Support a cumulative "interpretation" of Unknown metrics

[dashpole]

Prometheus Unknown metrics are metrics that have no type information, such as metrics from a third party system. Prometheus doesn't actually make use of type information, so Prometheus users can treat Unknown metrics as either counters or gauges at query time.

Under the current proposed specification, open-telemetry/opentelemetry-specification#2266, OpenTelemetry will convert Unknown-typed metrics to OTLP Gauges. While this is arguably more useful than dropping metrics, it does not preserve the fact that the metric type was initially unknown, and may make it difficult in some backends for users to treat them as counters.

If possible, we should attempt to preserve the fact that these metrics were initially Unknown-typed.

[dashpole]

One idea presented by @Aneurysm9 at the prometheus-wg meeting today is to add a resource metric attribute as a way to indicate that the metric was original unknown as a convention. This would not require data model changes.

[Aneurysm9]

Not a resource attribute, but a metric attribute since the attribute would not necessarily apply to all metrics from that resource.

[damemi]

reviving this topic, is there any opposition to a proposal for how to address this? I am willing to put something together for the spec

[dashpole]

Already fixed by open-telemetry/opentelemetry-collector-contrib#32605