kit
kit copied to clipboard
go-kit
metrics/dogstatsd: Add support for Distribution metrics
Datadog has a Distribution metric type that is not currently supported by go-kit metrics. This PR adds support for this metric type.
The Dogstatsd datagram format documentation does not list Distributions as supporting sampling rates, but Datadog's own dogstatsd client does submit a sampling rate, so I am assuming that it is supported.
References:
- https://github.com/DataDog/datadog-go/blob/v4.6.0/statsd/statsd.go#L570-L579
- https://github.com/DataDog/datadog-go/blob/v4.6.0/statsd/format.go#L139-L141
This is an isolated enough change that I felt opening an issue was not necessary, but if you disagree let me know and I will open one.
The purpose of the Go kit metrics packages is to offer a common interface of Counters, Gauges, and Histograms, which can abstract over different backends. If a backend supports more than those 3 types, that's fine, but why would the Go kit package need to also provide that support?
Perhaps I used the wrong term when I said "metric type". From go-kit's perspective, this is another implementation of the metrics.Histogram interface. To a package consuming this as a dependency, there would be no difference, but in Datadog one would be able to access features that are only provided by this new "Distribution" implementation.
This is similar to the prometheus package, which offers NewHistogram() and NewSummary() as implementations of metrics.Histogram. The dogstatsd package also offers two implementations via NewHistogram() and NewTiming().
Oh, sorry, I didn't understand it was a Histogram. Let me look closer.
No problem.
I admit, this is basically a copy+paste of the Histogram implementation, with a d as the metric type submitted in the payload to dogstastd instead of h. If you would like, I am happy to try reduce some duplication (in this PR or another one).
Wait, is it literally identical?
Why... does Dogstat do something different with this exactly identical data?
Yeah, let's have a flag or something in the Histogram constructor? Wild.