Allow exemplar-only metrics?

[LouisStAmour]

What are you trying to achieve?

I'd like to build a trace processor that can export real-time metrics but also follow up with sampled exemplars later. I can see how Prometheus could support this use case by publishing a list of exemplars with real-time metric data but if Otel Collector publishes metrics to another collector for export to Prometheus, it cannot do so using OTLP unless a convention is established to export exemplars without adjusting real-time metrics.

Additional context.

The equivalent feature would be an API call to AddExemplar directly without observing any values, which I've requested in the Prometheus Go Client community: prometheus/client_golang#868 That request also indicates what my use case is.

An alternative would be to not support publishing exemplars separately using OTLP and instead encourage clients receiving metrics to ignore the observed value contained in the metric if a flag is set or the date of the exemplar is before the date of the observation, perhaps.

It's also possible that rather than changing the spec, the Otel Collector could internally publish exemplars from trace processors without adjusting real-time metric data so that exemplars could be combined later into a proper Otel Metric with the real-time data sourced from a different metric. I could see this being the case with the Otel Collector Prometheus exporter that already knows how to combine multiple metrics to export them for Prometheus, it could use a hypothetical AddExemplar function to add the exemplar captured from a trace processor/sampler. There's an argument that publishing an exemplar could also include publishing the real-time metric used to compare and ensure the exemplar's value is an outlier. That's true, but it's extra data on the wire if you've separately published the real-time data and it's possible you're not looking at real-time data but actually the data at the timestamp associated with the exemplar - historical data - if kept in a circular buffer, say.