Impl EncodeLabelValue for Option<T>

[emschwartz]

Alternatively, the EncodeLabelSet derive macro could add code that checks whether the value of the Option and only encodes the Some value

[drbrain]

I've implemented an equivalent of this where I use this crate to allow metric_name{option_label=""}, It'll be great to have this as a built-in.

For derived EncodeLabelSet I think this could produce either metric_name{} (a None means "no label at all") or metric_name{option_label=""} (None means "label with empty value") depending on if you wanted a label to be joinable with other metrics or not. I've done some work in #136 which might make it easier to configure the derive macro to choose the desired behavior.

[mxinden]

Can you expand on the use-case for a metric with a label key but an empty string as label value? Ideally the use-case and the corresponding Prometheus query?

Thank you for your upstream work!

[emschwartz]

I'm working on the Autometrics project, which generates metrics from function names.

We have a feature where you can define Service-Level Objectives in your code and then attach them to metrics. This works by adding some "objective"-related labels to the metrics produced by each function. My label set thus contains some values that will always be there and some that are optional.

To be honest, it doesn't matter to me whether they show up as label keys with empty values or just aren't attached at all (and I'm not sure if it makes a difference to Prometheus as querying label_key="" apparently has the same effect as leaving off that part of the query). It might make slightly more sense to leave off the label keys entirely.

Here is a detailed write-up of the SLO feature in autometrics-rs: https://fiberplane.com/blog/autometrics-rs-0-3-defining-service-level-objectives-in-rust-source-code

[mxinden]

Thanks @emschwartz and @drbrain for the additional input.

and I'm not sure if it makes a difference to Prometheus as querying label_key="" apparently has the same effect as leaving off that part of the query

Backing this up:

Empty label values SHOULD be treated as if the label was not present.

https://github.com/OpenObservability/OpenMetrics/blob/main/specification/OpenMetrics.md#label

With that, I am fine either way. In case neither of you objects to this pull request as is, would you @emschwartz mind bumping the minor version and adding a changelog entry?

[mxinden]

Released through #141 as v0.21.1.

[emschwartz]

Thanks @mxinden!