Support exponential histograms in the prometheus bridge

[dashpole]

Changes

Support converting exponential histograms in the prometheus bridge. To use exponential histograms, users can enable the NativeHistogram... options when creating their prometheus histogram.

When prometheus native histograms are enabled, BOTH fixed-bucket and native histograms are available in the protobuf in a "combined" histogram format. Since we can only support one-or-the-other, we always use exponential histograms if they are available.

Alternatives

As an alternative, we could consider adding an option, like WithExponentialHistograms() to the bridge that a user would additionally need to enable to get exponential histograms. If the prometheus client decided in the future to enable native histograms by default, users of the bridge would suddenly start getting exponential histograms instead of fixed-bucket histograms. If their backend doesn't support native histograms, this would be breaking for them. If there was not a way to disable that behavior in the Prometheus client, they would be stuck.

I have not implemented it here because we can always add it later if needed.

Implementation references

This follows the inverse of this specification: https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/compatibility/prometheus_and_openmetrics.md#exponential-histograms.

It is also the inverse of the conversion here: https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/pkg/translator/prometheusremotewrite/histograms.go

Thankfully, OTel doesn't have a min or max scale, so we don't have to handle scale up/scale down.

Future work

Exemplars on native histograms are not yet supported in the prometheus client. They were added to the protocol here: prometheus/client_model#80, but hasn't been released yet, and hasn't been implemented in the prometheus go client yet.

@bwplotka do you think we can/should rely on NativeHistogram.* options in HistogramOpts to determine whether users get an exponential vs fixed-bucket histogram? Or should we introduce our own option on the bridge to be safe.

[codecov]

Codecov Report

Attention: 13 lines in your changes are missing coverage. Please review.

Comparison is base (dabfd13) 61.0% compared to head (c721ec7) 61.3%.
Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##            main   #5093     +/-   ##
=======================================
+ Coverage   61.0%   61.3%   +0.2%     
=======================================
  Files        185     185             
  Lines      13760   13882    +122     
=======================================
+ Hits        8405    8511    +106     
- Misses      5160    5171     +11     
- Partials     195     200      +5     

Files	Coverage Δ
bridges/prometheus/producer.go	89.0% <89.6%> (+0.1%)	⬆️

... and 1 file with indirect coverage changes