feature request/RFC: leave out a label (from at least `*MetricFamily` samples) when it's value is set to `None`

Hey.

This is somewhat related to https://github.com/prometheus/docs/issues/2977, i.e. “what to do if some values of a sample aren’t available”, which obviously may not just be the actual metric value but also values of labels.

AFAIU, labels values are always strings, so sometimes it may be possible to use a special label value to indicate an error, for example if the label would be a host, one might simply use some characters that cannot appear in hostnames or addresses and thus either an empty string or perhaps <unknown> or similar values depending on the error.

But sometimes this may not be possible, in particular when the possible label values are arbitrary (like I have a case where the label is a description which may contain any string including the empty one).

Now in the error case (at least in my use case) it's better to have the rest of the data, but it still feels wrong to use a value for it that would be valid, even though I couldn't determine it on that particular scrape.
The time series is likely anyway already interrupted, so what's IMO the cleanest solution is simply leaving out the respective label in the error case, even if I have another metric which somehow indicates that sometime is fishy (like _up or so).

(Or any better ideas?)

Now I guess often, exporters will simply have a set of *MetricFamily-objects and .add_metrics to it, which requires however all the values for the defined labels to be given.

My proposal/RFC here is: Would it make sense to special-handle a value of None (which currently seems to simply lead to an error), causing it to leave that particular label out from the printed data for the particular label only?

I guess I could look in implementing this, but on a first short glance it didn't seem straightforward, so I wanted to ask what people even think about it.

Cheers,
Chris.

首先定位 *MetricFamily 类、add_metric 以及输出序列化路径的实现，然后检查当前如何验证 None 标签值。在确认项目维护者同意此 RFC 后，定义并测试仅省略受影响标签、同时保留其余样本数据的行为。