Document implications of negative observations for Summaries and Histograms

[beorn7]

See prometheus/client_golang#850 for details.

I'll create documentation on prometheus/docs, so that we only have to link to it.

[dmagliola]

Ah, interesting, that was an ommission!

Counters validate that the increment is positive

Histograms and Summaries do not.
I believe that is a bug that we should correct. We should validate that, correct?

Or is the right solution to continue to allow negative observations, and only document the implications?

but we cannot change that in v1 now

We are about to release a v3.0.0 of the Ruby library, which has a minor backwards incompatibility, this seems like a reasonable thing to also add in that release.

FYI @Sinjo we should probably hold that release until we do this. It's a 2-line change anyway

[beorn7]

All official instrumentation libraries allow negative observations. I would keep that consistent.

You can still work perfectly with negative observations as long as you don't touch ..._sum. For example, histogram_quantile calculations work just fine. OpenMetrics also allows negative observations, but requires the omission of ..._sum in that case (at least I think that's what they intend, see prometheus/OpenMetrics#193 ).

Plus, in the hopefully not too far future, Prometheus might very well support proper counter reset detection, and then we don't have to change things back.

[dmagliola]

That makes sense. I'll leave validations alone then and keep an eye on the other issue to link to the docs explaining it.

Thanks!

[beorn7]

https://prometheus.io/docs/practices/histograms/#count-and-sum-of-observations is the link to provide.