Document convention that Counter metrics shouldn't end with .count?

[trask]

This isn't mentioned in https://github.com/open-telemetry/semantic-conventions/blob/main/docs/general/naming.md#naming-rules-for-counters-and-updowncounters, but the convention has been followed pretty closely to only use .count suffix on UpDownCounter metrics (and not to use it on Counter metrics).

I'm not thinking of reasons right now, maybe related to the ambiguity of Counters being commonly exported as both delta and cumulative, and maybe related to the Prometheus exporter converting these to _count_total(?)

There are just a few metrics that violate this convention:

• otel.sdk.span.ended.count
• otel.sdk.processor.span.processed.count
• otel.sdk.exporter.span.exported.count
• dotnet.thread_pool.work_item.count - this one is stable, can't rename it

One option for the otel.sdk.* metric names could be:

• otel.sdk.ended_spans
• otel.sdk.processor.processed_spans
• otel.sdk.exporter.exported_spans

cc @JonasKunz

[JonasKunz]

otel.sdk.ended_spans

Sounds good to me, but I think we should also rename otel.sdk.span.live.count -> otel.sdk.live_spans for consistency?

otel.sdk.exporter.exported_spans

Initially it was intended to have different namespaces per signal: E.g. otel.sdk.exporter.span.* otel.sdk.exporter.log.* etc. However I think it is okay to share metrics across signal types for exporters, see my at reasoning #1921.

So this change is fine with me, we might even remove the _spans suffix and introduce an attribute instead to distinguish the telemetry type.

`otel.sdk.processor.processed_spans``

Here I'm not that that happy about dropping the span in the namespace. There is no general concept of processor in the spec, there is just span-processors and log-processors, which are different things and evolve independently. This also shows with metrics: There is no MetricProcessor, just MetricReaders. Tehre's also some more reasoning given in #1921.

So I'm not opposing to removing the .count suffix, I'm just concerned about dropping span from the metric namespace turning it from "span processor" to just "processor"

[trask]

I'm just concerned about dropping span from the metric namespace turning it from "span processor" to just "processor"

ah, no problem, my only goal here was to get rid of the .count suffix 👍

[JonasKunz]

I've adjusted the names for log health metrics in #1921. Based on the outcomes of the naming discussions in that PR, I'll open up a follow-up PR to apply the same name pattern to the existing span health metrics.