Support Prometheus histograms

[RichiH]

Prometheus uses le, OTel uses ge bounded buckets. The two are mathematically incompatible and impossible to transform from one into the other.

[jmacd]

See also this current discussion about histograms in OTLP:

open-telemetry/opentelemetry-proto#226 (comment)

To be clear, that discussion has practically fallen apart, but it's a debate over an independent problem about locating the boundaries and encoding them compactly, not about >= vs <= question.

Whenever this has come up in the past, the issue gets sidelined with some technical argument about how histograms represent approximate a continuous probability distribution, not a discrete one, essentially to say that "it doesn't matter". I'm willing to say, at this point, that if "it doesn't matter" then OTLP should simply adopt the Prometheus "le" convention.

[RichiH]

Agreed. There's nuance between the four possible boundary choices, but they are effectively/roughly equivalent if applied consistently on floats. Integer or discrete values are a niche use case and arguably better served by other mechanisms depending on specifics.

My main concern is user trust if their metrics pipelines behave subtly differently and they need to write complex queries to re-cast data on the fly.

Using le would make the whole problem space simply go away.

[bogdandrutu]

I think this is the same as open-telemetry/opentelemetry-proto#258

[RichiH]

Agreed, it's the same issue. I don't know what the plan with temporary duplications and deduplication is.

[bogdandrutu]

I think this is fixed, since OTLP changed to use "le". Please reopen if that is not the case