Per-instrument metric resources: OK or non-standard?

[jmacd]

I was reflecting on how we handle Resources in the metrics pipelines and we haven't added support in the exporters yet. I noticed that we support resources on a per-instrument basis or fall-back to the SDK's resources. I wonder if this is "off spec"? I would note that Prometheus supports "constant labels" which is roughly the same, so maybe it's fine.

The question came about because I'm trying to figure out how to cache the encoding of a resource value in an exporter.

[jmacd]

One of the glaring points is that the api/metric package is importing the sdk/resource package. The API shouldn't have this dependency, so either we move Resource into the API or remove per-instrument resources. I think the spec would say we shouldn't have resources in the API.

[jmacd]

@bogdandrutu Please review.

[bogdandrutu]

Currently based on the specs there is one resource per SDK instance (metric provider), see here https://github.com/open-telemetry/opentelemetry-specification/blob/master/specification/resource/sdk.md#resource-sdk.

[jmacd]

This will happen after #640 is re-done following #651.

[jmacd]

The same has been done in the OTLP protocol.