Add histogram for response time percentiles

[JorritSalverda]

The current response time summary can't be aggregated over multiple machines so it's hard to graph response time percentiles. A histogram with configurable buckets would be very useful to add. See https://prometheus.io/docs/practices/histograms/

[mateusz-jablonski94]

Hi,

We decided to use summary metrics, because computing quantiles is done on client-side.
That results in low server-side cost, which was our main priority.
Thanks for the feedback, we will think about it.

[JorritSalverda]

According to https://www.robustperception.io/how-does-a-prometheus-histogram-work the histogram is actually cheaper than a summary, although the use of the word client in the blog post is confusing.

In addition to being aggregatable, histograms are cheaper on the client too as counters are fast to increment.