Define conversion mapping from OTel Exponential Histograms to Prometheus Native Histograms.

[kovrus]

Fixes #
#3078

Changes

Define conversion mapping from OTel Exponential Histograms to Prometheus Native Histograms. See, the implementation open-telemetry/opentelemetry-collector-contrib#17370

Related issues #

open-telemetry/opentelemetry-collector-contrib#16207
open-telemetry/opentelemetry-proto#440

Related OTEP(s) #
https://github.com/open-telemetry/oteps/blob/main/text/0149-exponential-histogram.md

[kovrus]

@dashpole @Aneurysm9 @gouthamve please take a look.

[jsuereth]

Note: I approve of the conversion, however I would like to see us attempt to rescale histograms as best we can to fit prometheus's current boundaries/limitations.

One of the goals of these histograms is to easily/efficiently rescale so we can balance performance + resolution. In that vein, I'd be a fan of something like:

If `Scale` is > 8 then  exponential histogram data points SHOULD be downscaled to a scale accepted by prometheus (in range [-4,8]).  Any data point unable to be rescaled to an acceptable range MUST be dropped.

This is a non-blocking request though.

[jack-berg]

however I would like to see us attempt to rescale histograms as best we can to fit prometheus's current boundaries/limitations.

We can reduce the scale if its greater than 8, but how would we change the scale if its less than -4? One option would be to introduce a MinScale config parameter such that the prometheus exporter can request that the SDK never rescale below -4. But what does it do if at scale -4 the measurements don't fit into the max number of allowed buckets? We could introduce a reset. When scale is at -4 and its still not enough to fit the measurements in the allowed buckets, trigger a reset. This isn't without problems though- what the measurements don't fit in the allowed buckets at scale -4 within a single collection interval? Resetting would cause data loss. We could introduce an infinity_bucket (similar to the zero_bucket), and increment infinity_bucket and zero_bucket when measurements don't fit within the allowed buckets at the allowed scale. Would have to populate zero_threshold and infinity_threshold with the smallest and largest respective values that are representable given the scale and allowed buckets.

[github-actions]

This PR was marked stale due to lack of activity. It will be closed in 7 days.