish counters and gauges in the upcoming v1 API.
>
>
> https://github.com/apache/mesos/blob/master/include/mesos/v1/master/master.proto#L265
>
>
>
> On Mon, Jun 27, 2016 at 5:09 PM, Benjamin Mahler <bmah...@apache.org>
> wrote:
>
> > It's worth doing because it has come up before [1] [2]. For our own
> > education, in your case why is the metrics collection system interested
> in
> > the type? Is it for optimization? Which metrics collection system are you
> > using?
> >
> > What are the types that the metrics collection system understands? The
> > types that seem to show up consistently are counter, gauges, and
> > histograms.
> >
> > [1] https://prometheus.io/docs/concepts/metric_types/
> > [2] http://docs.datadoghq.com/guides/metrics/
> >
> > On Mon, Jun 27, 2016 at 4:37 PM, Tuan-Anh Hoang-Vu <hvtuan...@gmail.com>
> > wrote:
> >
> > > Hi Vinod,
> > >
> > > Yes, my metric collection system need to distinguish between different
> > > metrics.
> > > Currently we hard coded all metric types into our system, but it's not
> > very
> > > scalable and error prone.
> > > For example I recently discovered that we misclassified `timestamp` as
> > > gauge instead of counter.
> > >
> > > On Mon, Jun 27, 2016 at 4:31 PM, Vinod Kone <vinodk...@apache.org>
> > wrote:
> > >
> > > > Hey Tuan. Can I ask what you are trying to accomplish here? Does your
> > > > metric collection system need to distinguish between different
> metrics?
> > > >
> > > > On Mon, Jun 27, 2016 at 10:35 AM, Tuan-Anh Hoang-Vu <
> > hvtuan...@gmail.com
> > > >
> > > > wrote:
> > > >
> > > > > Hello,
> > > > >
> > > > > I am planning to extend current metrics endpoint
> (/metrics/snapshot)
> > to
> > > > > only
> > > > > return a specific type of metrics (either counter, gauge or timer).
> > It
> > > > > would be great if I
> > > > > can get feedback from you if this is worth pursuing, or there is an
> > > > > existing solution,
> > > > > or my implementation plan is sound.
> > > > >
> > > > > Current /metrics/snapshot endpoint returns all metrics as key-value
> > > pairs
> > > > > for both
> > > > > master and agent. There is an additional timeout parameter for that
> > > > > endpoint.
> > > > > For backward compatibility, I want to add another parameter (`type`
> > or
> > > > > `metric_type`)
> > > > > that will take a value of (`counter`, `gauge` or `timer`) and the
> > > > endpoint
> > > > > will
> > > > > only return metrics of that type.
> > > > >
> > > > > My implement plan is to extend libprocess to support that,
> > > particularly:
> > > > >
> > > > > 1. Introduce a variable called `type` in metric.hpp
> > > > > 2. Extend constructors of Timer, Counter and Gauge classes to set
> > that
> > > > > variable
> > > > > 3. Extend Future<http::Response> MetricsProcess::_snapshot()
> function
> > > in
> > > > > metrics.cpp to support additional query parameter
> > > > >
> > > > > How does that sound? I would love to hear feedback from you!
> > > > >
> > > > > Thanks,
> > > > > Tuan.
> > > > >
> > > >
> > >
> >
>
