Hi Aditya,

Thank you for the KIP.

A few questions:

NS1: how were the metrics picked? There are other latency metrics as
well, which are not in the list, for example
NS2: Do I understand it correctly that the histograms will be
available only via a custom exporter, and not via JMX? Should we
consider adding percentiles?
NS3: It looks like the KIP proposes cumulative histograms that never
reset. If we want to calculate percentiles, making them useful will be
a significant challenge. Applications can run for weeks and months. If
we keep counting without resets, a p99 calculated over a month isn't
very useful.
I understand the challenge with custom exporters, but maybe we can
rework the idea to let exporters get a predictable delta, so we don't
need to keep the history forever? Then the percentiles can be
calculated for the configured window, and the future KIP-714 work will
be easier as well.
NS4: Exponential(Geometrical) histograms have a nice property: each
bin at one scale is exactly two bins at the next finer scale. Right
now, the KIP proposes having a max, which is very hard to set
correctly for latencies. Have you considered not having a cap and
automatically adjusting the scale depending on the data?

On Mon, Oct 26, 2026 at 11:54 PM Aditya Kousik via dev
<[email protected]> wrote:
>
> Hi all,
>
> I'd like to start a discussion on KIP-1387, which adds a stat that counts 
> latency and size values into fixed bins and gives the bin counts to the 
> metrics reporter.
>
> KIP link: https://cwiki.apache.org/confluence/x/XwKDGw
>
> Client latency and size metrics today are an average and a maximum over a 
> 30-second window. We don’t get a view into how the values are spread, a p99 
> can't be derived from them. A percentile computed inside the client can't be 
> combined across clients or across windows whereas bin counts can. A reporter 
> can publish them as a Prom or OTel histogram, and a backend query can compute 
> any percentile, over any period across the whole fleet.
>
> The KIP adds DistributionStat, whose value is a DistributionSnapshot of 
> cumulative bin counts. Seven distributions are registered at INFO next to the 
> existing -avg and -max metrics, with more at DEBUG. Existing metrics don't 
> change, and JmxReporter skips distributions, so a client reporting over JMX 
> sees nothing new.
>
> Looking forward to your feedback.
>
> Thanks,
Aditya
>
>

Reply via email to