Can you implement Hoeffding Bound function

[vkuznet]

Is your feature request related to a problem? Please describe.
We would like to perform statistical analysis of time series using Hoeffding Bound function, see https://www.youtube.com/watch?v=6UwcqiNsZ8U&feature=youtu.be&t=1237
Since it is general type function which can be applied to any metric it would be nice to have in VM.

[valyala]

@vkuznet , the commit 8b14572 adds the following functions:

• hoeffding_bound_upper(phi, m[d]) - returns upper bound for values from m time series over duration d for phi in the range [0..1] (phi=0 returns values equal to avg_over_time(m[d])).
• hoeffding_bound_lower(phi, m[d]) - returns lower bound for values from m time series over duration d for phi in the range [0..1] (phi=0 returns values equal to avg_over_time(m[d])).

For example, the following query would return data points for x, which exceed 99.9% Hoeffding bound calculated over data points from the last minute:

x > hoeffding_bound_upper(0.999, x[1m])

Could you verify whether these functions are properly implemented (see commit 8b14572 ) and whether they match your needs? You can build VictoriaMetrics from this commit using these instructions.

[vkuznet]

I was able to build, deploy and run few queries with this patch. Everything looks fine, but I don't have enough statistics in my dev Prometheus instance and for production we need an official docker image. Can you build and tag one that we can use? So far, we used image: victoriametrics/victoria-metrics.

[valyala]

Can you build and tag one that we can use? So far, we used image: victoriametrics/victoria-metrics.

I'm planning to publish the next release with hoeffding_bound_* functions in the next few days.

[valyala]

@vkuznet , hoeffding_bound_* functions are available starting from v1.32.3.

Closing this issue as resolved. Feel free re-opening it if the current implementation of hoeffding_bound_* functions works incorrectly.