Question About Histograms

[Anil K <se...@gmail.com>]

Hi,
I was doing some experimentation using Histograms, had a few questions
mostly related to fault tolerance and restarts. I am looking for a way to
calculate p95 over 30days. Since histograms are pushed as a summary into
prometheus, will not be able to do the aggregation for 30 days at
Prometheus' side from smaller windows. Also there is a high chance that the
job may restart in between.

So Are Histograms Checkpointed ? I am looking for a very large
SlidingTimeWindowReservoir of 30days and I don't want to lose the histogram
updates till then and start from scratch after restart.

If Histograms are not checkpointed and are reset, are there any other
possible ways to calculate Histograms/P95 on a large window like 30days?

Thanks & Regards
Anil

Re: Question About Histograms

[Anil K <se...@gmail.com>]

Hello Prasanna,
Thanks for your response,
Could you elaborate on what you meant by "overriding the Prometheus
Histogram class provided "? if possible with any samples?

Regards,
Anil

On Tue, Apr 5, 2022 at 1:11 AM Prasanna kumar <pr...@gmail.com>
wrote:

> Anil,
>
> Flink Histograms are actually summaries .. You need to override the
> Prometheus Histogram class provided to write it into different buckets to
> Prometheus .. Then you can write prom queries to calculate different
> quantiles accordingly ...  Checkpointing The histograms is not a
> recommended option/solution
>
> Thanks,
> Prasanna.
>
> On Tue, 5 Apr 2022, 01:26 Anil K, <se...@gmail.com> wrote:
>
>> Hi,
>> I was doing some experimentation using Histograms, had a few questions
>> mostly related to fault tolerance and restarts. I am looking for a way to
>> calculate p95 over 30days. Since histograms are pushed as a summary into
>> prometheus, will not be able to do the aggregation for 30 days at
>> Prometheus' side from smaller windows. Also there is a high chance that
>> the job may restart in between.
>>
>> So Are Histograms Checkpointed ? I am looking for a very large
>> SlidingTimeWindowReservoir of 30days and I don't want to lose the histogram
>> updates till then and start from scratch after restart.
>>
>> If Histograms are not checkpointed and are reset, are there any other
>> possible ways to calculate Histograms/P95 on a large window like 30days?
>>
>> Thanks & Regards
>> Anil
>>
>>
>>
>>
>>

Re: Question About Histograms

[Prasanna kumar <pr...@gmail.com>]

Anil,

Flink Histograms are actually summaries .. You need to override the
Prometheus Histogram class provided to write it into different buckets to
Prometheus .. Then you can write prom queries to calculate different
quantiles accordingly ...  Checkpointing The histograms is not a
recommended option/solution

Thanks,
Prasanna.

On Tue, 5 Apr 2022, 01:26 Anil K, <se...@gmail.com> wrote:

> Hi,
> I was doing some experimentation using Histograms, had a few questions
> mostly related to fault tolerance and restarts. I am looking for a way to
> calculate p95 over 30days. Since histograms are pushed as a summary into
> prometheus, will not be able to do the aggregation for 30 days at
> Prometheus' side from smaller windows. Also there is a high chance that
> the job may restart in between.
>
> So Are Histograms Checkpointed ? I am looking for a very large
> SlidingTimeWindowReservoir of 30days and I don't want to lose the histogram
> updates till then and start from scratch after restart.
>
> If Histograms are not checkpointed and are reset, are there any other
> possible ways to calculate Histograms/P95 on a large window like 30days?
>
> Thanks & Regards
> Anil
>
>
>
>
>