Now only basic variant of summary metric is implemented - with sum and count of observations.
What's left:
[ ] Quantiles;
[ ] Configurable sliding window
Prometheus summaries usually use highly biased quantile and since we deal with streaming data so they always implement the same algorithm from paper [0]. Example implementations: [1-3].
Sliding window seems to be implemented using N numbers of age-overlapped summaries maintained in parallel. Of course only 'top' one is rendered during scraping.
Due to the huge overhead maybe it's good idea to keep current simple implementation as an option?
So I found this: https://github.com/odo/quantile_estimator
It's rather old but seems to build, at least on OTP-21.
I'll take a look if it's possible to integrate to the summary metric.
Now only basic variant of summary metric is implemented - with sum and count of observations.
What's left:
Prometheus summaries usually use highly biased quantile and since we deal with streaming data so they always implement the same algorithm from paper [0]. Example implementations: [1-3].
Sliding window seems to be implemented using N numbers of age-overlapped summaries maintained in parallel. Of course only 'top' one is rendered during scraping.
Due to the huge overhead maybe it's good idea to keep current simple implementation as an option?
References: