OM 2.0: Float valued histograms

[krajorama]

Problem Statement

Native and classic histograms currently only allow positive integer counters for the Count and Bucket values (aka bucket count) in the Histogram type and positive integer gauges in the GaugeHistogram.

There's a use case for storing and exposing (gauge) histograms that are a result of arithmetic operations. For example a recording rule might store the rate of observations, which will result in fractions, or any other arithmetic operations such as division, negation, etc.

Proposed solution

Allow fractions in the count and buckets in Histograms.
Allow fractions and negative values in count and buckets for GaugeHistograms.

Alternatives considered

Do nothing

Cannot expose valid use cases.

Create new metric types

What would be the added value?

Depend on negotiated feature

Do not allow fractions/negative values by default. Only if negotiated.
Maybe something to consider.

[dashpole]

Discussed at WG:

• Consensus: We need to support float-flavored classic histograms, native histograms, and summaries.
• Note: These will all be "complex values" after that change is made.

Options:

• Allow integer or floats in those fields
• Allow floats in the field, even if they are usually an integer
• Some other option that differentiates a "float-flavored" type from the "integer-flavored" type

Considerations:

• Floats are more expensive to parse and store.
• Sometimes the server uses the heuristic that if the count is an integer, then it must be an integer histogram.