Metric standard SDK behavior specification

[jmacd]

In #250 we've tried to steer away from specifying the behavior of the library (as there are so many possibilities) and instead focus on the semantics of the events.

It's become clear that people are confused about the behavior of the SDK because it's literally unspecified as part of the API. I believe we should prepare a companion specification to explain what the initial OpenTelemetry SDK is targeting for metrics export. This could go as far as describing an API to configure views and aggregations, or a YAML syntax to support standard configuration.

For a prometheus exporter, we need to decide on certain behaviors that are "not semantic" and therefore haven't been described in the API specification. For each of the 6 sub-kinds of metric, we need to choose an export type for Prometheus. For example, our monotonic counter becomes a prometheus counter, but our non-monotonic counter needs a firm mapping. For measure metrics, I believe we should choose a histogram metric export w/ default bucket ranges. I could imagine making the choice of Histogram vs. Summary metric be an option for the API or YAML syntax mentioned above, with support for configurable bucket ranges and configurable quantiles.

[bogdandrutu]

I am confused "non-monotonic counter needs a firm mapping", this is the Prometheus Gauge and probably that's why people are also confused about our definition of Gauge.

[jmacd]

1. OTel monotonic Counter->Prometheus Counter
2. OTel non-monotonic Counter->Prometheus Gauge using its Add() and Dec() methods
3. OTel monotonic Gauge->Prometheus Gauge using its Set() method (?)
4. OTel non-monotonic Gauge->Prometheus Gauge using its Set() method
5. OTel non-negative Measure->Prometheus Histogram
6. OTel signed Measure->(not sure)

@bogdandrutu what do you think about case 3?

[bogdandrutu]

7. Observer non-monotonic -> Prometheus Gauge
8. Observer monotonic -> Prometheus Counter

For 3 that is a Prometheus Counter.

Everything that is monotonic is a Counter in Prometheus. We have two options to implement keep the last value and record delta if we want to use their main API or implement our own registry in Prometheus that gathers the right metricsfamilies.

[bogdandrutu]

Also Prometheus is moving to a non-negative histogram/summary based on OpenMetrics.

[jmacd]

"For 3 that is a Prometheus Counter". The problem is that Prometheus Counter doesn't support Set(), only Add(). To turn a monotonic Gauge into a Prometheus Counter we have to remember the last value.

[bogdandrutu]

Correct, but we need to understand the difference between API and data model in Prometheus. From data model perspective that is a Counter and in order to achieve this as I mentioned in the previous post we can use a Registry instead of using directly the API. This is how they also do it for another counter cpu-usage.

[bogdandrutu]

Sorry I used the wrong term "Registry". We will use Collector

Here is an example for the cpu-usage and few other metrics: https://github.com/prometheus/client_golang/blob/3525612fea19680dd3d49c768802ec082ca853b2/prometheus/process_collector.go

[jmacd]

Cool, this makes sense.

[bogdandrutu]

Here is a table that tries to summarize the discussion so far:

OpenTelemetry	Prometheus
Monotonic Counter	Counter
Non-Monotonic Counter	Gauge
Monotonic Gauge	Counter
Non-Monotonic Gauge	Gauge
Monotonic Observer	Counter
Non-Monotonic Observer	Gauge
Signed Measure	Histogram or Summary
Non-Negative Measure	Histogram or Summary

[c24t]

I believe we should prepare a companion specification to explain what the initial OpenTelemetry SDK is targeting for metrics export

It sounds to me like this should just be part of the spec, in the same way that sdk-tracing describes the span exporter interface now.

If the argument for keeping aggregation and export logic out of the API package is the same as for tracing then it seems like the spec would be incomplete without a description of the SDK. We've effectively got two APIs, one for library devs in the "API" package, and another for vendors in the "SDK" package.