New metric type fixed histogram

[mixalturek]

It would be nice to have at least basic support for histograms. I know it's one of the non-goals, but the note was probably meant as expensive computation of various percentiles (which are in fact no histograms). https://en.wikipedia.org/wiki/Histogram

A histogram built on top of a small set of Counters with fixed predefined ranges should be quite simple to implement using API that Dipstick already provides and still quite powerful and with a great performance.

// Ranges and counters for: 'less', 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 'more'
let histogram = metrics.histogram("hist", Histogram::unit_step(0, 10));

// Ranges and counters for: 'less', 0-2, 3-5, 6-8, 9-11, 'more'
let histogram = metrics.histogram("hist", Histogram::const_step(0, 11, 3));

// Ranges and counters for: 'less', 0, 1, 2-4, 5-9, 10-49, 50-99, 100-999, 'more'
let histogram = metrics.histogram("hist", Histogram::steps(vec![0, 1, 2, 5, 10, 50, 100, 1000]));

What do you think? I'm ready to contribute, I'm asking in advance...

[fralalonde]

I know other libraries do it this way but I would not define a Histogram as a metric type, this is "premature configuration" i.e. the code defining the instrument should not have to care about what statistics will be tracked, not to mention what range of values are expected - all those things that could be very well put in a dynamic config file.

If Dipstick was to do histograms, it would be at the Bucket level. If you have a very fast way of doing it it could be an optional set of stats maintained by the AtomicBucket, or it could be new separate bucket type that can be joined to an AtomicBucket (if regular stats are also needed) using MultiInput. Or there could be a special multi-bucket wrapper to allow combining different buckets and have them act a single entity with regards to drain / flush.